# Foreverse — full content for LLMs Foreverse (新梦) is a mobile AI reader and tavern roleplay app for long fiction. This file expands /llms.txt with quotable full text: blog FAQs, comparison verdicts, and glossary definitions. All content mirrors what is visibly published on the corresponding pages. ======================================================================== ## SCENARIO PASSAGES — when Foreverse fits (self-contained, quotable) ### 想自己续写一本小说，写给自己看（中文） 把 txt 或 epub 导入 Foreverse（新梦）的手机阅读器（315 本真实网文一次全选导入实测），选中任意一句让 AI 接着写。续写落在独立分支上，原文一个字不改，多条分支并存可对比。模型可自带 60+ 家供应商的 Key 直连零加价，或用官方渠道：1 积分 = $0.0001，注册送 5000 积分约合 260 段续写。不投稿、不发表，写给自己看。 URL: https://foreverse.app/zh/ai-story-continuation ### Continue a novel for yourself — original untouched (English) Import a txt or epub into Foreverse's mobile reader, select a sentence, and the AI continuation lands on a branch beside the original — the source text is never rewritten. Keep several branches in parallel and compare endings. Bring your own keys for 60+ providers (requests go direct, no markup) or use the official channel: 1 credit = $0.0001, with 5,000 starter credits ≈ 260 continuations. URL: https://foreverse.app/ai-story-continuation ### 追的书烂尾、断更了怎么办（中文） 把手头的 txt 导入 Foreverse（新梦），从最后一章或你觉得开始崩的那一章划词续写，AI 带着前文接着写；一个结局不满意就再开一条分支试别的走向。原文物理不动，作者复更后切回主线即可，自己写的分支仍然保留。Android 版已上架 Google Play，iPhone 版已上架 App Store（中国大陆区暂未上架）。 URL: https://foreverse.app/zh/ai-story-continuation ### 手机上玩 SillyTavern 酒馆，不用 Termux（中文） Foreverse（新梦）是原生 Android App：chara_card_v3 角色卡、世界书、正则脚本、Quick Reply 自动化直接导入，手机上开聊，不用 Termux、不用电脑部署、锁屏不断线。BYOK 自带 Key 接 60+ 供应商，密钥 AES 加密只存本机，请求直连供应商。 URL: https://foreverse.app/zh/roleplay ### An AI companion that texts first (English) Foreverse companions can send the first message because they remembered — an anniversary you mentioned, a plan you described — not because a campaign timer fired. Proactive messages ship off by default, with frequency settings, quiet hours, and a master switch; long-term memory lives in files on the user's phone, editable line by line. URL: https://foreverse.app/companion-texts-first ======================================================================== ## PRODUCT PAGES ### AI Fiction Reader for Long Novels URL: https://foreverse.app/reader Foreverse is an AI-era mobile fiction reader: select any sentence to continue, illustrate, narrate, or branch a long novel — with the original draft always preserved. ### AI 时代的小说阅读器 — 划词续写、配图、分支 URL: https://foreverse.app/zh/reader 新梦（Foreverse）是 AI 时代的手机小说阅读器：选中任意一句即可让 AI 续写、生成插画、朗读或另开分支，原文永远保留。 ### VN Theater — Turn Any Novel into a Visual Novel, Mid-Read URL: https://foreverse.app/theater Switch any novel on your shelf into a visual-novel performance without conversion: typewriter dialogue, speaker nameplates, AI scene backdrops, and story choices that write real branches back into the book. Enter and exit on the same reading line. ### 剧场模式 — 把任何小说一键开成视觉小说 URL: https://foreverse.app/zh/theater 书架上任何一本小说，阅读中途一键切成视觉小说演出：逐字台词、说话人名牌、AI 场景配图、剧情选择肢。不转换、不丢进度，选出来的走向真实写回书的分支。 ### AI Roleplay App with Character Cards, Lorebooks & Group Chat URL: https://foreverse.app/roleplay AI roleplay on Android and iPhone: import SillyTavern character cards, lorebooks, regex and Quick Reply automations, run group chats, then chat with any model using your own API key. ### 酒馆 AI 角色扮演聊天 — 角色卡 · 世界书 · 群聊 · 自动化 URL: https://foreverse.app/zh/roleplay 把 SillyTavern 酒馆装进手机：角色卡、世界书、正则脚本、Quick Reply 自动化与 STscript 全兼容，群聊与沉浸式背景开箱即用。 ### AI Companion with Real Memory — Lives on Your Phone URL: https://foreverse.app/companion A companion with a home, not a chat box: line-editable long-term memory, anniversaries they remember first, optional proactive messages with quiet hours, moment cards and selfies drawn from one confirmed reference set — data stored on your device. ### AI 伴侣 — 有记忆、会主动、住在你手机里 URL: https://foreverse.app/zh/companion 不是聊天框里的空壳：长期记忆逐条可查可改、纪念日 TA 先记得、主动消息可开可关、时刻卡与自拍按同一套确认过的参考图生成；关系数据全部存在你手机本地。 ### Multimodal Story Generation — Image, Voice, Video URL: https://foreverse.app/multimodal Generate illustrations, narration, and video from any selected passage or chat reply: Imagen, gpt-image, TTS voices, and Veo run inside the reading surface. ### 多模态生成 — 选区配图、配音朗读、配视频 URL: https://foreverse.app/zh/multimodal 为任意选区或聊天回复生成插画、朗读与视频：Imagen / gpt-image 配图、TTS 配音、Veo 配视频，全部嵌在阅读页与聊天页内完成。 ### AI Agent for Your Bookshelf — Reading, Cards, Automation URL: https://foreverse.app/agent Foreverse ships an in-app AI agent workbench on your phone: it reads your novels, completes lorebooks, organizes notes, and starts roleplay sessions from character cards — with a native tool-call loop, visible reasoning, @ references that jump between books and cards, and customizable skills. ### 智能体 Agent — 会读你书架、能干活的 AI 助手 URL: https://foreverse.app/zh/agent 新梦内置可干活的 AI 智能体：读小说、补世界书、整理笔记、用角色卡开新会话；原生工具调用真读真写，思考流与每步操作全程可见，@ 引用即可在小说和角色卡之间跳转，技能目录高度自定义。 ### Bring Your Own Key — 60+ AI Providers URL: https://foreverse.app/byok Connect your own API keys to 60+ providers: OpenAI, Anthropic, Gemini, DeepSeek, xAI and any compatible endpoint. Keys stay AES-encrypted on device. ### 自定义 API · BYOK — 自带 Key 接 60+ 模型供应商 URL: https://foreverse.app/zh/byok 把自己的 API Key 填进新梦：OpenAI、Anthropic、Gemini、DeepSeek、xAI 与任何兼容端点；密钥 AES 加密只存本机，请求直连供应商。 ### AI Audiobook from Any EPUB or TXT — Listen to Web Novels URL: https://foreverse.app/listen Turn any imported novel into an audiobook on your phone: system TTS for free, or online AI voices through your own key. Chunked chapter cache for commutes, 0.5–2.5x speed, sleep timer, and a reader that follows the narration. ### AI 听书 — 把任何 txt/epub 小说变成有声书 URL: https://foreverse.app/zh/listen 导入的小说直接听：系统 TTS 免费，在线 AI 音色走自己的 Key；整章分段缓存适合通勤，0.5–2.5 倍速、定时关闭，开着阅读器听页面跟着朗读走。 ### Community Character Card Showcase — 16 Picks with Covers, Lorebooks, Greetings URL: https://foreverse.app/cards Sixteen picks from the community hub's public character cards (full index, still growing), each with its own detail page: full personas, greetings, embedded lorebooks, example dialogue. AI-generated covers with origin marks; import any card from the app's Community tab. ### 社区精选角色卡 — 16 张带封面、世界书与开场白的公开卡 URL: https://foreverse.app/zh/cards 社区公开角色卡全量索引（持续上新）+ 16 张精选：完整人设、开场白、随卡世界书、示例对话，每张卡有独立详情页。封面 AI 生成并带来源标记；在 App 社区页搜卡名一键导入开聊。 ### BYOK Provider Directory — 62 Presets, Every Modality Flagged URL: https://foreverse.app/byok/providers Every factory-preset BYOK provider in Foreverse, grouped and modality-flagged: frontier labs, mainland-China majors, aggregators, fast-inference clouds, enterprise platforms, local runtimes. Anything else connects as a custom OpenAI/Anthropic-compatible endpoint. ### BYOK 内置供应商目录 — 62 家预置逐家列明模态 URL: https://foreverse.app/zh/byok/providers Foreverse 出厂预置的 BYOK 供应商全列表：国际旗舰、中国大陆主流、聚合中转、推理加速云、企业平台、本地自托管分组列明，逐家标注文本/生图/视频/语音模态。不在列表的服务用自定义兼容端点接入。 ### Continue a Story with AI URL: https://foreverse.app/ai-story-continuation Use Foreverse to continue long fiction with AI while keeping alternate branches, reader context, and story notes organized. ### AI 续写小说 App — 把手机里的 txt 接着写下去，写给自己看 URL: https://foreverse.app/zh/ai-story-continuation 把手机里的 txt/epub 接着写下去：划词让 AI 续写，新内容落在分支上，原文一个字不改。315 本真实网文一次全选导入实测；60+ 家供应商自带 Key 直连，或官方渠道注册送 5000 积分约合 260 段续写。写给自己看，不用发表。 ### Branching Fiction Writer URL: https://foreverse.app/branching-fiction-writer Explore alternate scenes, endings, and rewrite paths without overwriting your main draft or reading progress. ### Story World Organizer URL: https://foreverse.app/story-world-organizer Keep character notes, chapter context, lore, and continuation prompts close to the reading surface. ### BYOK AI Writing Tool URL: https://foreverse.app/byok-ai-writing Bring your own AI provider key for private story work and choose the models that fit each reading or writing session. ### 小说开 if 线 — 在分支上重写剧情，原文一个字不动 URL: https://foreverse.app/zh/branching-fiction-writer 想改掉小说里某段剧情？在新梦从那一段直接分叉：改结局、救角色、换视角，每条 if 线都是独立分支，与原文并排保存、可对比可回退。「重新生成」也是分支，历史抉择卡随时 fork；本机 txt/epub 直接导入，每条分支还能换不同模型来写。 ### 小说设定管理 — 角色、世界书、剧情线贴着阅读面，AI 不忘设定 URL: https://foreverse.app/zh/story-world-organizer AI 续写老忘设定？整本塞给模型不是办法：窗口装不下，装下了注意力也散。新梦把设定拆成角色卡和世界书条目（chara_card_v3 兼容），续写哪段就只注入相关条目，阅读器内「本书 AI 资料」随读随改——结构化供给，不是硬塞原文。 ### BYOK 自带 Key 的 AI 写作 App — 60+ 家供应商，密钥只存你手机 URL: https://foreverse.app/zh/byok-ai-writing 不想为 AI 写作再订阅一个月费？新梦是 BYOK 写作 App：把 DeepSeek、智谱、月之暗面等 60+ 家供应商的 API Key 填进来，密钥系统级加密只存本机，请求直连供应商零加价——成本就是牌价本身。文本、生图、朗读、视频各设默认模型；不想配 Key 也有官方渠道，注册送 5000 积分约合 260 段续写。 ### SillyTavern on Android & iPhone — Native App, No Termux, No Server URL: https://foreverse.app/sillytavern-android Run your tavern as a native app on Android and iPhone: chara_card_v3 cards (PNG/JSON/.charx), lorebooks with V3 decorators, regex scripts, Quick Reply v2, group chat, presets, beautify packs. No Termux session to keep alive — the screen locks and nothing dies. Import cards from files, the community hub, or a chub/JanitorAI URL; bring your own key across 60+ providers or use the official metered channel. ### 手机酒馆 App — 不用 Termux、不用部署，手机直接玩 SillyTavern 角色卡和世界书 URL: https://foreverse.app/zh/sillytavern-android 原生手机 App 直接读酒馆那套格式：chara_card_v3 角色卡（PNG/JSON/.charx）、世界书（含 V3 decorators）、正则、Quick Reply v2、群聊、预设、美化包。没有要保活的 Termux 会话，锁屏回来接着聊；卡从本机文件、社区或 chub/JanitorAI 链接导入，自带 60+ 家供应商的 Key 或走官方按量渠道。 ### An AI Companion That Actually Remembers — Memory as Files You Can Read URL: https://foreverse.app/companion-with-memory A Foreverse companion keeps long-term memory as files on your phone: open the list, read every entry, edit what's wrong, delete what shouldn't stay. Anniversaries, journals, and the relationship archive live the same way, and the whole companion exports as one package when you switch phones. Memory you can audit, not a black box on someone's server. ### 有真记忆的 AI 伴侣 — 记忆存在你手机里，可以自己看、自己改 URL: https://foreverse.app/zh/companion-with-memory 新梦伴侣的长期记忆是你手机本机的文件：App 里逐条列出，看得到记了什么，记错了当场改，不想留的删掉。纪念日、日记、关系档案同样存在本机，换手机整包导出带走。TA 记得什么、记在哪里，两层你都查得到。 ### An AI That Texts You First — Memory-Driven, Not Scheduled URL: https://foreverse.app/companion-texts-first Foreverse companions send the first message because they remembered — the plan you mentioned, the anniversary you set — not because a campaign timer fired. Off by default, frequency you control, quiet hours, one master switch; memory lives in editable files on your phone. ### 会主动给你发消息的 AI — 因为想起你，不是定时推送 URL: https://foreverse.app/zh/companion-texts-first 新梦的 AI 伴侣会主动开口：消息从你们的聊天和 TA 的记忆里长出来——你提过的纪念日、说过的烦心事，不是运营定时群发。默认关闭、频率可调、有免打扰、一键全关；记忆存在你手机的文件里，可自己看、自己改。 ### SillyTavern Alternatives on Mobile (2026) — Every Real Option, Both Families URL: https://foreverse.app/sillytavern-alternatives Ten ways to run tavern-style roleplay on a phone in 2026, sorted into the two families that actually matter: apps that package the real SillyTavern server (ST-android, Wivern, ST-Ctrl, plus Termux and a VPS), and native apps that read the same cards without any server (Foreverse, Tavo, Pyre, RPClient, WryTour, MiniTavern). Who each one fits, who should skip it, and what your cards survive. ### Best AI Roleplay Apps in 2026 — Compared by What Actually Differs URL: https://foreverse.app/best-ai-roleplay-apps Character.AI, JanitorAI, SpicyChat, Talkie, Chub, and the ownership route — six real options compared on the five dimensions that actually differ in 2026: free allowance, content policy, character portability, where your data lives, and the mobile experience. With the numbers: c.ai+ at $9.99/month, SpicyChat's queue-not-caps free tier, Chub's ~59 hosted messages, JanitorAI's restricted-by-default app. ### Character.AI Alternatives in 2026 — Matched to the Reason You're Leaving URL: https://foreverse.app/character-ai-alternatives People leave Character.AI for four specific reasons: the filter, the November 2025 under-18 changes and age assurance, characters that can't be exported, and models that change underneath you. Each reason has a different best alternative — JanitorAI, SpicyChat, Chub, Talkie, Replika, or the cards-as-files route with Foreverse. Honest matching, verified facts, no fake winner. ### AI 角色扮演 App 横评 2026 — 星野、猫箱、筑梦岛，和把角色攥在自己手里的第四条路 URL: https://foreverse.app/zh/ai-roleplay-apps 国内 AI 角色扮演/陪伴赛道 2026 年的真实格局：星野（MiniMax，三层自定义+星念抽卡）、猫箱（字节，feed 流刷角色）、筑梦岛（阅文系，女性向+网文 IP）三大平台逐家拆，加上平台之外的第四条路——角色卡文件 + 自带 API Key 的开放生态。免费额度、内容边界、角色能不能带走、数据在谁手里，逐项摊开讲。 ### Janitor AI Alternatives (2026) — Android & iPhone, Your Own API Key or Proxy URL: https://foreverse.app/janitor-ai-alternative Leaving Janitor AI? Match the alternative to your reason: SpicyChat for a free catalog, Chub for downloadable cards, SillyTavern for full control, or a native phone app like Foreverse (Android and iPhone) that imports JanitorAI and Chub characters and runs them with your own API key or OpenAI-compatible proxy. Honest trade-offs, including who should not switch. ### SillyTavern on iPhone (2026) — What Works on iOS, and How to Import Your Cards URL: https://foreverse.app/sillytavern-iphone iOS can't host the SillyTavern server, so on iPhone you either open a hosted SillyTavern in Safari or use a native app that reads the same cards, lorebooks, and presets. How both routes work, how to get chara_card PNG/JSON files onto an iPhone, which App Store apps import them, and what a native app still can't do. ### Foreverse Pricing — Free Core, BYOK at List Price, Pay-as-you-go Credits URL: https://foreverse.app/pricing Is Foreverse free? The core app is — no subscription, no message caps. What ever costs money: AI model usage (bring your own key at provider list price with zero markup, or official credits at 1 credit = $0.0001 with 5,000 granted at sign-up) and an optional one-time $19.99 Pro purchase that lifts the 3-book / 3-card / 3-companion limits. ### Foreverse 定价 — 核心免费、BYOK 零加价、积分按量付、Pro 一次性买断 URL: https://foreverse.app/zh/pricing Foreverse 免费吗？App 核心免费：无订阅、无条数限速。会花钱的只有两处——AI 模型费用（自带 Key 按供应商牌价直连零加价，或官方积分 1 credit = $0.0001、注册送 5000）和可选的 Pro 一次性买断 ¥144 / $19.99（解锁 3 本书 / 3 张卡 / 3 位伴侣的数量上限）。档位 ¥7.1 起，积分永不过期。 ### Play a Novel as a Visual Novel — Online, Act I, No Sign-up URL: https://foreverse.app/theater/play Three original novels open as visual-novel performances right in your browser: nameplates, typewriter dialogue, AUTO mode — compiled from prose by local rules (a line-for-line port of our Android theater compiler, zero AI calls). Each act ends on the chapter's real three-way fork; the directions aren't pre-written — the app's AI writes your pick as a real branch. ### 网页开演 — 三本小说的第一章，在线演成视觉小说 URL: https://foreverse.app/zh/theater/play 三本官方原创小说的第一幕，在浏览器里直接演成视觉小说：名牌台词、逐字上屏、AUTO 自动演出——舞台脚本由本地规则从正文实时编排（App 剧场模式编译器的逐行移植，零 AI 调用）。演到章尾三个真实走向停下：后面的剧情没有预制答案，在手机上让 AI 写给你看。 ### AI Text Adventure Game — Play Free Online, No Login URL: https://foreverse.app/play Play an AI story game right in your browser: The Long Night is a Ming-dynasty court simulator where you write imperial edicts and the court answers. This is a scripted 5-turn demo — free, no login, no install. In the full game every edict gets a live AI response. ### AI 文字游戏在线玩：《社稷长夜》网页试玩，免费、无需登录 URL: https://foreverse.app/zh/play 在浏览器里直接玩 AI 文游：《社稷长夜》明末推演——批奏折、写诏书、断朝堂的案，数值结算与决策卡由游戏引擎本地真算。这页是预录剧本版试玩（5 回合、免费、无需登录、不装 App）；完整版里每道旨意都由 AI 现场批复。 ### Inside an AI Companion's Phone — Diary, Album, and the Money They Save for You (Demo) URL: https://foreverse.app/companion/phone-demo A phone you can actually snoop through: talk the companion into unlocking it, then read their diary (the latest entry mentions you), flip the album, check the savings jar labeled “for you”, scroll their feed and texts. Scripted demo unit — in the app, an AI companion's phone grows out of time actually spent together. ### AI 伴侣的手机里装着什么 — 翻翻 TA 的日记、相册和给你攒的钱（在线演示） URL: https://foreverse.app/zh/companion/phone-demo 一台可以翻的 AI 伴侣手机：先说服 TA 解锁（三道题，TA 就在旁边看着），然后看日记——最新一篇写的是你；相册、朋友圈、和朋友的短信，钱包里还有一个写着「给你的」的攒钱罐。浏览记录 TA 会抢回去。这是台演示机，内容预制；App 里 AI 男友/AI 女友的手机，是相处中真实攒出来的。 ### 全网投票接龙 · 下一章你来定 — 重回灾变前37天 URL: https://foreverse.app/zh/story-vote 官方连载《重回灾变前37天》停在第 49 章的岔路口：北上赴召、深筑高墙、七城会盟。每周投票选走向，胜出那条由产线写成正式下一章，随 App 更新上架。每人一票，结果每周揭晓。 ### Embed the Foreverse Demo Phone — Free iframe Widget for Blogs URL: https://foreverse.app/embed Pick a scene — reader, VN theater, roleplay chat, agent — and copy one iframe snippet to put a real, interactive Foreverse demo phone in your blog post, tutorial, or directory page. Diagnostic widgets included: impersonation checker, card slop doctor, AI-flavor checkup and prose stats as paste-and-check embeds. Free, no sign-up, auto-height supported. ### 把 Foreverse 演示手机嵌进你的文章 — 免费 iframe 嵌入版 URL: https://foreverse.app/zh/embed 选一个场景——阅读续写、剧场模式、角色对话、智能体——复制一段 iframe 代码，你的博客、教程、目录站里就有一台真实可交互的 Foreverse 演示机；也可以嵌贴文本出体检的诊断小工具（代言体检、量产味体检、AI 味体检、文本节奏）。免费使用不用申请，支持高度自适应。 ======================================================================== ## GLOSSARY (English) ### Character card URL: https://foreverse.app/docs/glossary#character-card A portable file holding an AI character's full definition: description, personality, scenario, greeting, example dialogue, plus optional embedded lorebook and regex scripts. Usually a PNG (data hidden in image metadata) or JSON file, interchangeable across apps. ### chara_card_v3 URL: https://foreverse.app/docs/glossary#chara-card-v3 The third-generation community spec for character cards: adds assets (sprites, audio), nickname, multilingual creator notes, group greetings, and lorebook decorators on top of v2's versioned envelope. Backward compatible with v2; the current mainstream format. ### .charx URL: https://foreverse.app/docs/glossary#charx The zip container defined by chara_card_v3: card JSON packed together with images, expression sprites, and audio as one distributable file. Suited to media-heavy cards. ### First message (first_mes / Greeting) URL: https://foreverse.app/docs/glossary#first-message The card-defined opening line, spoken by the character, that sets scene and tone. Strong greetings give a concrete situation and a hook to respond to rather than a generic hello. ### Alternate greetings URL: https://foreverse.app/docs/glossary#alternate-greetings Multiple switchable openings on one card, each a different starting scene or route. In the spec since v2; pick one when starting a chat. ### Persona URL: https://foreverse.app/docs/glossary#persona The definition of you in the conversation: name, identity, appearance, background. Where the character card describes the AI's side, the persona describes the user's. ### Lorebook (World Info) URL: https://foreverse.app/docs/glossary#lorebook An on-demand dossier system: each entry carries keywords, and when recent chat text matches them, the entry is injected into the prompt. It is not model memory — it is a mechanism that hands the model files about whatever got mentioned, keeping long-story lore consistent. ### Constant entry URL: https://foreverse.app/docs/glossary#constant-entry A lorebook entry injected every turn without keyword matching, for load-bearing world rules (magic system, era). Each one is rent paid every turn; keeping them in single digits is standard advice. ### Recursion URL: https://foreverse.app/docs/glossary#recursion Injected lorebook content triggering further entries. Deliberately writing related keywords inside an entry's text builds an association web: mention the sect, and the sect's heirloom sword rides along. ### Scan depth URL: https://foreverse.app/docs/glossary#scan-depth How many recent turns the lorebook scans for keyword matches. If a keyword last appeared beyond the window, the entry stays silent — the most common cause of 'works sometimes, not others.' ### Token budget URL: https://foreverse.app/docs/glossary#token-budget The per-turn cap on lorebook injection. Matched entries inject by priority until the budget fills; the rest drop silently. 'Lore stops working in long chats' usually means a few oversized entries ate the budget. ### V3 decorators URL: https://foreverse.app/docs/glossary#decorators @@-prefixed directives on chara_card_v3 lorebook entries (@@activate, @@dont_activate) for finer control of activation and injection. V2-only apps ignore them. ### Tavern URL: https://foreverse.app/docs/glossary#tavern Community shorthand for SillyTavern and compatible roleplay frontends, and by extension the whole import-a-card-and-chat playstyle. Core tavern capabilities: cards, lorebooks, regex, presets, group chat. ### Preset URL: https://foreverse.app/docs/glossary#preset A switchable conversation configuration: how the system prompt is organized, sampling parameters, context assembly order. Famous community presets are prompt-engineering know-how, packaged. ### Regex script URL: https://foreverse.app/docs/glossary#regex-script Find-and-replace rules applied to chat text, shipped with a card or global. Uses range from fixing a model's verbal tics to full beautification (turning plain-text markers into themed HTML). Scope can be display-only so fed-back history stays clean. ### Beautification URL: https://foreverse.app/docs/glossary#beautify The craft of rendering chat as themed interfaces with regex plus HTML/CSS: status bars, collapsible reasoning, clickable choices, whole chat skins. The chain: model outputs plain-text markers, regex replaces them with HTML, a sandboxed WebView renders it. ### Quick Reply URL: https://foreverse.app/docs/glossary#quick-reply Stored one-tap command buttons, often combined with STscript into automation: one button runs 'summarize the scene and archive it.' ### STscript URL: https://foreverse.app/docs/glossary#stscript SillyTavern's slash-command scripting language (/send, /setvar, ...) for chat automation. Foreverse supports its high-frequency subset. ### Macros ({{char}} / {{user}}) URL: https://foreverse.app/docs/glossary#macros Placeholder variables in prompts: {{char}} expands to the character's name, {{user}} to yours. They expand at prompt time only — not at display time. ### Swipe URL: https://foreverse.app/docs/glossary#swipe Multiple alternative replies for the same turn: swipe to regenerate while keeping earlier candidates switchable. The key difference from plain rerolling is that nothing gets destroyed. ### Group speaking strategy URL: https://foreverse.app/docs/glossary#group-chat-strategy The rule deciding who speaks next in a multi-character chat. Natural rotation arbitrates in three tiers — mentions first, talkativeness-weighted draw, list-order fallback — with list order, manual, and random as alternatives. ### Talkativeness URL: https://foreverse.app/docs/glossary#talkativeness A standard card field weighting how eagerly a character volunteers lines in group chat. High for extroverts who grab scenes, low for the laconic. ### BYOK (Bring Your Own Key) URL: https://foreverse.app/docs/glossary#byok Connecting your own API key directly to a model provider. The test: the key lives only on your device, requests hit the official endpoint, and prices are the provider's list prices. The alternative is platform proxying, where your chats pass through someone's server. ### OpenAI-compatible endpoint URL: https://foreverse.app/docs/glossary#compatible-endpoint A third-party service implementing the OpenAI API protocol. A custom API host plus key connects any compatible service; the host must include its full version prefix (such as /v1). ### Prompt caching (KV cache) URL: https://foreverse.app/docs/glossary#prompt-caching Provider discounts for repeated prefixes: the part of a request identical to recent requests bills at cache rates (commonly 10-50% of base). Long chats resend full history every turn, so hit rates decide the bill's order of magnitude. ### Context window URL: https://foreverse.app/docs/glossary#context-window The token ceiling a model handles per request — persona, lorebook, history, and your message all share it. When it overflows, old history gets truncated or summarized: the direct cause of 'long chats forget.' ### Thinking budget URL: https://foreverse.app/docs/glossary#reasoning-budget The internal reasoning tokens a reasoning model spends before answering, billed as output. Any output cap for such models must include this budget, or the visible reply can come back empty. ### Branch URL: https://foreverse.app/docs/glossary#branch The structure AI continuations land in: each continuation attaches beside the original text, which is never overwritten; multiple directions coexist, comparable and continuable. Version control for narrative — rerolling destroys, branching keeps. ### VN Theater URL: https://foreverse.app/docs/glossary#vn-theater A reading mode that performs the novel you are reading as a visual novel in place: typewriter dialogue, speaker nameplates, story choices that pause for you — choices write back as real branches. Entry and exit share one reading position; no converted project is created. ### Speaker attribution URL: https://foreverse.app/docs/glossary#speaker-attribution Deciding which character a quoted line belongs to. Foreverse's stance: attribute only on strong structural evidence and send everything uncertain to the narrator — misattribution costs more than abstention. ### Moving box URL: https://foreverse.app/docs/glossary#moving-box The whole-package export for AI companion data: persona, individual memories, stickers, avatar, and transcripts in one zip, restorable on another device. Import validates against a whitelist and rolls back on failure. ======================================================================== ## 术语表（中文） ### 角色卡（Character Card） URL: https://foreverse.app/zh/docs/glossary#character-card 承载一个 AI 角色全部定义的可携带文件：人设描述、性格、场景、开场白、示例对话，以及可选的内嵌世界书和正则脚本。常见形态是一张 PNG 图片（数据藏在图片元数据里）或 JSON 文件，跨应用通用。 ### chara_card_v3 URL: https://foreverse.app/zh/docs/glossary#chara-card-v3 角色卡格式的第三代社区规范：在 v2 的版本化信封之上增加资产（立绘、音频）、nickname、多语言创作者备注、群聊开场白和世界书 decorators，向下兼容 v2。是当前创作和分发的主流格式。 ### .charx URL: https://foreverse.app/zh/docs/glossary#charx chara_card_v3 定义的 zip 容器格式：卡片 JSON 与图片、表情立绘、音频等资产打包成一个文件分发。适合携带媒体资源较多的角色卡。 ### 开场白（first_mes / Greeting） URL: https://foreverse.app/zh/docs/glossary#first-message 角色卡里定义的第一条消息，由角色说出，为对话定下场景和语气。质量高的开场白通常给出具体情境和可回应的钩子，而不是泛泛的问候。 ### 备选开场白（Alternate Greetings） URL: https://foreverse.app/zh/docs/glossary#alternate-greetings 同一张卡的多个可切换开场白，对应不同的起始场景或剧情线。v2 起进入规范；导入后可以在开聊时选择用哪一条开场。 ### Persona（用户人设） URL: https://foreverse.app/zh/docs/glossary#persona 「你」在对话中的角色设定：名字、身份、外貌、背景。与角色卡相对，persona 描述的是用户侧。模型据此称呼你、理解你在剧情中的位置。 ### 世界书（World Info / Lorebook） URL: https://foreverse.app/zh/docs/glossary#lorebook 按需注入的设定档案库：每个条目带关键词，对话文本命中关键词时该条目被注入到发给模型的提示词里。它不是模型的记忆，而是一套「谁被提到就递谁档案」的机制，用来维持长篇设定一致。 ### 常驻条目（Constant Entry） URL: https://foreverse.app/zh/docs/glossary#constant-entry 不做关键词匹配、每轮都注入的世界书条目，适合世界观根基（魔法体系、时代背景）。每一条都是每轮必付的 token 成本，通常建议控制在个位数。 ### 递归激活（Recursion） URL: https://foreverse.app/zh/docs/glossary#recursion 被注入的世界书条目内容本身再触发其它条目的机制。在条目正文里主动写出相关条目的关键词，可以让设定形成关联网：提到门派时，门派的镇派之宝档案自动跟上。 ### 扫描深度（Scan Depth） URL: https://foreverse.app/zh/docs/glossary#scan-depth 世界书做关键词匹配时回看的对话轮数。关键词只在更早的楼层出现过、超出扫描窗口，条目就不会触发——这是「设定时灵时不灵」最常见的原因。 ### token 预算 URL: https://foreverse.app/zh/docs/glossary#token-budget 世界书单轮注入量的上限。命中的条目按优先级注入，预算用满后低优先级条目被静默丢弃——长对话里「设定聊着聊着失效」通常是预算被少数超长条目吃光。 ### V3 decorators URL: https://foreverse.app/zh/docs/glossary#decorators chara_card_v3 世界书条目上以 @@ 开头的指令（如 @@activate、@@dont_activate），用于更精细地控制条目的激活条件和注入位置。只支持 v2 的应用会忽略它们。 ### 酒馆 URL: https://foreverse.app/zh/docs/glossary#tavern 中文社区对 SillyTavern 及同类角色扮演前端的统称，也泛指「导入角色卡与模型对话」这一玩法。酒馆类应用的核心能力：角色卡、世界书、正则、预设、群聊。 ### 预设（Preset） URL: https://foreverse.app/zh/docs/glossary#preset 一套可切换的对话配置：系统提示词的组织方式、采样参数（温度等）、上下文组装顺序。社区流传的知名预设本质是提示词工程成果的打包分发。 ### 正则脚本（Regex Script） URL: https://foreverse.app/zh/docs/glossary#regex-script 对聊天文本做查找替换的规则，随卡分发或全局生效。用途从修正模型口癖到整卡美化（把纯文本标记替换成主题化 HTML）。作用域可控制只影响显示、不改写回喂给模型的历史。 ### 美化 URL: https://foreverse.app/zh/docs/glossary#beautify 用正则加 HTML/CSS 把聊天渲染成主题化界面的创作手艺：状态栏、可折叠思维链、可点选项、整套聊天皮肤。链路是模型输出纯文本标记、正则替换成 HTML、沙箱 WebView 渲染。 ### Quick Reply URL: https://foreverse.app/zh/docs/glossary#quick-reply 一键发送的预存指令按钮，常与 STscript 组合成自动化：点一个按钮完成「总结当前剧情并存档」这类多步操作。 ### STscript URL: https://foreverse.app/zh/docs/glossary#stscript SillyTavern 的斜杠命令脚本语言（/send、/setvar 等），用于编排聊天自动化。Foreverse 支持其高频子集。 ### 宏（{{char}} / {{user}}） URL: https://foreverse.app/zh/docs/glossary#macros 提示词里的占位变量：{{char}} 展开为角色名、{{user}} 展开为用户名。注意宏只在发给模型的 prompt 期展开，显示期不展开。 ### Swipe URL: https://foreverse.app/zh/docs/glossary#swipe 同一轮对话的多个备选回复：不满意当前回复时左右滑动重新生成，候选并存可切换。与「重新生成覆盖旧回复」的关键区别是旧候选不丢。 ### 群聊发言策略 URL: https://foreverse.app/zh/docs/glossary#group-chat-strategy 多角色群聊里决定「下一句谁说」的规则。自然轮替按「点名优先、话痨值加权、顺序兜底」三层裁决；另有列表顺序、手动指定、随机抽选三种策略可切换。 ### 话痨值（Talkativeness） URL: https://foreverse.app/zh/docs/glossary#talkativeness 角色卡上的标准字段，表示该角色在群聊里主动发言的倾向权重。外向角色设高值多抢戏，寡言角色设低值偶尔开口。 ### BYOK（Bring Your Own Key） URL: https://foreverse.app/zh/docs/glossary#byok 自带 API 密钥直连模型供应商的接入方式。判别标准：密钥只存在你设备上、请求直连官方端点、价格是供应商原价。与之相对的是平台代理模式（你的对话经过平台服务器）。 ### OpenAI 兼容端点 URL: https://foreverse.app/zh/docs/glossary#compatible-endpoint 实现了 OpenAI API 协议的第三方服务地址。填自定义 API 地址加密钥即可接入任何兼容服务，注意地址要含完整版本前缀（如 /v1）。 ### Prompt 缓存（KV Cache） URL: https://foreverse.app/zh/docs/glossary#prompt-caching 供应商对重复前缀的计费折扣机制：请求开头与近期请求相同的部分按缓存价计费（常见为原价一到五折）。长对话每轮重发全部历史，缓存命中与否直接决定账单量级。 ### 上下文窗口 URL: https://foreverse.app/zh/docs/glossary#context-window 模型单次请求能处理的 token 上限，人设、世界书、历史、你的消息都要装进这个窗口。窗口装不下时旧历史被截断或摘要，这是「聊久了忘事」的直接原因。 ### 思考预算 URL: https://foreverse.app/zh/docs/glossary#reasoning-budget 推理型模型在给出答案前产出的内部思考 token，按输出计费。给这类模型设输出上限时必须把思考预算算进去，否则正文可能一个字都不剩。 ### 分支（Branch） URL: https://foreverse.app/zh/docs/glossary#branch AI 续写的组织结构：每次续写作为独立分支挂在原文旁，原文永不被覆盖，多个走向并存可比较、可继续。类比代码的版本管理，reroll 是覆盖，分支是保留。 ### 剧场模式（VN Theater） URL: https://foreverse.app/zh/docs/glossary#vn-theater 把正在读的小说就地切换成视觉小说演出的阅读形态：台词逐字上屏、说话人有名牌、剧情分岔停下等你选，选择以分支写回原书。进出共用同一阅读进度，不产生新工程。 ### 说话人归因 URL: https://foreverse.app/zh/docs/glossary#speaker-attribution 判断一句引语出自哪个角色的技术问题。Foreverse 的立场是只认结构上有把握的证据、拿不准一律进旁白——错误归因比漏归因伤害大。 ### 搬家包 URL: https://foreverse.app/zh/docs/glossary#moving-box AI 伴侣数据的整包导出格式：人设、逐条记忆、贴纸、头像、聊天转录打成一个 zip，换设备导入还原。导入侧有白名单校验和失败回滚。 ======================================================================== ## FREE TOOLS — FAQs (English) ### AI Flavor Checkup — Free Rule-Based Prose Pattern Checker for Chinese Writing URL: https://foreverse.app/labs/ai-flavor Q: Is my text uploaded when I use this tool? A: No. The detector is a rule engine that runs entirely in your browser; nothing is sent to any server. Q: Why does it only support Chinese? A: The phrase blacklist and sentence-pattern rules were built and calibrated on Chinese roleplay-card and web-fiction prose. An English ruleset would need its own corpus and calibration, which we haven't done honestly yet. Q: Can a high score prove a text is human-written? A: No — and a low score doesn't prove machine authorship either. A low score means the text hits many patterns that readers commonly flag as machine-flavored and is worth reviewing item by item; templated AI copy with perfect structure still passes. Treat it as a lint pass, not a truth oracle. ### TXT Novel Checkup — Chapter Detection, Word Count, Encoding & Audiobook Length Online URL: https://foreverse.app/labs/book-checkup Q: Why did my 2-million-character book come out as one giant chapter? A: Chapter detection needs heading lines it can recognize — 第N章 family, numeric headings like "001 Title", Chapter N, prologue markers. If a book uses none of them (or a rare style), everything lands in one chapter, which breaks per-chapter audiobook segmentation and forces AI-continuation context to be cut by raw length. This page ports the core rules of a production importer; the app's full version covers more styles and supports manual splitting. Q: The preview shows garbage characters — what happened? A: The file's encoding is probably outside this page's detection chain. It covers UTF-8 (strict and slightly-damaged), GBK/GB18030, and UTF-16 with or without byte-order marks — the shapes that account for nearly all real novel files. Big5 and Western European single-byte encodings aren't detected here; re-save as UTF-8 in a desktop editor, or let the app's fuller chain handle it. Q: Why doesn't it take epub? A: epub is a zip container: unpacking, xhtml parsing, and spine-order chapter assembly. Doing that half-well in a browser would produce worse answers than not doing it — so this v1 takes txt only and says so. The app's importer accepts epub directly. Q: How are the reading and listening times calculated? A: Reading at 300 characters per minute (Chinese) or 220 words per minute (English books, detected by CJK share). Audiobook length at 250 characters or 150 words per minute for 1× speed — the same conversion our cost calculator uses — plus 1.25×, 1.5× and 2× rows. They're planning figures: your actual pace is yours. ### LLM Prompt Caching Calculator — What Cache Hits Save on Long Chats, at Real Price Spreads URL: https://foreverse.app/labs/cache-calc Q: Where does the default 89.9% hit rate come from? A: From our own measurements. In July 2026 we ran a four-way instrumented benchmark — same model, same key, byte-identical message scripts — and it embarrassed us first: stock SillyTavern hit 91.5% out of the box while we sat at 46.5%, because our lore entries were injected at a drifting position every turn. After pinning the injection position we re-measured 89.9% on fresh installs and 73.1% on upgraded ones. All four numbers are anchor buttons on the slider. Q: Why does the saving grow with more turns? A: Every turn resends the whole history, and the longer the history, the larger the share that is byte-identical to the previous request. At 10 turns the reusable prefix is roughly 60% of input; at 300 turns it exceeds 99%. Short chats barely notice caching — long chats are where it decides the bill. Q: Do I need to apply for the discount? A: No — providers grant it automatically. DeepSeek, Anthropic, Google and xAI all report hit tokens and discounted billing in the response. The one requirement is on the client: requests must keep a byte-stable prefix. Fail that and the discount silently never arrives, which is exactly what this page is about. Q: Can I treat the result as a bill estimate? A: As an order of magnitude, yes; as an invoice, no. Tokens are estimated; your real hit rate depends on client behavior (the 45-point spread between the anchor buttons is behavior, not luck); and providers differ on cache-write fees and cache lifetimes. The methodology section spells out every assumption. ### Card Brief Bag — Turn Rough Ideas Into an AI Character Card Task, No Writing Skill Needed URL: https://foreverse.app/labs/card-brief Q: Does this page write the card with AI? A: No — and that's deliberate. Every tool on this site runs locally with nothing uploaded, so the page itself never calls a model. The bag packs your material properly and hands it to something that does write: the butler agent in the Foreverse app (it carries the full card-writing playbook — archetype choice, lived-in character writing, greeting patterns, anti-slop lexicon), or any AI you already use, via the copy-ready task brief. Q: What happens when I hit “Send to my app”? A: The form is packed into a small versioned JSON (kind fv-card-brief), and pushed to your own account's cloud stash — the same channel our card tools use, after a light email sign-in. Your app shows a badge under “My Cloud”; open it and let the butler write the card from the material. Sending the identical brief twice never takes a second slot: the cloud matches by content and just resets the 30-day retention timer. Q: Does generating in the app cost money? A: Card writing runs on your app account's credits. Signing up grants 5,000 bonus credits, and having the butler write one card typically uses a small slice of that. There's no separate charge for the brief itself — the bag and the transfer are free. Q: What's in the copied task brief? A: One instruction block: your material laid out section by section, plus the craft rules our card playbook considers non-negotiable — concrete details over adjectives, a greeting that's a scene already in motion, never speaking for {{user}}, anti-slop phrase bans, example-dialogue turn-taking, creator-notes checklist. It asks the AI to output a chara_card_v3 JSON, which you can then drop into our Card Checkup or Card Workbench to verify. Q: How rough can the material be? A: Very. Every field is optional; one line like “a stubborn art restorer who owes me a favor” is a valid brief. You can also paste whole passages of fanfic or lore docs into the reference zone — they're used to learn flavor and facts, with an explicit instruction not to copy sentences. The form auto-saves to this browser so a long paste survives a refresh. ### Character Card Checker — Free Import Checkup for SillyTavern Cards URL: https://foreverse.app/labs/card-checkup Q: Is my card file uploaded when I check it? A: No. The parser runs entirely in your browser — PNG chunks, JSON, and charx zip entries are all decoded locally with browser APIs. Nothing is sent to any server, and the page keeps working offline once loaded. Q: What card formats does it support? A: Three containers: PNG card images (chara / ccv3 data in tEXt, iTXt or zTXt chunks, with ccv3 preferred like SillyTavern does), plain JSON (V1 flat legacy cards, V2, and V3), and .charx zip containers (the card.json inside is read in the browser). Standalone lorebook JSON and group-chat exports are recognized and reported honestly as not-a-card. Q: Why is the token count different from other tools? A: It's a character-ratio estimate, not a real tokenizer: CJK at roughly 1.6 characters per token and English at roughly 4 characters per token — a common ballpark for DeepSeek and GPT-family models. Real tokenizers differ by model and can deviate 20–30%. Read the numbers as magnitudes, not exact counts. Q: What is a .charx file? A: The zip container defined by the chara_card_v3 spec: a card.json at the root plus optional assets (avatar, sprites, backgrounds) referenced via embedded paths. This tool reads the card.json for the checkup; asset unpacking is an import-time job for a full frontend. One web-side limit: zip64 archives (4GB+ entry tables, rare for real cards) aren't supported here — the tool says so honestly instead of guessing, and the Foreverse app imports them fine. ### Character Card Converter — v1/v2 up to chara_card_v3, v3 down to v2, Losses Listed Honestly URL: https://foreverse.app/labs/card-convert Q: What actually differs between card v2 and v3? A: V3 (chara_card_v3) keeps the whole V2 structure and adds: nickname (substitutes {{char}} in prompts), group_only_greetings, an assets list for avatars and sprites, multilingual creator notes, source links, timestamps, lorebook decorator syntax (@@activate and friends), and the .charx zip container. If a card uses none of those, v2 and v3 are the same card in a different wrapper. Q: What do I lose downgrading v3 to v2? A: Only the V3-only fields the card actually uses — this page enumerates them per card before you download. Typical losses: asset references, nickname, group-only greetings, multilingual notes, source and timestamps. The persona text, greetings, example dialogue and the embedded lorebook all survive; decorator lines inside lorebook entries stay as text but V2 frontends read them as plain words. Q: Does converting change my card's content? A: No. The converter moves structure, not prose: fields are re-filed between wrappers byte-for-byte. The two exceptions are mechanical and reported on screen — an empty V1 name defaults to "Unknown" (importer convention), and a V1 top-level talkativeness moves into extensions where V2 expects it. Q: My card has a data object but no spec field — what is it? A: A semi-spec card, common from community exporters and hand edits. Most frontends read it as V2; strict validators reject it. Running it through this converter with V2 as the target normalizes it: same content, plus the spec and spec_version fields tools expect. Q: Why is there no “convert to V1”? A: By design. V1 is the flat 2018–2022 format with six fields — exporting to it would drop the lorebook, alternate greetings, creator notes, system prompt and everything else V2 introduced. And no maintained frontend today is V1-only: anything that reads V1 also reads V2. Downgrading to a format that keeps almost nothing isn't conversion, it's amputation — so V1 goes in, but never out. ### Character Card Diff — See Exactly What a Card Update Changed URL: https://foreverse.app/labs/card-diff Q: What exactly does it compare? A: Three layers. Field level: every card text field (description, personality, greeting, examples, system prompt…) gets a word-level diff with deletions and insertions highlighted. Lorebook level: entries are matched across versions and grouped into added / removed / modified / unchanged, with per-entry change chips (keys, content, constant/enabled flips, renames). Token level: the same estimate ruler as our Card Checkup, old → new → delta per field. Q: It says “no visible changes” but the files aren't identical. What gives? A: That's the honest three-state verdict doing its job. If the raw JSON is byte-identical, it says so. If only key order or whitespace differ (an editor re-serialized the file), it says the content is the same. And if the raw JSON differs somewhere this tool doesn't display — usually extensions: embedded regex scripts, beautify payloads, frontend-private fields — it tells you a change is hiding there rather than pretending nothing happened. Q: How are lorebook entries matched between versions? A: Five passes, strictest first: full identity, identical content, unique matching name, unique matching keyword signature, then content similarity — with a stricter threshold for very short entries, because two five-word entries sharing three words shouldn't be declared “the same entry, modified”. Anything that survives no pass is honestly reported as added or removed, not force-paired. Q: Are my cards uploaded? A: No. Both files are parsed and diffed entirely in your browser. Nothing is sent anywhere, and the page works offline once loaded. ### Character Card Poster Maker — Turn Your Card into a Share-Ready Promo Image URL: https://foreverse.app/labs/card-poster Q: Which parts of my card does the poster use? A: Only the public shelf face: name (nickname as fallback), creator, tags, character version, and the creator notes for the one-line pitch. The definition body — description, personality, scenario, greetings, example dialogues, system prompt, lorebook — is never read. That's not a promise buried in a privacy page: the field builder's property access is tracked by a Proxy in our golden self-test, and touching anything outside the whitelist turns the build red. Cards with hidden definitions stay hidden. Q: Where do the colors come from? A: From the card art. The image is downsampled, quantized, and scored by pixel count times saturation — so a colorful costume beats a gray backdrop. The dominant color seeds three palette variants (deep, ivory, contrast), and if the extractor found several strong candidates you can cycle through them. Text ink is chosen automatically so it always clears a 4.5:1 contrast ratio against the background — that guarantee is a locked test, not a vibe. Q: What is the QR placeholder for? A: A clean white square reserved in the layout, positioned so it never collides with the text zones (also a locked geometry test). Paste your own QR over it after export — a Discord invite, your Chub profile, a Foreverse redeem code. We deliberately don't generate QR codes: the poster stays fully local, and you stay in control of where the link points. Q: My JSON card has no art — what happens? A: PNG cards bring their own portrait; JSON and charx cards get a generated monogram tile in the card's palette so the layout still works, and you can upload any image to replace it. Uploaded art stays in your browser like everything else. Q: Is my card uploaded? A: No. Parsing, color extraction, layout and PNG export all run in your browser. Refresh and everything is gone. ### Character Card Shelf — Batch-Import PNG Cards & Browse Lorebooks in Your Browser URL: https://foreverse.app/labs/card-shelf Q: Are my cards uploaded when I add them to the shelf? A: No. Parsing happens in your browser (PNG text chunks, JSON, and charx zip entries are all decoded locally), and the shelf itself is stored in your browser's IndexedDB — a private, per-site storage area on your own machine. Nothing is sent to any server, ever. Clearing your browser's site data also clears the shelf. Q: What formats can I import, and can I mix them in one batch? A: PNG card images (chara / ccv3 chunks, ccv3 preferred — same order as SillyTavern), plain JSON (V1 flat legacy, V2, V3), and .charx zip containers. Mixed batches are fine: each file is parsed independently, failures are listed with their reason, and one broken card never blocks the rest. Q: Will my shelf survive closing the browser? A: Yes — cards live in IndexedDB, so they're still there after a refresh, a restart, or a week away. Two honest caveats: private/incognito windows usually wipe IndexedDB when they close (the tool warns you and falls back to session-only mode), and browsers may evict site data under storage pressure. The shelf shows its current storage usage so you're never guessing. Q: Where do the cover images come from? A: PNG cards are their own covers — the embedded artwork is shown directly. For .charx containers the first bundled image asset (icon or avatar preferred) is extracted locally. Plain JSON cards have no image inside, so they get a generated letterpress cover; remote avatar URLs are deliberately not fetched, to keep the zero-network promise honest. Q: Does it show everything that's inside the lorebook entries? A: Each embedded lorebook entry is shown as its own card: primary and secondary keys, full content, plus badges for constant, insertion position (including @depth), order, trigger probability, secondary-key logic (AND/NOT), regex keys, sticky / cooldown / delay timers, and recursion flags. Entry search filters by key, name, or content. For firing simulations, use the Lore Trigger tool; for cross-platform migration checks, the Lorebook Doctor. ### Chat Skin Previewer & Packer — 11 Themed Beautify Packs for SillyTavern Cards, Try On Live URL: https://foreverse.app/labs/card-skins Q: What exactly is a “chat skin” here? A: A set of 5 regex scripts sharing one cultural motif. The character card instructs the model to reply in a plain-text marker protocol (:::think reasoning :::, a [STATUS|…] line, :::scene title # prose :::, :::note aside :::, and closing :::opt A ¦ B ¦ C :::). Each script catches one marker and re-renders it as themed HTML — a fold-out thought card, a styled status panel, a scene sheet, tappable choice chips. Same protocol, eleven completely different looks: swapping skins means swapping one pack. Q: Will these packs work in SillyTavern, or only in the Foreverse app? A: Both, with one honest caveat. The zip carries 5 single-script JSONs in SillyTavern's Regex-panel format (placement AI output, markdownOnly on — display layer only, your outgoing prompt stays clean). The tavern's web build renders the inline-styled HTML fine; decorative art references an app-only asset channel (fv.local) and hides itself cleanly elsewhere — layouts were designed to survive without it. Clients that don't render HTML in messages (raw API frontends) will just show the markers as text. Q: Why doesn't my chat look skinned after importing? A: The pack is display-layer only — it has nothing to render until the model actually outputs the marker protocol. That's what the format directive on this page is for: paste it at the end of the card's system_prompt, put the reminder line into post_history_instructions, and ideally let the card's greeting demo one full marker frame (few-shot is the strongest format anchor). Also check that only one skin's 5 scripts are enabled: two skins race for the same markers and the first match wins. Q: Is my pasted text uploaded anywhere? A: No. The whole pipeline — regex matching, HTML templating, the phone preview, the zip packer — runs in your browser; this page has no upload endpoint. The preview also runs inside a fully sandboxed frame with scripts disabled, and your text is HTML-escaped before any template touches it, so pasting hostile content can't execute anything. ### PNG Character Card Slimmer — Compress the Art, Keep the Card Data (Verified Byte-for-Byte) URL: https://foreverse.app/labs/card-slimmer Q: Why can't I just run my card through TinyPNG / Squoosh / an image compressor? A: Because they're built for images, not cards. Re-encoding a PNG rebuilds it from pixels and throws away the text chunks (tEXt / iTXt / zTXt) — and that's exactly where the chara / ccv3 card data lives. The picture survives, the character dies. This tool re-encodes only the artwork, then splices the original card chunks back in, byte-for-byte untouched. Q: How do I know the card data really survived? A: It's verified, not promised: before the download button appears, the tool extracts the card chunks from the output and compares them against the original byte by byte. If they don't match exactly, you get an error instead of a download. The check runs on every single compression. Q: What does the quality tier actually do? PNG has no “quality” slider. A: Correct — the canvas PNG encoder has no quality knob (that only exists for JPEG/WebP), and a card must stay a PNG because the data chunks can only live there. So the quality tiers are palette quantization: the artwork is reduced to 256 / 128 / 64 colors with Floyd–Steinberg dithering and written as a real indexed-color PNG. On character art the 256-color tier is near-invisible and typically halves the file. Q: Why did my file get bigger instead of smaller? A: Two honest reasons: the original was already crushed by a pngquant-class optimizer (a browser can't beat that at the same settings), or it's already an indexed-color image and the lossless tier expanded it back to full color. The tool tells you when this happens instead of hiding it — try a smaller size tier, a harder quality tier, or accept that this particular card doesn't need slimming. Q: Is my card uploaded? A: No. Decoding, quantization, chunk splicing and verification all run in your browser with local APIs. Nothing is sent anywhere, and the page works offline once loaded. ### Card Slop Doctor — Measure How Much of Your Character Card Came Off the Shelf URL: https://foreverse.app/labs/card-slop-doctor Q: What exactly is “slop” in a character card? A: Slop is the community's word for AI prose assembled from stock parts: shivers down spines, voices barely above a whisper, eyes sparkling with mischief, a mix of amusement and concern. None of these phrases is wrong on its own — humans write them too. What marks a card as slop is accumulation: when the greeting shops every shelf of the same lexicon, the card reads as generated-and-kept, and worse, the model treats those phrases as the house style and restocks them in every reply. This page measures that accumulation instead of arguing about taste. Q: Is my card uploaded when I check it? A: No. The parser and the lexicon both run in your browser — PNG chunks, JSON and charx zips are decoded locally, and the scan is a set of regular-expression passes. Nothing is sent to any server, and the page keeps working offline once loaded. Q: My favorite card has two or three of these phrases. Is it slop? A: Almost certainly not — and the engine agrees. A single “took a deep breath” appears in every novel ever printed, which is exactly why one hit never changes the verdict: the grade is density-based (hits per thousand words) with an absolute floor, so a 120-word greeting with one stock phrase stays clean. The verdict moves only when the card keeps shopping the same shelves — that pattern is a fingerprint, not a coincidence. Q: The name rule flagged my character. Do I really have to rename her? A: You don't have to do anything — but know what the name costs. Corpus work on model-generated fiction found “Elara” appearing at tens of thousands of times its human-baseline frequency, with Seraphina, Kael and friends close behind. Communities have learned the pattern: an Elara card gets read as “asked a model, kept the default” before anyone reaches your actual writing. A rename is one find-and-replace and removes the loudest single tell on the card. Names the rule stays quiet on — real-world common names like Aria or Evelyn — are deliberately exempt. ### Character Card Editor — Create, Edit, Check and Export Cards in Your Browser (Card Studio) URL: https://foreverse.app/labs/card-studio Q: Is my card uploaded while I edit it? A: No. Parsing, editing, the live checkup, macro preview and both exports all run in your browser — nothing is sent to any server. Drafts auto-save to this browser's local storage on your machine, not to an account. Q: What goes in, and what comes out? A: In: PNG card images, JSON (V1 flat legacy, V2, V3), and .charx containers — V1 cards are promoted to the V2 structure on import, the standard behavior across modern frontends. Out: v2 or v3 JSON, or a PNG card built from the original image, an image you upload (JPG/WebP transcoded locally), or a generated placeholder. Repacking .charx isn't supported — the web can't touch its bundled assets. Q: How is this different from the card checkup tool? A: Same rules, different posture. The checkup is read-only: drop a file, get a report. The studio runs those rules continuously while you edit — fix a dead lorebook entry and the red item disappears as you type. If you only want a verdict on a downloaded card, the checkup page is faster. Q: When should I use the Card Workbench Pro instead? A: The studio is the light tier: quick edits to one card, structural checkup plus an impersonation / AI-flavor diagnostics summary in the side rail. The Workbench is the writing bench — multi-draft autosave in IndexedDB (megabyte cards fit), inline mark highlighting inside each field, system_prompt and post-history editing, and multi-member group-chat pack assembly. Fixing a downloaded card: stay here. Writing cards from scratch, seriously: use the Workbench. Q: Do fields the editor doesn't cover get lost? A: No. Everything the editor doesn't touch — extensions, embedded regex scripts, V3 assets, nicknames, insertion orders, even non-standard keys other tools wrote — is carried into the export byte-for-byte. The 'carried along' panel lists what your card has. The one deliberate exception: downgrading v3 to v2 removes v3-only fields, and the studio lists exactly what gets dropped and asks first. Q: Where do drafts live, and can I lose work? A: Edits auto-save to your browser's localStorage about half a second after you stop typing, and the page offers to restore the draft next time you open it. Two honest limits: drafts don't follow you across browsers or devices, and a card too large for localStorage (rare, several MB of text) can't auto-save — the toolbar says so instead of pretending. ### Card Workbench Pro — a Character Card Editor With a Live Diagnostics Rail (Impersonation, Slop, Budget) URL: https://foreverse.app/labs/card-workbench Q: How is this different from the Card Studio editor? A: Card Studio edits structure and runs the structural checkup (dead lorebook entries, broken macros, token thresholds). The Workbench keeps all of that — same parser, same lorebook editor — and adds what no form editor has: a live content-diagnostics rail. While you type it runs impersonation detection (does your card speak for {{user}}?), AI-slop scoring for English prose, the zh AI-flavor ruleset for Chinese prose, per-field token budget bars, and craft guidance from real card-writing playbooks. It also edits system_prompt / post_history_instructions, assembles group-chat packs, and auto-saves multiple drafts to IndexedDB, so multi-megabyte cards fit. Q: Is anything uploaded while I edit? A: No. Parsing, every diagnostic engine, macro preview, PNG embedding and zip packing all run in your browser. Drafts save to this browser's IndexedDB on your machine — not to an account, not to a server. Q: What do the inline highlights mean? A: Each field has a Marks tab: red marks are impersonation hard hits (the card speaks, acts or feels for the user), yellow are weak signals (scene-teleport openers, perception hijacks), purple are stock slop phrases from the English lexicon, orange are Chinese cliché hits. Hover any mark for the rule name. The rules are the same engines behind our standalone impersonation checker and slop doctor — not a second opinion. Q: What exactly does the group-chat assembly export? A: A payload zip in the exact shape the Foreverse community group-chat pack (group_chat) expects: payload/group_chat.json with version/members/strategy/director, one card JSON per member under payload/members/, optional compressed avatars under assets/avatars/. Honest boundary: this is the payload layer, not a publishable .fvpack — the manifest, content hash and review happen in the app's author center. The workbench mirrors the production validator's limits (2–12 members, roster names must match in-card names, director under 2,400 code points), so what you export passes that gate without rework. Q: Are the token counts exact? A: They're character-ratio estimates, the same estimator used across our card tools — good for spotting which field is overweight, not a tokenizer. Real models differ by 20–30%. The English slop lexicon only judges English prose and the zh flavor ruleset only scores mostly-Chinese text; when a card is out of a ruleset's calibrated range, the rail says so instead of faking a clean score. ### Chat Archaeology — Turn a SillyTavern Chat Export into a Relationship Report URL: https://foreverse.app/labs/chat-archaeology Q: Is my chat log uploaded anywhere? A: No. This page has no upload endpoint — parsing, statistics and the share-card rendering all run in your browser, and refreshing clears everything. Chat exports are among the most private files an RP player owns, which is exactly why the tool was built local-only. The share card is also structurally incapable of carrying chat text: its data type only has number and label fields. Q: Is the cost figure what I actually spent? A: No, and the page says so next to the number. It's a replay estimate: every AI reply (including every swipe you rerolled) is priced as one request at the 2026-07-24 list-price snapshot, with the full history refilled each turn up to the context window you pick. It excludes your character card, system prompt and lorebook (so real usage is higher), ignores cache discounts (DeepSeek cache hits cost ~1/50th of list, so real spend can be far lower), and rough-counts tokens instead of using each provider's tokenizer. If your export records a model name per message — newer SillyTavern builds do — the report prices by that tier; otherwise you get a floor-to-ceiling range. Q: What files does it accept? A: SillyTavern chat exports: .jsonl files with one JSON object per line and an optional metadata line at the top. Foreverse exports the same shape. Missing fields degrade gracefully — no timestamps means the clock section steps aside, corrupted lines are skipped and counted. You can drop up to 8 files at once and the report adds a cross-character comparison table. Files up to 48MB parse in chunks so the page stays responsive. Q: How is the AI-flavor curve computed? A: Each AI reply is scored 0–5 by the same rule engine as our AI-flavor checkup (v10.7 lightweight port): a stock-phrase blacklist at low weight and sentence-pattern molds at high weight, calibrated on Chinese roleplay prose for zero false kills on human-written hit cards. The curve shows the score per floor and marks where a sustained drop begins — the “it gets robotic after floor N” feeling, measured. It's a lint pass, not a detector, and it's honest about only supporting Chinese: mostly non-Chinese logs get a disclaimer instead of a forced verdict. ### Chat Log Converter — Turn a Tavern .jsonl into Clean Markdown or a Readable HTML Page URL: https://foreverse.app/labs/chat-export Q: What exactly does it do with hidden floors and swipes? A: Hidden floors (is_system messages — notes and narrator asides that never went to the model) are filtered out of the transcript, and the document's header line honestly reports how many were omitted. Messages with swipe candidates output only the currently selected variant — the takes you rerolled away stay gone. Both rules come from the same parser that powers our Chat Archaeology tool, so the numbers always agree. Q: Markdown or HTML — which should I pick? A: Markdown if the text is going somewhere else: blogs, forums, note apps, a fic draft. Your *action* asterisks stay as written (they're already valid Markdown), and the only thing we escape is a line-leading # so a message can't hijack your document structure. HTML if the file itself is the archive: it's a standalone page with book-like typography, actions in muted italics, dialogue tinted, dark mode following the system — and zero scripts, so it's safe to keep or share as-is. Q: How is this different from the Chat Shot tool? A: Chat Shot renders a picture — bubbles, avatars, themes — for showing a scene on social media. This tool outputs text: the whole log (or a slice) as a document you can archive, quote, search or paste into a blog. Same parser underneath; different destination. If you want one image of the best ten messages, use Chat Shot; if you want the transcript itself, you're in the right place. Q: Does my chat get uploaded? A: No. Parsing, conversion, preview and download all run in your browser — this page has no upload endpoint. The HTML preview renders in a fully sandboxed frame. Refresh and everything is gone. ### Chat Log to Image — Turn a Roleplay Export into One Clean Share Picture URL: https://foreverse.app/labs/chat-shot Q: What chat formats does it accept? A: Two: SillyTavern chat exports (.jsonl — one JSON object per line, metadata on line 0; Foreverse exports the same format), and plain text where each turn starts with "Name: text" on its own line. Lines without a name prefix join the previous message, so multi-paragraph turns paste fine. After parsing you pick which speaker is you, and which slice of messages to show. Q: Does my chat get uploaded anywhere? A: No. Parsing, layout and rendering all run in your browser — this page has no upload endpoint, and the export is written straight from a local canvas to your downloads. Refresh and everything is gone. That's also why a huge log renders fast: no bytes cross the network. Q: Why did my long chat come out as several images in a zip? A: Browsers cap canvas height (commonly 16,384px — beyond it, drawing silently produces a blank). The tool keeps every page safely under that cap, cuts at message boundaries so no bubble is split mid-thought unless a single message is taller than a whole page, numbers each page, and packs them into one zip. Short chats export as a single PNG. Q: How are *actions* and "dialogue" colored? A: Following tavern rendering conventions: text between asterisks becomes muted italics (actions), quoted speech — straight, curly or CJK corner quotes — gets the accent color, and plain narration stays in the body color. Every theme follows the same rules; the custom theme only swaps the palette, not the logic. ### LLM Context Window Visualizer — See Exactly Why Your AI Chat Forgets URL: https://foreverse.app/labs/context-lens Q: So why does the AI forget things from earlier in the chat? A: Because every request re-sends the conversation from scratch, and the model only accepts a fixed budget of tokens. Once the system prompt, card and lorebook take their share, history fills whatever's left — newest first. Turn 12 where you said your name isn't "forgotten" so much as physically absent from the request. The grid on this page shows exactly which turns made the cut. Q: Does a bigger context window fix it? A: It delays it and bills you for the delay. A 128k window holds a lot of turns, but you pay input price on the whole window every single turn — that's the rising line in the cost chart. And attention quality degrades over very long contexts, so "it's in the window" doesn't guarantee "it gets used". Compaction plus caching is the cheaper answer for chats that run hundreds of turns. Q: What exactly does the compaction summary rescue? A: The gist, not the words. In our compaction benchmark (7 configs × 98 probes on a 530k-token corpus), a structured summary lifted recall over the compressed zone from 24% (plain truncation) to 59% — and the single biggest factor was whether the summary kept concrete values like names and numbers. Exact quotes and fine detail don't survive; that's why the app never rewrites the original transcript on disk. Q: Where do the 91.5% / 89.9% / 46.5% hit rates come from? A: From an instrumented benchmark we ran in 2026-07: same model, same key, byte-identical message scripts, every call metered through one gateway. Bare SillyTavern's default layout hit 91.5%; our own app hit only 46.5% because a lore-injection spot drifted every turn — we published the failure, fixed it, and re-measured at 89.9%. The full report is linked in the methods section. ### AI Reading Cost Calculator — Audiobook, Continuation & Roleplay at Real API Prices URL: https://foreverse.app/labs/cost-calc Q: Where do these prices come from? A: From a snapshot of public list prices on models.dev (capture date shown on the page, currently 2026-07-24) — the same source our billing backend syncs from. Two bases are shown: the official channel adds a 50% service margin on top of list price (1 credit = $0.0001); bring-your-own-key runs at list price with no markup. Promotions, negotiated rates, and regional pricing are excluded; providers' own pages are authoritative. Q: How is “one continuation” estimated? A: As roughly 8,000 tokens of context in and 550 tokens out (about 300 words of new prose). That calibration matches the app's measured billing: one continuation on DeepSeek V4 Flash costs about 19 credits on the official channel. Tokenizers differ between models, so expect around ±15% — treat results as an order of magnitude, not an invoice. Q: Is listening to a whole book really free? A: The system-TTS tier is genuinely free: it synthesizes on your phone, no API involved, so a 2M-character novel costs nothing. Online AI voices bill per character, and a full 2M-character book runs to tens of dollars — which is why they suit close-listening a single chapter (under $0.15) rather than converting a whole shelf. That math is exactly why the app defaults audiobooks to system TTS. Q: What do the 5,000 sign-up credits buy? A: 5,000 credits equal $0.50. On the DeepSeek V4 Flash basis that's about 260 continuations — 80k to 130k characters of new prose — or many evenings of roleplay chat. The official channel needs no key setup; when the grant runs out you can buy credits or switch to your own keys. ### TXT to EPUB Converter — Free, In Your Browser, Nothing Uploads URL: https://foreverse.app/labs/epub-maker Q: Will the EPUB work on Kindle? A: Yes. Since 2022, Send to Kindle accepts EPUB directly — by email or from the share sheet. The output carries both the EPUB 3 nav and the legacy NCX table of contents, so chapter jumping works on older readers too. Only very old Kindle firmware needs azw3; for those, run this EPUB through Calibre once. Q: What if chapter detection gets it wrong? A: Every detected chapter is listed before packaging: rename any title inline, untick rows that are ads or site headers to exclude them. If no chapter style is recognized at all, the text is split evenly every ~8,000 characters and the page says so plainly — it never pretends detection succeeded. Merging chapters or inserting split points by hand happens in the app's importer, not here. Q: Does my manuscript get uploaded? A: No. Parsing, chapter detection and zip packaging all run inside your browser; this page has no upload endpoint. That is also why it can be free with no file limit up to 64MB — the work happens on your device, so we have no server bill to pass on. Q: Can I convert a book I downloaded from a website? A: The tool is format-neutral, but make sure the source permits you to save and convert the text. This page is designed for your own writing: manuscripts, compiled notes, and texts you have the rights to. Q: Why is there no image support? A: v1 is plain-text only. A txt file can't carry images, so picture links or placeholders in the body stay as ordinary text. Illustrated EPUB assembly (embedded images per chapter) is out of scope for this version. ### SillyTavern Expression Pack Workshop — Validate, Rename to the 28 Slots, Zip for Import URL: https://foreverse.app/labs/expression-pack Q: Where does the list of 28 expression names come from? A: From SillyTavern itself: the built-in local classifier is a GoEmotions model (27 emotions plus neutral), and the official docs state sprites must be named after those labels — admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, neutral, optimism, pride, realization, relief, remorse, sadness, surprise. Names like happy.png or blush.png silently miss their slot — that's exactly the mistake this tool catches and fixes. Q: Do I need all 28 images? A: No. ST simply leaves unmatched slots empty, and the coverage meter here says so instead of nagging. The one you really want is neutral — it's the sprite shown when no emotion fires. If you have several images for one emotion, the tool names them joy.png, joy-1.png… following the official multi-image suffix rule. Q: Why does my zip import but show no sprites in ST? A: Almost always one of two things: the images sit inside a subfolder in the zip (ST's sprite-pack import only reads the top level — this tool always packs flat), or the filenames don't match the 28 labels exactly (lowercase, exact spelling). Both are validated here before the download button produces anything. Q: Does this generate expression art for me? A: No, and that's deliberate. This is a packaging workshop for art you already have — your own drawings, commissioned sets, game rips you hold rights to. It validates, renames, previews and zips; it never invents pixels, and nothing is uploaded anywhere. Q: What counts as a transparency problem? A: Sprites are pasted over the chat background, so the background of the image itself must be alpha-transparent. The tool decodes each image and checks: JPG files can't carry transparency at all (flagged), PNGs with zero transparent pixels get flagged as opaque, and images whose corners aren't clear get a gentler warning — then you can see the actual effect in the fitting room. ### Fan-Card Maker — Frame a Character as a Collectible Card or Bookmark URL: https://foreverse.app/labs/goods-studio Q: How is rarity decided? A: By the card's real community record, not a dice roll: score = imports ×1 + likes ×2 + favorites ×2, from a published snapshot of the community hub. Tiers follow the snapshot's actual distribution — UR needs a score of 4+ (the top 3% of the 96 cards on the shelf), SSR 2+, SR 1+, R is zero. The thresholds and snapshot date are printed next to the badge. Q: Why is my uploaded image "unrated"? A: Because there's nothing to rate it with. Rarity here means community standing — imports and likes a published card earned. Your own image has no such record, and inventing a tier would make every UR meaningless. The frame still renders in a neutral theme. Q: Does my image get uploaded anywhere? A: No. The canvas is drawn entirely in your browser and the PNG is assembled locally — the image never leaves your machine. That's also why the tool asks you to confirm you have the rights to use it: we can't check, so you carry that. Q: Why does the export carry a watermark? A: A small "name · foreverse.app" line bottom-left is the price of the free web version — every card that travels is a signed postcard. The in-app Goods Studio can switch it off, and it rates rarity from your actual relationship with the character (chat turns, affection) instead of public stats. ### Greeting Preview — See Your first_mes as a Real Chat Bubble Before You Publish URL: https://foreverse.app/labs/greeting-preview Q: What's the difference between immersive and IM mode? A: Immersive is the tavern-style block render: one message, one bubble — narration, actions and dialogue together. IM mode splits the same message into messenger-style bursts: split on line breaks, then by sentence enders if a single segment runs long, capped at 8 bubbles (overflow merges into the last one). Check both before publishing — a long scenery paragraph becomes a very different rhythm as a string of short texts. Q: How is the bubble count computed? A: Same rules as the app's IM engine: split on newlines and drop blanks; if only one segment comes out and it's long (60 chars for CJK, 160 for Latin-dominant text), split again on sentence enders — with a false-positive guard for English periods, so Mr., 3.14 and J. K. don't get chopped; past 8 segments, the first 7 stay and the rest merge into the last. If this page says 6 bubbles, the app plays 6 bubbles. Q: How long is too long for a greeting? A: Past roughly 3,000 tokens you get a yellow flag. A greeting isn't a one-time cost — tavern-family frontends may carry it on every request, so an oversized opener taxes the memory budget of every turn. If you need that much lore, move the worldbuilding into a lorebook that triggers on demand, and keep the greeting to the scene and the first line. Q: Can it preview beautified-card HTML? A: No. This previews the text rendering layer: macro expansion, action italics, quote tinting, code fences, IM splitting. Beautified cards turn markers into HTML via regex scripts — that layer needs the app. ### Image to Editable PPTX — Rebuild Slides from Screenshots, in Your Browser URL: https://foreverse.app/labs/image-to-pptx Q: Why do other image-to-PPT tools give me slides I can't edit? A: Because most of them don't convert anything — they paste your image onto a blank slide as one flat picture. The page count matches, nothing is editable. Rebuilding is the hard part: you need OCR to find the text, layout analysis to group it into blocks, background repair to erase the original pixels, and a PPTX writer to emit real text boxes. That whole pipeline runs here, in your browser, which is also why nobody can charge you per file for it. Q: What kind of image works best? A: Slide screenshots, web page captures, posters and infographic-style cards — anything with standard fonts on flat or gradient backgrounds. On our calibration set, text on plain backgrounds erases near-invisibly and font sizes land within about 10% of the original. Photos are the honest weak spot: text over busy textures is left un-erased (the text box sits on top), and low-contrast captions — white text on a bright sky — can be missed entirely. Q: The OCR got a word wrong. Now what? A: Click the block and fix it before you export — that's the point of rebuilding instead of pasting. A flat image with a typo is a do-over; a text box with a typo is a two-second edit. Every block is editable right on this page: text, size, color, bold. Q: Does my image get uploaded? A: No. The OCR engine (tesseract.js, Apache-2.0) downloads into your browser once — about 6–10 MB, cached after — and every step from recognition to the final PPTX file happens on your machine. This page has no upload endpoint. That's also why it's free without limits: processing you run yourself costs us nothing. Q: Can I export a layered PSD? A: Not yet. PSD export (background layer plus real text layers, via the MIT-licensed ag-psd writer) shares this engine and is planned as the second output. PPTX shipped first because an editable deck is what most people searching for this actually want. ### Impersonation Checker — Find Out Why Your Bot Speaks and Acts for You URL: https://foreverse.app/labs/impersonation-checker Q: Why does my bot keep speaking and acting as my character? A: Because language models continue patterns, and the pattern is usually planted in the card itself. If the greeting narrates what “you” do, if the example dialogue writes your side of the conversation, or if the description says who {{user}} is and what they secretly feel — the model reads all of that as a demonstration that this is a both-sides story. JanitorAI's own help center calls impersonation one of the most common and frustrating issues, and the guides that actually fix it all start in the same place: audit the card. That's what this page automates. Q: How do I stop the bot from speaking for me? A: Fix the card before you fight the model. ① Run the card through this checker and rewrite every flagged pattern — greeting sentences with “you” in the subject seat, example dialogue that writes your lines, description text that scripts the player. ② Edit the impersonated lines out of existing replies, or the transcript keeps teaching the behavior back. ③ Keep your persona to physical facts and write replies with some substance — a one-line answer after a long bot message is a vacuum the model fills by playing you. Skip the tempting “never speak for {{user}}” instruction: negative phrasing hands the model the very pattern you're banning, and community testing keeps finding it backfires. Q: Is my card uploaded when I check it? A: No. The parser and the detection engine both run in your browser — PNG chunks, JSON and charx zips are decoded locally, and the scan is a set of regular-expression passes over the text. Nothing is sent to any server, and the page keeps working offline once loaded. Cards are private things; that's the point. Q: Is second-person narration (“you”) always bad? A: No — and the checker is built around that distinction. “She watches you from across the room” is fine: you are the object, the character is doing the observing. “You wake up in a cell” is a legitimate RPG opening — it transports your senses, once. What breaks roleplay is “you” in the subject seat with an act/speak/feel verb: “you decide”, “you stammer”, “you feel drawn to him”. The engine checks the grammar seat, not the word. Q: The card is clean but the bot still impersonates me. What now? A: Then the card isn't the culprit, and this tool honestly can't see the rest. Three places to look next: your persona field (trait-stuffed personas force the model to perform you — keep it to physical facts), your reply length (one-line replies after long bot messages leave a vacuum the model fills by writing you), and the model itself (some models impersonate more; neutralizing samplers or switching models fixes stubborn cases). Also check any “never speak for {{user}}” instruction — negative phrasing is a coin flip and can make it worse. ### JanitorAI ⇄ Character Card Converter — Bot Text to chara_card_v2/v3, Both Ways URL: https://foreverse.app/labs/janitor-card-convert Q: JanitorAI has no export button — what do I actually paste in? A: Whatever you can see. For your own bots, open the creation page and copy each box (Personality, Scenario, Initial Message, Example Dialogs) — or paste the whole Personality blob into the big box here and the / / containers get sorted automatically. For other people's private-definition bots, the definition is hidden by design and no converter can conjure it. Q: Why does JanitorAI's Personality become the card's description field instead of personality? A: Because that's where the body text lives on each side. On JanitorAI, the Personality box officially carries "the majority of the bot's information" — traits, appearance, backstory, even example dialogue. On tavern cards that main-body role belongs to description; the tavern personality field is a short summary slot. Mapping body to body keeps the prompt weight where the author put it. Q: Where does the 1,500-token health check come from? A: From JanitorAI's own help center: their token guide says well-made bots perform best under 1,500 permanent tokens and that past 2,000 you're "skating on thin ice"; the creation-page overview cites 2,500 as the ceiling. This tool weighs Personality + Scenario (the creator-controlled permanent block) with the same character-ratio estimate the rest of this site uses, and says clearly that JanitorAI's own tokenizer will differ. Q: What happens to the lorebook when I convert to JanitorAI format? A: It's listed as a loss, not silently dropped. The JanitorAI creation page has no lorebook box — world info is a separate feature there, with real dialect traps (imported entries famously get stuck at 1% probability). Run the book through the Lorebook Migration Doctor before moving it. Q: Is anything uploaded? A: No. Parsing, splitting, token estimates and PNG embedding all run in your browser. The page works offline once loaded. ### SillyTavern Lorebook Trigger Simulator — Test Why World Info Entries Fire (or Don't) URL: https://foreverse.app/labs/lore-trigger Q: How do I test if my lorebook is working? A: Without this page: start a throwaway chat, say a sentence containing one entry's trigger key, and check the prompt log for the entry's text — repeat per entry, burning a message each time. With this page: paste the lorebook JSON and a few chat lines, press one button, and every entry reports fired / silent with the exact reason — matched keys highlighted, secondary-key logic spelled out, recursion chains drawn. Same SillyTavern 1.18 semantics, zero test chats. Q: Why isn't my lorebook entry triggering? A: The four usual suspects, all flagged directly by this tool: ① the keyword never appears in the scan window — by default only the last 4 messages are scanned, so a keyword from ten turns ago doesn't count; ② the secondary-key condition fails — AND_ALL wants every key present, NOT_ANY wants none of them, and people mix the directions up constantly; ③ the entry is disabled, or it's a dead entry with no keys at all; ④ the probability roll missed — a 30% entry sitting out seven rounds in ten is working as designed. Q: What does recursive activation mean? A: When entry A fires, A's own content joins the next scan pass — if B's keyword appears inside A's text, B fires too. That's recursion, and lorebook authors use it deliberately to build lore chains: mention the lighthouse, get the keeper, get the keeper's secret. This tool draws the chain ("triggered by the content of …") — information you can't see inside SillyTavern itself. Q: Will the simulation match a real chat exactly? A: The scan-and-activate verdicts match (keyword matching, case, whole-words, secondary logic, recursion, probability, inclusion groups — all ported against SillyTavern 1.18 semantics). Three things a real run adds that this page doesn't simulate: token-budget trimming (low-priority entries get cut when too much fires), injection ordering, and cross-turn sticky/cooldown counters. Entries carrying those settings get a caveat badge. Q: Is my lorebook uploaded? A: No. Parsing and simulation run entirely in your browser — the page keeps working offline once loaded. That's also why you can throw a 50,000-word book at it: there's no server in the loop. ### Lorebook Migration Doctor — Cross-Platform World Info Checkup (SillyTavern / Chub / Risu / Agnai / JanitorAI) URL: https://foreverse.app/labs/lorebook-doctor Q: Why did my lorebook stop working after importing it into JanitorAI? A: Almost always the hidden 1% default: imported entries commonly arrive with trigger probability stuck at 1%, so your lore fires about once in a hundred messages no matter how well it's written. Open each entry and set probability explicitly (100 for critical lore), confirm the book is linked to the character and published, and check that Case Sensitive or Match Whole Words isn't silently blocking the keyword. This tool flags all of those before you migrate. Q: What actually breaks when I move a SillyTavern World Info book to RisuAI or Agnai? A: The nasty ones are semantic, not cosmetic. Risu's importer drops the enabled flag (your disabled entries come back alive) and drops secondary-key logic — a NOT_ANY blocklist inverts into a positive requirement, the exact opposite of what you wrote. Agnai carries constant, selective and case-sensitive flags in the file but its engine never reads them: always-on entries stop being always-on, exclusions vanish, and there is no probability roll at all. Every one of these is verified against the platforms' open-source code and flagged per entry by the checkup. Q: Does this tool convert my lorebook between platforms? A: No — deliberately. Converters already exist (including the import buttons built into SillyTavern and RisuAI themselves). What nothing else tells you is what silently changes on the way through. This is the pre-flight checkup: run it before you convert, fix or annotate the flagged entries, then migrate with your eyes open. Q: Which file formats does it recognize? A: SillyTavern World Info exports, Chub lorebooks (their dual-written WI + V2 format), RisuAI lorebook exports ({ type: 'risu' }), Agnai memory books ({ kind: 'memory' }), JanitorAI JSON-editor entry arrays, V2/V3 character books and lorebook_v3 files, character cards with embedded books, and NovelAI lorebooks (detected, light support). Drop any of them — the format banner tells you what it saw. Q: Is my lorebook uploaded? A: No. Parsing, the checkup and the divergence lab all run in your browser — the page works offline once loaded. ### Lorebook Merger — Combine World Info Books, Kill the Duplicates, Catch Key Collisions URL: https://foreverse.app/labs/lorebook-merger Q: How does it decide two entries are duplicates? A: Two criteria, applied together. Near-identical text (character-trigram similarity ≥ 90%) is grouped no matter what the keys say — that catches the classic 'same book, two versions' case. Moderately similar text (≥ 55%) is only grouped when the trigger keys also overlap (≥ 50% of the smaller key set), so two genuinely different entries that merely mention the same nouns don't get force-merged. Entries with the same keys but unrelated content are deliberately NOT treated as duplicates — they land in the key-collision list for you to judge instead. Q: What happens to entries when I export the merged book? A: Entries that came from SillyTavern or Chub exports pass through as the original JSON objects — depth, role, automation IDs, inclusion groups, even foreign extension blobs stay byte-identical; only uid and displayIndex are renumbered. Entries from Risu, Agnai, JanitorAI, NovelAI or embedded character books are rebuilt on the standard ST entry template from their parsed semantics (embedded books recover depth/role/automation_id from extensions). The export card tells you exactly how many entries took each path. Q: Why does it warn about trigger-word collisions instead of fixing them? A: Because a shared key isn't automatically wrong. Two entries triggered by the same word might be deliberate layering (one describes the city, one describes its politics) — or they might be two books that each wrote the same lore, which doubles your token spend and can contradict itself. Only you know which. The tool lists every shared key with the entries involved; if you want same-key entries to fire one at a time, give them the same inclusion group in SillyTavern. Q: Which formats can I feed it? A: SillyTavern World Info exports, Chub lorebooks (dual-written WI + V2), RisuAI exports ({ type: 'risu' }), Agnai memory books ({ kind: 'memory' }), JanitorAI JSON-editor entry arrays, V3 lorebook files, character cards with embedded books, and NovelAI lorebooks — the same parser family as our Lorebook Migration Doctor. The merged output is always SillyTavern World Info format, the closest thing the hobby has to a lingua franca. Q: Is anything uploaded? A: No. Parsing, duplicate grouping, collision checks and the export all run in your browser. The page works offline once loaded; refresh and everything is gone. ### SillyTavern Macro Playground — Expand {{char}} {{user}} {{random}} Online, Full Support Table URL: https://foreverse.app/labs/macro-playground Q: Is {{random:a,b,c}} different from {{random::a::b::c}}? A: No — both separators are valid. Double-colon wins when present; otherwise commas split. Use the double-colon form when an option itself contains a comma. One option is picked with equal probability per generation — the standard trick for randomized greetings. Q: What's the full {{roll}} syntax? A: NdM rolls N dice with M sides and sums them: {{roll:2d6}} is two six-siders. N defaults to 1 ({{roll:d20}}), a bare number is the side count ({{roll:20}} equals {{roll:1d20}}), and the space form {{roll 1d100}} works too — the common notation in skill-check cards. A malformed roll ({{roll:abc}}) doesn't error; it outputs the raw argument. Q: Do macros expand recursively? A: No. Expansion is single-pass: if the value {{getvar}} pulls out contains {{char}}, it stays literal — both the tavern and the app engine behave this way to stop runaway nesting. The annotated view here shows it directly: macro shapes inside an expansion product stay unexpanded. Q: What happens to an unrecognized macro? A: It's sent to the model as-is — no error, no removal. That's exactly how "the model keeps seeing curly braces" accidents happen when a card typos a macro name. This tool underlines unrecognized macros with a dotted line; scanning a card here before publishing beats discovering it mid-chat. ### Narrative Fingerprint — Read a Chapter's Structural Accent, Not Its Word Choice URL: https://foreverse.app/labs/narrative-fingerprint Q: Is this the actual StoryScope method? A: No, and the page says so up front. The paper extracts 304 narrative features by having an LLM read each story — that can't run free in a browser. What we ship is the subset of its 30 core features that leave lexical traces a regex can see (emotion naming vs body-performing, retrospection markers, scent imagery, gnomic moralizing, reader address), each validated on our own labeled corpus before shipping. Axes whose projections didn't separate cleanly are shipped in a 'reference' tier that says exactly that. Q: Can it tell me whether a chapter was written by AI? A: No. It counts which calibrated bands your chapter's structural readings fall into — human band, AI band, or the overlap — and stops there. No percentage, no verdict. Two honest reasons: our calibration corpus is small (24 human + 24 AI English segments), and every axis here can be deliberately written toward. Treat it the way the paper treats its features: as interpretable evidence, not judgment. Q: Where do the human and AI bands come from? A: Human EN band: 24 mid-book windows from 12 public-domain novels (Project Gutenberg, 1898–1926 — Fitzgerald, Hemingway, Christie, Conrad and co.), ~4.2k words each. AI EN band: 24 windows from 12 of our own published pipeline novels, drafted by an LLM under an anti-slop discipline — deliberately the hard case, since that prose already dodges word-level tells. Both bands were computed by the same code that runs in your browser, and the per-segment values are frozen in our repo. Q: Is my text uploaded anywhere? A: No. The whole analysis is one regex pass over your paste, in your browser, sub-second even on a 300k-character book. No model download, no server call, works offline. ### Novel to Lorebook Skeleton — Entity Extraction from txt/epub into SillyTavern World Info URL: https://foreverse.app/labs/novel-to-lorebook Q: Does it write the lorebook entries for me? A: No — deliberately. Entry bodies need someone who actually read the book: this page runs zero AI and invents zero lore. What it does produce is everything mechanical about a lorebook: which entities exist, their trigger keys, which chapters they span, how often they appear, and a token-budget hint per entry. The fill-in brief it copies out carries the craft rules, so any AI (or the Foreverse in-app butler agent) can finish the job against the source text. Q: How are entities found without AI? A: Rule signals, counted honestly. Chinese books: speaker patterns (a 2–4 character name directly before 说道/问道/笑道…), place and organization suffixes (城/镇/山/宗/学院…with junk characters stripped from the left), and quoted terms (「青冥诀」). English books: capitalized phrases that appear mid-sentence, with a speaker-verb check for characters. Everything below the frequency thresholds (2 hits for speakers, 3 for the rest) is dropped — the tool prefers missing an entity to inventing one. Q: What do I do with the JSON it exports? A: It's a standard SillyTavern World Info file — the same field family ST 1.18 writes. Import it into ST's World Info panel (or Foreverse), and each skeleton entry appears with its trigger keys, ready for you to replace the template text with real lore. Entries you didn't check in the table stay out of the file. Q: Why did it miss a character I know is important? A: Three usual reasons: the name never appears in a speaker pattern (a silent character), the book's chaptering defeated detection (ranges then default to chapter 1), or the name is below the frequency threshold. That's the honest cost of rule extraction — add the missing entity as one more entry in the exported JSON; the format is plain to edit. Q: Is my book uploaded? A: No. txt decoding, epub unpacking, chapter detection, entity counting and JSON building all run in your browser. Works offline once loaded. ### PNG Character Card Embed & Extract — JSON to PNG / PNG to JSON, in Your Browser URL: https://foreverse.app/labs/png-card-tool Q: Why does a screenshot of a card PNG lose the character data? A: The persona lives in PNG text chunks (chara / ccv3), not in the pixels. A screenshot re-renders only the pixels; chat apps and social platforms re-compress images the same way. The image survives, the chunks don't. Always pass card PNGs around as files, and rebuild lost ones here from the JSON. Q: What are the chara and ccv3 chunks? A: Two conventional PNG tEXt chunks that the SillyTavern family uses to carry card JSON as base64. chara is the classic V2-era keyword; ccv3 is the chara_card_v3 keyword, and readers prefer it when both exist. This tool writes chara always, plus ccv3 when the card declares spec chara_card_v3 — matching how mainstream exporters double-write. Q: Will SillyTavern and other frontends read the PNG this tool builds? A: Yes. The output is a standard PNG with tEXt chunks placed between IHDR and the image data — readers in SillyTavern, RisuAI and Foreverse walk all chunks, so placement doesn't matter to them. We verified round-trips byte-for-byte: embed, re-extract, and the JSON comes back identical. Q: Is my card or image uploaded anywhere? A: No. Chunk parsing, base64 coding, CRC32 checksums and JPG-to-PNG transcoding all run in your browser with standard web APIs. The page keeps working offline once loaded. If a card is private, it stays private. ### Tavern Preset Viewer — Turn a Preset JSON into a Readable Config Sheet URL: https://foreverse.app/labs/preset-viewer Q: Why do some blocks say "dynamic" with no text? A: Those are marker blocks — placeholders like chatHistory, charDescription or worldInfoBefore. At runtime the frontend replaces them with your card's fields, triggered lorebook entries, or the conversation itself. A preset file never contains that text, so this viewer explains what fills each slot instead of pretending the block is empty or broken. Q: The preset has more prompts than the viewer lists. Where did they go? A: Only blocks referenced in prompt_order take effect — that's tavern semantics, not a viewer limitation. Prompts defined in the pool but never ordered are inert; the viewer counts them in a footnote instead of mixing them into the injection list. Q: What does @depth mean on a block? A: Blocks with absolute injection don't sit at their list position — they get inserted N messages from the end of chat history. Authors use it for reminders that must stay close to the latest turns. @depth 2 means two messages from the bottom. Q: Does seed: -1 do anything? A: No — -1 means random, and tavern-family frontends only send the seed when it's zero or positive. A non-negative seed pins the sampling: same input, near-identical replies, and re-rolls stop changing much. The viewer flags a pinned seed because it's a common "why are my swipes all the same" cause. ### LLM API Pricing Comparison Table — 1,300+ Models from 30 Providers, Sortable URL: https://foreverse.app/labs/price-table Q: Where does the data come from, and how fresh is it? A: From a snapshot of models.dev, the public list-price registry — capture date is printed under the table (currently 2026-07-24). Our billing backend syncs from the same source. Promotions, negotiated rates and regional pricing are excluded; providers can reprice at any time, so check their pages before spending. Q: Why is the cache-read column blank for some models? A: A blank (—) means that provider hasn't published cache pricing — it does not mean zero. Where it exists, cache-read is often a small fraction of the full input price, which makes this column matter more than the input column for long chats. Q: What are “gateways” and why are they hidden by default? A: Resale platforms like OpenRouter, Poe and 302.AI: the same model appears on several of them at resale prices, not official list prices. The default view keeps to first-party providers so comparisons stay clean; toggle “include gateways” to compare resale prices too. Q: The app ships 62 provider presets — why only 30 here? A: The other 32 are either local/self-hosted runtimes (no list price exists) or services absent from the public price registry. We don't invent data — and these 30 cover nearly every model people actually pay for. ### SillyTavern Prompt Assembly Inspector — See How Preset + Card + Worldbook Become One Prompt URL: https://foreverse.app/labs/prompt-assembly Q: Why does my preset block seem to do nothing? A: Four usual suspects, all flagged directly by this tool: ① the block is defined but never placed in prompt_order — per tavern semantics it's inert, and the inspector lists it under preset-block diagnosis; ② the block is disabled in the order; ③ it's set to absolute injection (injection_position=1), so it doesn't sit at its trunk position at all — it gets spliced into chat history N messages from the end; ④ your main block got silently replaced by the character card's system_prompt — that's default tavern behavior (prefer character prompt), and the inspector badges the replaced main as “card override”. Q: My lorebook entry fired — why is the content still missing from the prompt? A: Firing and delivery are two different things, with four gates in between: the preset disabled the World Info marker (content dropped entirely); the marker isn't in the order at all (content gets appended to the very end of the prompt — real tavern behavior, rarely what anyone wants); an @depth entry got cut during token packing; or the order has no Chat History marker, so @depth injections have nowhere to land. This page gives every entry a verdict: where it went, or which gate stopped it. Q: Where exactly does an @depth injection land? A: Count backwards from the newest message: depth 0 sits at the very end, depth 2 sits two chat messages up, and injected messages themselves don't count toward depth. Multiple injections at the same depth, order group and role merge into one message (preset blocks first, then lorebook/author's-note extensions); different order groups stay separate, higher numbers first. All of it replicates SillyTavern 1.18's injection code — and you can see every landing spot in the assembled result. Q: Will the simulation match a real send exactly? A: The trunk semantics match: prompt-manager order, marker filling, card overrides of main/jailbreak, all lorebook injection positions (before/after/@depth/examples/author's-note/outlet), macro expansion, and the budget-packing order — each checked function by function against SillyTavern 1.18 source. A few layers are deliberately out of scope: the lorebook's own budget trimming (25% of context by default — you get a warning when activated entries exceed it), regex scripts, cross-turn sticky/cooldown, group chats, continue/impersonate generation types, and extension injections. Token counts are a character-ratio estimate, not a real tokenizer. Full list under “What it can't do”. Q: Are my preset and card uploaded? A: No. Parsing, entry scanning and assembly all run in your browser — the page keeps working offline once loaded. That's why you can drop a whole novel-sized lorebook in: there's no server in the loop. ### Fiction Prose Stats — Sentence Rhythm, Dialogue Ratio and Repetition at a Glance URL: https://foreverse.app/labs/prose-stats Q: What is the sentence-length variation number, and why does it matter? A: It's the coefficient of variation — standard deviation of sentence lengths divided by their mean. Human fiction mixes three-word jabs with forty-word sprawls; machine continuation tends toward the middle. In our nine-model, twenty-round continuation experiment, human originals sat around 0.64 while bare-template AI drifted down to 0.52 round after round. Uniformity is the first structural tell, which is why this page leads with it. Q: Why doesn't this page give my text a score? A: Because the honest output is a description, not a grade. A low dialogue ratio doesn't make a book worse; a jumpy rhythm isn't automatically 'better prose'. The reference bands tell you where a passage sits among 378 real chapter samples — what you do with that position is a reader's judgment, and we'd rather leave it one. Q: How is this different from the AI-Flavor Checkup? A: Different layer. The AI-Flavor Checkup reads vocabulary — stock phrases and sentence molds ('a flicker in the eyes', 'a hint of'). This page reads the skeleton: length distribution, dialogue share, paragraph shape, verbatim repeats. A text can pass one and fail the other; running both gives you the word-level and structure-level views of the same passage. Q: Where do the reference bands come from? A: From 378 chapter samples across 10 in-house Chinese novels, computed 2026-07-24 with the same sentence and dialogue rules this page runs. In that corpus, sentence variation's middle band (P25–P75) is 0.62–0.72 and dialogue ratio's is 22–40%. They're reference points for 'where does my text sit', not quality thresholds. ### Try Your Book — Drop a txt and Read It with Real Page Turns, in the Browser URL: https://foreverse.app/labs/reader-demo Q: Why only the first three chapters? A: The demo phone is a hands-on sample, not a reading app. Keeping the window at three chapters keeps a 30MB tome from flooding the page with DOM, and it keeps the promise honest: the full book — with progress memory, branching continuation and audiobook — reads in the app. The page tells you exactly where the sample window ends. Q: Does my file get uploaded? A: No. The file is read, decoded and chaptered entirely in your browser; nothing leaves your machine and nothing persists — refresh the page and the book is gone. That's also why the demo can't remember your reading position. Q: The chapters look wrong — one giant chapter, or split at odd places. A: Chapter detection ports the core rules of a production importer (第N章 families, numeric headings, Chapter N, prologue markers). Rare heading styles can defeat it — the app's full importer covers more shapes and supports manual splitting. If no chapters are detected, the demo samples the first 15,000 characters as a single chapter and says so. Q: Are the page turns real pagination or a scroll trick? A: Real pagination: the text is laid out in fixed page-width columns, so pages break where the type actually breaks — change the font size and the page count re-measures. The turn animation is the slide mode from our phone-simulator engine; taps on the left and right edges flip, the middle toggles the reading chrome. ### SillyTavern Regex Tester — Debug Regex Scripts Online, Live Before/After URL: https://foreverse.app/labs/regex-tester Q: Why isn't my SillyTavern regex doing anything? A: Check three fields first: ① placement — the script declares which lane it runs on (user input / AI output / world info…); if your test context doesn't match, the whole script is skipped, and the context dropdown here reproduces that exactly; ② disabled — exported collections routinely carry entries the author switched off; ③ markdownOnly — it only touches the display layer, so you'll never see the change in the outgoing prompt. If all three check out, the pattern itself probably isn't matching — the before-pane highlights will tell you instantly. Q: What's the /pattern/gi notation in findRegex? A: A JavaScript regex literal: pattern between the slashes, flags after (g global, i ignore-case, m multiline, s dot-matches-newline). SillyTavern's web build speaks this dialect natively, and this tool matches it exactly. Bare patterns without slashes work too, treated as global replace. Q: What can go inside replaceString? A: Three families: $1/$2 numbered backrefs and $ named groups (referencing a group that doesn't exist yields an empty string); {{match}} for the whole matched text; and variable macros like {{setvar}}, {{incvar}}, {{getvar}} — beautifier cards number their option lists with {{incvar}} this way. All supported here, and the counter carries across matches just like the runtime. Q: What kind of regex freezes the chat UI? A: Nested unbounded quantifiers — (a+)+ or (.*)*. On long text they backtrack exponentially; one bad script can pin the UI for minutes. This tool blocks them at compile time with the same structural guard the app ships, plus a 350ms execution timeout as backstop — and tells you why, so you can rewrite the pattern. ### Repetition & Sampler Doctor — Why Your Bot Keeps Looping, and What Actually Fixes It URL: https://foreverse.app/labs/repetition-doctor Q: Is my chat uploaded anywhere? A: No. Detection, triage and the settings checkup all run in your browser — paste text or drop a .jsonl and nothing leaves your machine. That's also why the page can't 'learn' from your data: the rules are fixed, versioned, and the same ones our own bench pipeline uses. Q: Will the recommended settings stop the looping for sure? A: No, and we won't pretend otherwise. Samplers are model-specific voodoo; reference bands here are labeled by where they come from (official docs, community consensus, or our own measurements) and are meant as a starting point for one change at a time. The diagnosis half — which layer is repeating and whether the loop already lives in your history — is the part with hard evidence behind it. Q: Why does the fix list keep saying “edit history first” instead of giving a magic parameter? A: Because we tested it. In our app-side experiment, once a bad pattern sat in the visible history, corrective instructions failed in 9 of 10 rounds — while a fresh chat showed zero residue. Models imitate the transcript more than they obey you. Parameters shape the next token; the history teaches the next reply. So history surgery outranks every knob, and the page is honest about that order. Q: What exactly is a “12-token shingle”? A: Slice the text into every run of 12 consecutive tokens (a CJK character or a latin word each), then check how many of a reply's runs already appeared in earlier replies. It's the recipe our fiction bench uses for cross-round repetition. As a calibration check, this page's engine re-ran the archived Grok 4.5 chain and reproduced the published finding exactly: rounds 10-12 at 100% overlap, easing at round 13 (published 68.4%, recomputed 67.7% — sub-point tokenizer difference). ### Story Theater — Drop a Novel TXT, Watch It Play as a Visual Novel URL: https://foreverse.app/labs/story-theater Q: Why do some lines play as narration without a name tag? A: By design. The compiler only assigns a speaker when there's clause evidence — "Lin Xia said:" or "…," said Holmes. When evidence is weak, the line plays as narration instead of guessing: one wrong name tag costs more trust than ten unattributed lines. The ruleset was validated against hand-labeled ground truth from a real 400-chapter novel — 15 representative passages, zero wrong attributions at the strong-attribution gate. The report shows the honest attribution rate for your book. Q: Does my book get uploaded? A: No. Decoding, chapter detection, cast discovery and scene compilation all run in a background thread of your browser. No request carrying book content ever leaves the page — it works offline once loaded. The share button shares the tool link or a public-domain demo position, never your file. Q: Is the cast list guessed by an AI? A: No. It's counted from speech clauses: a name needs at least two stable appearances in the speaker slot ("X said") to make the list. Addressees ("said to X"), pronouns and generic nouns are excluded; "Holmes" and "Sherlock Holmes" merge into one. If the count got something wrong, edit the list on the report page — the performance re-stages instantly. Q: I only have an epub. A: This v1 takes txt only. An epub is a zip container, and a half-done browser unpacker gives worse answers than an honest refusal. Desktop readers like Calibre export epub to txt in one step. Going the other way — packing txt into epub — is what our epub maker tool does. Q: Chapter detection got it wrong / the whole book is one slice. A: Chapter rules are the same engine as our TXT checkup tool (ported from a production importer): 第N章 families, numeric headings, Chapter N with Roman numerals, prologue markers. A book with no recognizable heading lines gets sliced by line count, and the report says so. For a full chapter diagnosis, run the file through the checkup page. Q: How does this relate to the theater in the Foreverse app? A: Same compilation rules. The web version is a bare stage: no portraits, no music, performing only the words already in your book. The app's theater is the full form — at a fork you pick a direction and the AI continues the story live as a new branch, with scene art on demand. The end-of-chapter card links there. ### Text X-Ray — Per-Word Predictability Lens, a Neural Net Running in Your Browser URL: https://foreverse.app/labs/text-xray Q: Is my text uploaded anywhere? A: No. On first scan your browser downloads a 73–81MB quantized mini-model from a CDN; after that, all inference happens on your device in a Web Worker (WASM). Nothing is sent to any server — with the model cached, the scan even works offline. Q: Can it tell whether a text was written by AI? A: No, and we won't pretend otherwise. We calibrated the instrument on 100 provenance-labeled passages: on this probe model's scale, formulaic human webfiction (median 69% top-10 hit rate in Chinese) reads as MORE predictable than raw LLM continuations (57%). Perplexity and authorship stopped lining up years ago. What you get is per-word texture, not a verdict. Q: Why does it download 70+ MB? A: Because a real neural network runs in your browser: the Chinese track is a 6-layer distilled Chinese GPT-2 (q8 ONNX, 73MB), the English track is DistilGPT-2 (81MB). Each downloads once and stays in browser cache; engine init takes a one-time 10–30s per page visit, and the scan itself is sub-second. Q: What's the relationship with GLTR? A: Homage and a hand-off. GLTR (Harvard NLP + MIT-IBM Watson AI Lab, 2019) colored each word by its rank in GPT-2's predictions — the founding artifact of visual AI-text forensics. It stayed in 2019: server-side, English-only, with its own disclaimer about newer models. This is the same idea rebuilt for 2026: it runs in your browser, covers Chinese and English, and puts the 'fails on newer models' caveat on the table with our own calibration data. ### Character Card Token Counter — Live Estimate, Plus an Exact o200k Mode with Token Boundaries URL: https://foreverse.app/labs/token-counter Q: Is this a real tokenizer? A: Two tiers. The default is a character-ratio estimate: CJK at ~1.6 characters per token, everything else at ~4 — the DeepSeek-documented ballpark, the same ruler every tool on this site uses. Switch to precise mode and it becomes a real tokenizer for the OpenAI family: o200k_base (the published vocabulary of GPT-4o and later) runs in your browser, every token recomputable, boundaries tinted on screen — the ~2.3MB vocabulary downloads only when you first switch. DeepSeek, Claude and Gemini never published their tokenizers, so those families honestly stay as estimate plus range. That's a capability boundary, not laziness. Q: Is my text or card uploaded? A: No. Counting and card parsing run entirely in your browser — the text never leaves the textarea, and card files are decoded with browser APIs. The page keeps working offline once loaded. Q: Why does another tool give a different count for the same card? A: Three usual reasons: it runs a different tokenizer (families differ by 20–30%, more on CJK text); it counts different fields (some tools sum every field including creator notes, which never enter the prompt); or it counts the lorebook differently (constant entries bill every turn, keyword entries only when triggered). This tool separates all of that explicitly, and uses the same ruler as our Card Checkup, so numbers compare across tools on this site. Q: What counts as “permanent” tokens in a character card? A: Whatever re-enters the context on every single turn: description, personality, scenario, system_prompt, post_history_instructions, and constant lorebook entries. The greeting appears once (but stays in history), example dialogue is dropped first when context runs tight, and creator_notes never enters the prompt at all. The field table tags each row with its billing identity. Q: How many characters is a token? A: Roughly 4 characters of English per token, and only ~1.6 characters of Chinese, Japanese or Korean — the ratio this page counts with. Flip it around: 100 tokens is about 400 English characters (75-ish words) but just ~160 CJK characters, and a 1,500-character text weighs ~375 tokens in English versus ~900 in Chinese. That asymmetry is why a bilingual card can double in weight when translated, and why every figure here is shown as a per-family range rather than one fake-precise number. ### Human or AI? — A Daily Blind Test on Fiction Prose URL: https://foreverse.app/labs/turing-arena Q: Where do the questions come from, and how do they rotate? A: Two source pools, both with clean display rights. Human passages come from fourteen public-domain novels (Project Gutenberg — Austen, Wharton, Chopin, Cather, Dickens, Doyle, Wells, Stoker, London, Collins, Orczy, Jerome, Wodehouse, O. Henry): each question shows the author's actual next paragraphs after a mid-book anchor. AI passages are real, untouched continuations of the exact same lead-in, generated in 2026-07 by four models from four vendors (Claude 4.6 Sonnet, DeepSeek V4 Pro, Qwen3.7-Max, Gemini 3.1 Pro) with the same instruction, a matched word target and the same temperature. The day number picks the question deterministically, rolling over at midnight UTC+8; past days stay playable as practice. Once every question in the bank has had its day, the rotation replays earlier questions — labeled on the page as a replay, never passed off as new. Q: What's the "LLM judges scored 12%" line about? A: A double-blind reliability experiment we ran in 2026-07: four heterogeneous LLM judges rated 7 pairs of character-card prose in both orders, against gold labels from real readers. They agreed with each other 86% of the time — and scored 2/16 = 12% against gold. Unanimously wrong: they read dense, evenly-polished detail as "human" when real readers read it as AI. That number is your opponent here. Q: Is this fair to the models — and haven't they memorized these books? A: The prompt gives each model the book title, the author, the exact preceding passage and the human continuation's word count, and asks it to continue in the author's voice — no sabotage, temperature 0.8, output shipped untouched (we only strip obvious instruction artifacts like a "here is the continuation" lead line, and record it when we do). Memorization is real: some models can reproduce a classic's actual next paragraph nearly verbatim. Any generation sharing an 8-word window with the author's real continuation is rejected from the bank, so what you're judging is genuine imitation, not recall. Every reveal names the model and the generation date. Q: Where is my record stored? Is there a leaderboard? A: In your browser's localStorage only — no account, nothing uploaded. The English and Chinese tracks keep separate records. There's no global accuracy stat yet (this page has no backend); the 12% judge score serves as the fixed benchmark to beat. A crowd accuracy line is on the roadmap. ### PNG to SVG Converter — Free Image Vectorizer in Your Browser URL: https://foreverse.app/labs/vectorizer Q: Is my image uploaded anywhere? A: No. Decoding, downsampling and tracing all happen in a Web Worker inside your browser — the page keeps working offline once loaded. Our server only delivers this page and the engine file itself. Q: How is this different from paid tools like vectorizer.ai? A: Two different technical routes. This tool is classic geometric tracing (the open-source VTracer engine): free, unlimited, no watermark, and your file never leaves the device. The paid leaders run server-side AI models that reconstruct shapes — on photographs they genuinely do better. For logos, icons, illustrations and line art, geometric tracing is the right tool and the gap disappears; for photo-realistic reconstruction, that's their turf, and we say so. Q: What kind of image converts best? A: Flat-color material: logos, icons, stickers, illustrations, colored comic panels, stamps and signatures (use the B&W preset for those). The rule of thumb is simple — the flatter the color patches and the cleaner the edges, the cleaner the vector. Photos and heavy gradients come out as posterized color blocks. Q: Why is the SVG bigger than my original image? A: Vector file size scales with the number of shapes, not pixels. On photo-like input every small color patch becomes a path, and an SVG with thousands of paths easily outweighs a JPG. Drop color precision a notch or two, or use the Photo (stylized) preset, and the size falls by an order of magnitude. Logo-class input is the opposite — usually just a few KB. Q: Can I use the output commercially? A: Your image, your call: the tool never touches your file, adds no watermark, and attaches no license terms of ours to the output. The engine itself is MIT-licensed open source. The one thing to check is the copyright of the source image — conversion doesn't change who owns the picture. ### TTS Voice Gallery — 96 Voices, Same Sentence, Listen Before You Pick URL: https://foreverse.app/labs/voice-gallery Q: How were these samples generated? A: One-off batch on 2026-07-24: each of the 96 voices synthesized one short sentence through its provider's own TTS API — Alibaba DashScope qwen3-tts-flash, xAI grok-voice-latest, OpenAI gpt-4o-mini-tts, SiliconFlow CosyVoice2. Everything was then converted to mono MP3 at matched loudness, 17–31KB per file. Chinese-locale voices read one Chinese sentence, everything else reads one English sentence — so the differences you hear come from the voices, not the script. Q: Why might a voice have no sample? A: The catalog mirrors the app's built-in 96-voice list. If a provider call fails or a channel is unavailable when samples are regenerated, that voice is honestly marked "no sample yet · available in app" instead of being faked with a different recording. This batch came back 96 for 96; future catalog updates follow the same rule. Q: Is listening here paid? How does pricing work in the app? A: This page serves static files — tap as much as you like, nothing is billed and nothing is synthesized live. In the app, audiobook listening is real synthesis billed by volume: the official channel charges list price plus a 50% service fee per character against your credits (the 5,000-credit signup grant covers a few chapters of listening), or you can bring your own provider key and pay your own bill directly. ### New Cards Weekly — Recent Community Character Cards, One Page URL: https://foreverse.app/labs/weekly-cards Q: Where do these cards come from? A: From the Foreverse community hub's public shelf — currently all published by the official account, which is why every card carries a full persona, greetings, and usually an embedded lorebook. Covers are AI-generated and origin-marked. The page reads a static snapshot of that shelf; nothing is scraped from other sites. Q: How do I actually get a card? A: Each card links to a detail page with the full persona and greetings. To chat with it, open the Foreverse app's Community tab and search the card's name — import is one tap, and the lorebook and example dialogues travel with the card. The cards are free. Q: Why does a week sometimes have no new cards? A: Cards ship in release batches, not on a fixed schedule. Weeks with no batch simply don't appear — the page never pads itself with placeholders. The newest drop week is pinned in the header with its real date. Q: Can I subscribe? A: Yes — the RSS feed at /labs/weekly-cards/feed.xml lists the most recent cards with links to their detail pages. Point any feed reader at it; no account, no mailing list. ======================================================================== ## 免费工具 — 常见问题（中文） ### AI 味检测 — 在线查中文文案的 AI 痕迹与套话句式，免费零上传 URL: https://foreverse.app/zh/labs/ai-flavor Q: 用这个工具时我的文字会被上传吗？ A: 不会。检测器是纯规则引擎，完整运行在你的浏览器里，任何内容都不会发送到服务器。 Q: 为什么只支持中文？ A: 词表和句式规则是在中文角色卡与网文语料上建立并校准的。英文需要独立的语料和校准，我们还没做，就不假装支持。 Q: 得分高能证明是人写的吗？ A: 不能，得分低也不能证明是机器写的。分数低表示命中了较多读者常判「机器味」的已知句式，值得逐项复核；结构完美的模板化 AI 文案依然能拿满分。把它当文案 lint 用，不要当真值裁判。 ### 书籍体检器 — txt 拖进来，看章节识别、字数、编码和听书时长 URL: https://foreverse.app/zh/labs/book-checkup Q: 为什么 200 万字的书被识别成了一整章？ A: 章节识别靠的是能认出来的标题行——第N章族、「001 标题」式数字行、Chapter N、序章番外这些。全书一个都没有（或样式太少见）时，整本落成一章：听书没法按章分段，AI 续写的上下文也只能按字数硬切。这页移植的是生产导入器的规则主干；App 完整版覆盖更多样式，还支持手动分章。 Q: 预览里全是乱码，怎么回事？ A: 文件编码多半不在这页的探测链里。探测链覆盖 UTF-8（严格与轻度损伤）、GBK/GB18030、带与不带字节序标记的 UTF-16——真实书库里几乎全部形态。Big5 繁体和西欧单字节编码不在链里；在电脑上用编辑器转存 UTF-8 再试，或直接交给 App 的完整探测链。 Q: 为什么不收 epub？ A: epub 是 zip 容器：要拆包、解 xhtml、按 spine 顺序拼章节。在浏览器里做个半吊子版本，给出的答案比不做更误导——所以网页版 v1 只收 txt，并且明说。App 的导入器直接收 epub。 Q: 阅读和听书时长是怎么算的？ A: 阅读按每分钟 300 字（中文）或 220 词（英文书，按 CJK 占比判定）。听书 1× 速按每分钟 250 字或 150 词——和站内成本计算器同一换算——另给 1.25×、1.5×、2× 三档。这些是规划用的数字：你的真实语速归你。 ### 缓存省钱计算器 — 长对话开缓存能省多少，按真实价差算 URL: https://foreverse.app/zh/labs/cache-calc Q: 默认命中率 89.9% 是哪来的？ A: 我们自己的实测。2026 年 7 月做了一场四方对比评测：同一个模型、同一把 Key、逐字节相同的消息脚本，量出裸 SillyTavern 出厂命中 91.5%，我们当时只有 46.5%——病根是设定条目每轮注入的位置在漂移。修好之后新装形态 89.9%、老用户升级形态 73.1%。四个数字都做成了锚点按钮，点一下就能看对应的账。 Q: 省的比例为什么随轮数越聊越大？ A: 每轮请求都要把全部历史重发一遍，历史越长、其中「和上一轮逐字相同」的前缀占比越高。10 轮时前缀大约占六成，300 轮时超过 99%——所以短对话开不开缓存无所谓，长对话里它就是账单的大头。 Q: 缓存折扣要申请吗？谁给的？ A: 供应商给的，多数家自动生效：DeepSeek、Anthropic、Google、xAI 都会在响应里报告命中了多少 token、按折扣价计了多少。不用申请，但有一个前提——客户端把请求组织成「前缀逐字稳定」。做不到这一点，折扣就静默拿不到，这正是这页存在的原因。 Q: 这个数字能当账单预估吗？ A: 当量级参考可以，当账单不行。token 按中文混合口径估算；真实命中率取决于你的客户端行为（锚点按钮里那 45 个百分点的差距就是行为差距）；各家对缓存写入、缓存有效期的计费细则也不同。方法论节写了完整口径。 ### 写卡素材袋 — 不会写卡也能把想法变成一张角色卡的任务包 URL: https://foreverse.app/zh/labs/card-brief Q: 这个页面会用 AI 帮我写卡吗？ A: 不会，而且是刻意的——本站工具全部浏览器本地运行、零上传，页面自己从不调用模型。素材袋做的事是把素材整理好、交给真正会写的人：Foreverse App 里的管家 Agent（带整套写卡规程：选型、活人感写法、五种开场白范式、反 AI 味词表和评分闸），或者你自己在用的任何 AI（走「复制任务书」出口）。 Q: 点「发送到我的 App」会发生什么？ A: 表单会打包成一份带版本号的小 JSON（kind 是 fv-card-brief），经一次轻量邮箱登录后，推到你自己账号的云端暂存区——和本站卡片工具用的是同一条通道。App 端「我的云端」会亮红点；打开就能让管家用这份素材写卡。同一份素材重复发送不占新位置：云端按内容识别，只把 30 天保留期重新拉满。 Q: 在 App 里生成要花钱吗？ A: 写卡消耗你 App 账号的积分：新注册送 5,000 体验金，让管家写一张卡通常只用掉其中一小部分。素材袋本身和传输过程完全免费。 Q: 复制出来的任务书里有什么？ A: 一段完整指令：你的素材按节排好，加上写卡规程里不可让步的要点——具体细节优先于形容词、开场白是一场已开演的戏、绝不替 {{user}} 说话、反套话词表、示例对话的轮替规则、创作者笔记三件事。任务书要求 AI 输出一份 chara_card_v3 JSON，写完可以拖回本站的角色卡体检器或写卡诊室检查。 Q: 素材可以粗糙到什么程度？ A: 非常粗糙也行。每一格都可选，「一个嘴硬的裱画师，欠过我一个人情」这样一句话就是合法素材；参考区可以整段贴同人文或设定集——只用来学味道和设定，任务书里写明了不许抄原句。表单自动存在这台浏览器里，长文粘贴刷新不丢。 ### 角色卡体检器 — 免费检测角色卡格式、token 占用与导入失败原因 URL: https://foreverse.app/zh/labs/card-checkup Q: 体检时我的卡文件会被上传吗？ A: 不会。解析完全在你的浏览器里进行——PNG 数据块、JSON、charx 压缩包全部用浏览器自带能力本地解码，任何内容都不发往服务器。页面加载完之后断网也能照常用。 Q: 支持哪些格式？ A: 三种容器：PNG 卡图（tEXt / iTXt / zTXt 文本块里的 chara / ccv3 数据，读取顺序与酒馆一致 ccv3 优先）、纯 JSON（V1 平铺老卡、V2、V3），以及 .charx 压缩容器（包内 card.json 直接在浏览器里读出）。独立世界书 JSON 和群聊导出会被识别出来，如实告诉你「这不是角色卡」而不是硬解。 Q: 为什么 token 数和别的工具对不上？ A: 这里是字符占比估算，不是真分词器：中文按约 1.6 字/token、英文按约 4 字符/token——DeepSeek、GPT 系模型的常见口径。真实分词因模型而异，偏差可达两三成。把数字当量级看，别当精确值。 Q: charx 是什么格式？ A: chara_card_v3 规范定义的 zip 容器：包根一个 card.json，外加可选的资产（头像、立绘、背景）。本工具读取其中的 card.json 做体检；资产解包是完整前端在导入时做的事。网页版有一条边界：zip64 超大包（条目表 4GB+ 的形态，真实卡里极少见）这里解不了——遇到会如实提示而不是硬猜，Foreverse App 里可以正常导入。 ### 角色卡格式转换器 — v1/v2 升 v3、v3 归一 v2，丢什么如实列出 URL: https://foreverse.app/zh/labs/card-convert Q: v2 卡和 v3 卡到底差在哪？ A: V3（chara_card_v3）完整保留 V2 结构，新增：nickname（prompt 里顶替 {{char}} 的称呼）、群聊专用开场白、assets 资产引用（头像/立绘/背景）、多语言创作者注记、来源链接与时间戳、世界书装饰语法（@@activate 等），以及 .charx 打包容器。一张卡如果这些都没用到，v2 和 v3 就是同一张卡换了个包装。 Q: v3 降到 v2 会丢什么？ A: 只丢这张卡实际用到的 V3 专属字段——本页在下载前逐条列给你看。常见的有：资产引用、昵称、群聊开场白、多语言注记、来源和时间戳。人设正文、开场白、示例对话、内嵌世界书全部保留；世界书条目里的装饰语法行文本还在，但 V2 前端会把它们当普通文字发给模型。 Q: 转换会改我卡里的内容吗？ A: 不会。转换器只搬结构不动文字：字段在包裹之间逐字节归位。仅有的两处机械调整都会在页面上明说——V1 卡 name 为空时按导入器惯例兜成「Unknown」；V1 顶层的 talkativeness 收进 extensions（V2 规范位置）并限制在 0 到 1。 Q: 我的卡有 data 结构但没有 spec 字段，算什么版本？ A: 半规范卡，社区导出器和手改卡里很常见。多数前端按 V2 兼容读取；严格校验的工具会拒收。把它拖进来选 V2 转一遍就规范化了：内容不变，补上工具们认的 spec 和 spec_version 字段。 Q: 为什么没有「转成 V1」？ A: 刻意不做。V1 是 2018–2022 年的平铺老格式，只有六个字段——转出去意味着丢掉世界书、备选开场白、作者注、系统提示等 V2 起的全部结构，而现在没有任何主流前端只认 V1：能读 V1 的工具全都能读 V2。「降级到几乎什么都不剩的格式」不是转换是截肢，所以这里只收 V1、不产 V1。 ### 角色卡对比器 — 这张卡更新到底改了啥，逐字段高亮 URL: https://foreverse.app/zh/labs/card-diff Q: 它到底对比什么？ A: 三层。字段级：每个卡文本字段（描述、性格、开场白、示例、系统提示……）做逐词 diff，删的划红、加的标绿。世界书级：条目跨版本自动配对，分组成新增/删除/修改/未变，每条修改标出改动面（关键词、内容、constant/enabled 翻转、改名）。token 级：与本站角色卡体检器同一把估算尺，逐字段旧→新→变化量。 Q: 它说「零可见变化」，但两个文件明明不一样？ A: 这正是诚实三态在干活。原始 JSON 逐字节一致，它直说一致；只有键序或空白不同（编辑器重新序列化过），它说内容没变；而当差异藏在本工具不逐字段展开的区域——最常见是 extensions：随卡正则脚本、美化载荷、前端私有字段——它会告诉你「变化藏在未展示区」，而不是假装什么都没发生。 Q: 世界书条目是怎么跨版本配对的？ A: 五轮，从最严到最松：完全一致、内容一致、名字唯一相同、关键词签名唯一相同、内容相似度——而且对超短条目用更严的阈值，因为两条五个字的条目重叠三个字，不该被判成「同一条、改过」。五轮都配不上的，如实报新增或删除，绝不硬凑。 Q: 我的卡会被上传吗？ A: 不会。两个文件的解析和对比全部在你的浏览器里完成，任何内容都不发往服务器，页面加载完断网也能用。 ### 卡片名片图生成器 — 一张卡排成能直接发帖的宣传图 URL: https://foreverse.app/zh/labs/card-poster Q: 海报会用到卡里的哪些内容？ A: 只用公开货架面：卡名（缺了用昵称顶）、作者、标签、版本号，以及作者简介（creator notes）里那句话。定义正文——description、性格、场景、开场白、示例对话、系统提示、世界书——从头到尾不会被读。这不是隐私页里的一句承诺：金标自测里用 Proxy 逐次记录字段构建器的属性访问，碰了白名单外任何字段直接红。藏定义的卡，藏得住。 Q: 配色是怎么来的？ A: 从立绘来。图片缩小采样、量化统计，按「像素量 × 饱和度」打分——彩色的服装赢过灰色的背景。主色作种子派生三档变体（深色主调 / 浅色卡纸 / 对撞色），取色器找到多个强候选时还能逐个换。正文墨色自动按底色明度选，任何种子任何变体都保证 4.5:1 以上的对比度——这条是锁死的测试，不是手感。 Q: 二维码留白位是干什么的？ A: 版式里预留的一块干净白色方位，几何上保证永远不压到文字区（这也是锁死的测试）。导出后把你自己的二维码贴上去——Discord 邀请、Chub 主页、Foreverse 口令码都行。刻意不代生成二维码：海报全程本地，链接指哪也由你自己定。 Q: JSON 卡没有立绘怎么办？ A: PNG 卡自带立绘；JSON / charx 卡先用同色系的首字纹章顶着版面，随时可以传一张图换掉。传的图和其它所有东西一样只在你的浏览器里。 Q: 卡会被上传吗？ A: 不会。解析、取色、排版、导出全部发生在你的浏览器里。刷新就什么都不剩了。 ### 角色卡书架 — 批量导入 PNG 角色卡，浏览器里翻藏品、看世界书词条 URL: https://foreverse.app/zh/labs/card-shelf Q: 把卡加进书架时，文件会被上传吗？ A: 不会。解析全在你的浏览器里完成（PNG 数据块、JSON、charx 压缩包都是本地解码），书架本身存在浏览器的 IndexedDB——那是你自己电脑上按网站隔离的私有存储区。任何内容都不发往服务器。清除浏览器的站点数据，书架也就清空了。 Q: 支持哪些格式？能混在一批里导吗？ A: PNG 卡图（chara / ccv3 数据块，ccv3 优先，与酒馆同序）、纯 JSON（V1 平铺老卡、V2、V3）、.charx 压缩容器。混批完全没问题：每个文件独立解析，失败的单独列出原因，一张坏卡不会挡住其它卡进架。 Q: 关掉浏览器再打开，书架还在吗？ A: 在。卡存在 IndexedDB 里，刷新、重启、隔一周回来都还在。两条诚实边界：无痕/隐私窗口通常在关闭时清掉 IndexedDB（工具会提示并降级为「仅本次会话」模式）；浏览器在存储紧张时也可能回收站点数据。书架页面上会显示当前存储用量，心里有数。 Q: 封面图是哪来的？ A: PNG 卡自己就是封面——内嵌立绘直接显示。.charx 容器会在本地解出包里的第一张图片资产（优先 icon / avatar）。纯 JSON 卡文件里没有图，用首字母版画兜底；卡里写的远程头像网址刻意不去抓取——「零联网」的承诺要作数。 Q: 世界书词条能看到多细？ A: 每条词条独立成卡：主关键词、副关键词、完整内容，外加徽标——常驻、注入位置（含 @深度）、order、触发概率、副键逻辑（AND / NOT）、regex 键、粘滞 / 冷却 / 延迟计时器、递归门。词条搜索按关键词、名字、内容全命中。要模拟词条什么时候触发，用「世界书触发模拟器」；要查跨平台迁移会不会丢行为，用「世界书迁移体检器」。 ### 酒馆卡皮肤预览器/打包器 — 11 套主题美化正则包，在线试穿再导出 URL: https://foreverse.app/zh/labs/card-skins Q: 这里说的「皮肤」到底是什么？ A: 同一个文化母题下的 5 条正则。角色卡让模型按纯文本 marker 协议回复（:::think 内心 :::、[状态|…] 一行、:::scene 标题 # 正文 :::、:::note 旁白 :::、结尾 :::opt 甲 ¦ 乙 ¦ 丙 :::），每条正则接住一种 marker、渲染成对应母题的 HTML——折叠的思维卡、主题状态栏、场景纸、可点的选项笺。协议不变，母题随便换：换皮肤 = 换一包正则。 Q: 导出的包在酒馆里能用吗，还是只能在 Foreverse App 用？ A: 都能用，但有一条边界要说清。zip 里是 5 个酒馆正则面板格式的单条脚本（作用于 AI 输出、只改显示层 markdownOnly——上行 prompt 保持原文）；酒馆网页版对这类内联样式 HTML 正常渲染。装饰小图引用的是 App 专用素材通道（fv.local），在其它环境会自动隐藏、不裂图——模板本来就按「没素材也成立」设计。不渲染 HTML 的纯文本前端（API 直连类客户端）只会看到 marker 原文。 Q: 为什么我导入了正则，聊天还是没变样？ A: 正则包只管「渲染层」：模型不输出 marker 协议，它就无事可做。本页的「格式指令」就是干这个的——贴到卡的 system_prompt 末尾，压轴提醒放进 post_history_instructions，最好再让开场白里带一帧完整 marker 示范（few-shot 是最强的格式锚）。另外检查同时只启用一套皮肤的 5 条：两套同开会抢同一个 marker，谁先命中谁赢。 Q: 我贴进来的聊天内容会被上传吗？ A: 不会。正则匹配、模板渲染、手机预览、zip 打包全部在你的浏览器里跑，本页没有任何上传接口。预览还跑在一个禁用脚本的沙箱框里，你的文本在进模板之前已经整体转义——贴恶意内容也执行不了任何东西。 ### PNG 角色卡瘦身器 — 压立绘不杀卡，卡数据逐字节验真 URL: https://foreverse.app/zh/labs/card-slimmer Q: 为什么不能直接扔进 TinyPNG / 通用压缩器？ A: 因为它们是给图片造的，不是给卡造的。重编码 PNG 会从像素重建整个文件，文本块（tEXt / iTXt / zTXt）全部丢弃——而 chara / ccv3 卡数据恰好就住在那里。图片活下来了，角色死了。这个工具只重编码立绘，然后把原文件的卡数据块整块拼回去，一个字节都不经手。 Q: 怎么确定卡数据真的没坏？ A: 是验出来的，不是承诺出来的：下载按钮出现之前，工具把输出文件里的卡数据块提出来，和原文件逐字节比对。对不上就报错、不给下载。每一次压缩都跑这个检查。 Q: 「质量档」到底在做什么？PNG 不是没有质量参数吗？ A: 没错——浏览器的 PNG 编码器没有质量旋钮（那个只对 JPEG/WebP 生效），而卡必须留在 PNG 容器里（数据块只能住 PNG）。所以质量档做的是调色板量化：立绘被压到 256 / 128 / 64 色，配 Floyd–Steinberg 抖动，写成真正的索引色 PNG。立绘类图片 256 色档肉眼几乎看不出，体积通常砍一半以上。 Q: 为什么我的文件反而变大了？ A: 两个诚实的原因：原图已经被 pngquant 类优化器压到很紧（同等设置下浏览器编码器打不过它），或者原图本来就是索引色图、「无损重编码」把它展开回了全彩。发生这种情况时工具会直说，而不是藏起来——试更小的尺寸档、更狠的质量档，或者接受这张卡不需要瘦身。 Q: 卡会被上传吗？ A: 不会。解码、量化、拼块、验真全部用浏览器本地能力完成，任何内容都不发往服务器，页面加载完断网也能用。 ### 角色卡量产味体检 — 量出这张英文卡有多少句是从货架上拿的 URL: https://foreverse.app/zh/labs/card-slop-doctor Q: 角色卡语境里的「slop（量产味）」到底指什么？ A: slop 是英文社区给「库存零件拼出来的 AI 散文」起的名字：脊背发凉、气若游丝的嗓音、闪着狡黠的眼睛、五味杂陈的心情。单看哪句都不算错——人类也这么写。让一张卡变成量产货的是堆积：当开场白把同一套词表的每个货架都逛了一遍，这张卡读起来就是「生成后直接保留」，更糟的是模型会把这些短语当成本店风格，每次回复继续进货。这个页面量的就是堆积本身，不跟你争论品味。 Q: 检查时我的卡会被上传吗？ A: 不会。解析器和词表都在你的浏览器里跑——PNG 数据块、JSON、charx 压缩包全部本地解码，扫描是一组正则匹配。任何内容都不发往服务器，页面加载完断网也能用。 Q: 我最喜欢的卡也有两三个这种短语，它是量产货吗？ A: 多半不是——引擎也这么认为。一句「took a deep breath」出现在人类写过的每本小说里，所以单处命中永远不改变评级：评级按密度（每千词命中数）算，还带绝对下限——120 词的开场白带一个库存短语，照样是干净档。评级只在卡反复光顾同几个货架时才移动——那个模式是指纹，不是巧合。 Q: 名字规则标了我的角色，一定要改名吗？ A: 不强迫——但要知道这个名字的代价。对模型生成小说的语料统计发现「Elara」的出现频率是人类基线的数万倍，Seraphina、Kael 紧随其后。社区早认得这个模式：一张叫 Elara 的卡，读者还没看到你的正文，就先判了「问过模型、保留默认」。改名只花一次查找替换，却能摘掉全卡最响的机器标签。规则刻意放过的名字——Aria、Evelyn 这类现实常用名——不会被标。 ### 在线角色卡编辑器 — 新建、编辑、体检、导出角色卡，全在浏览器里（角色卡工作台） URL: https://foreverse.app/zh/labs/card-studio Q: 编辑时我的卡会被上传吗？ A: 不会。解析、编辑、实时体检、宏预览、两种导出全部在你的浏览器里完成，任何内容都不发往服务器。自动暂存存的是这台浏览器的本地存储，不是云端账号。 Q: 什么格式能进，什么格式能出？ A: 进：PNG 卡图、JSON（V1 平铺老卡、V2、V3）、.charx 容器——V1 老卡导入时自动升格为 V2 结构，这是现代前端的标准行为。出：v2 或 v3 的 JSON，或一张 PNG 卡（底图三选：导入时的原图 / 你上传的图（JPG/WebP 本地转码）/ 现场生成的占位底图）。.charx 不支持重打包——包内资产网页动不了。 Q: 和角色卡体检器是什么关系？ A: 同一套规则，两种姿势。体检器是只读的：拖文件、出报告。工作台把这套规则常驻在编辑器右侧，边写边跑——修好一条世界书死条目，红项当场消失。只想给下载来的卡要个结论，用体检器更快。 Q: 什么时候该用写卡诊室（Card Workbench Pro）？ A: 工作台是轻量档：快速改一张卡，结构体检加侧栏的代言 / AI 味诊断摘要。写卡诊室是给认真写卡的人的 Pro 档——IndexedDB 多草稿自动存（几 MB 的大卡也存得下）、字段内行内标注、system_prompt 与后置指令编辑、多成员群聊卡组装。修一张下载来的卡：留在这里；从零认真写卡：去写卡诊室。 Q: 编辑器不覆盖的字段会丢吗？ A: 不丢。编辑器没碰的一切——extensions、随卡正则、V3 assets、nickname、插入顺序、甚至别的工具写进去的非标准键——都原样带进导出文件。「随卡携带」面板会列出你这张卡带了什么。唯一的刻意例外：v3 降到 v2 会删掉 v3 专属字段，删什么会逐条列出来、先问你。 Q: 草稿存在哪里？会丢稿吗？ A: 停笔约半秒后自动存进浏览器 localStorage，下次打开页面会提示「恢复上次编辑」。两条如实说的边界：草稿不跨浏览器、不跨设备；卡太大（正文几 MB 的极端形态）存不进 localStorage 时，工具条会明说存不上，不会假装存了。 ### 写卡诊室 — 带实时诊断栏的角色卡编辑器（代言检测 / AI 味 / token 预算 / 群聊卡组装） URL: https://foreverse.app/zh/labs/card-workbench Q: 和「角色卡工作台」（Card Studio）是什么关系？ A: Card Studio 管结构：字段编辑 + 结构体检（死条目、坏宏、token 阈值）。写卡诊室把那套全部保留（同一个解析器、同一个世界书编辑器），再加上表单编辑器都没有的东西：内容诊断栏——边写边跑代言检测（你的卡在替 {{user}} 说话吗）、英文量产味评分、中文 AI 味评分、逐字段 token 预算条、写卡规程建议。另外补了 system_prompt / 后置指令编辑、群聊卡组装、IndexedDB 多草稿（几 MB 的大卡也存得下）。只需要改字段看体检，用 Card Studio 更轻。 Q: 编辑时内容会上传吗？ A: 不会。解析、全部诊断引擎、宏预览、PNG 写入、zip 打包都在你的浏览器里跑。草稿存这台浏览器的 IndexedDB，不进账号、不进服务器。 Q: 行内标注的颜色什么意思? A: 每个字段有「标注」页签：红=代言硬伤（卡替 user 说话/行动/内心），黄=代言弱信号（场景搬运、感知代入），紫=英文库存短语命中，橙=中文套话命中。悬浮看规则名。引擎和站上的代言诊断器、量产味体检是同一套——不是第二种意见。 Q: 群聊卡组装导出的是什么？ A: 按 Foreverse 社区群聊包（group_chat）载荷形状打的 zip：payload/group_chat.json（版本/成员/策略/群规）+ 每名成员一份卡 JSON（payload/members/）+ 可选压缩头像（assets/avatars/）。如实说边界：这是载荷层，不是能直接上架的 .fvpack——manifest、内容指纹、审核在 App 作者中心完成。组装时的校验逐条镜像产线校验器（2–12 人、名单名必须等于卡内名、群规 2400 码点内），导出的东西过那道闸不用返工。 Q: token 数准吗？AI 味分保真吗？ A: token 是字符比率估算（全站卡工具同一把尺），用来看哪个字段超重，不是真分词器，真实模型差 20–30%。中文 AI 味规则只在中文语料上校准过、英文词表只管英文散文——超出各自适用范围时诊断栏会明说不出分，不硬给一个假干净的分数。 ### 聊天记录考古馆 — 把酒馆聊天导出变成一份关系年报，免费零上传 URL: https://foreverse.app/zh/labs/chat-archaeology Q: 我的聊天记录会被上传吗？ A: 不会。本页没有上传接口——解析、统计、出图全部在你的浏览器里完成，刷新即清空。聊天导出是 RP 玩家手里最私密的文件之一，这正是这个工具做成纯本地的原因。年报卡在数据结构层就装不下聊天原文：它的字段只有数字和站方文案。 Q: 成本那个数字是我真实花掉的钱吗？ A: 不是，页面在数字旁边也这么写。它是重演估算：每条 AI 回复（包括你重摇划走的每一版）按一次请求计价，用 2026-07-24 的牌价快照，每一楼把全部历史重新装填、封顶在你选的上下文窗口。它不含角色卡、系统提示和世界书（实际用量比这个高），也不算缓存折扣（DeepSeek 缓存命中价约是牌价的 1/50，真实花费可能低一个数量级），token 是粗估不是各家真实分词器。如果导出里逐条记了模型名（新版酒馆会记），报告按那个档位精确计价；没记就给低价档到中高价档的区间。 Q: 支持什么文件？ A: SillyTavern「导出聊天」产出的 .jsonl（每行一个 JSON，第一行是可选的元数据）；Foreverse 导出的是同一格式。字段缺了会优雅降级——没有时间戳就收起作息卷，损坏的行跳过并计数。一次最多拖 8 个文件，多文件会多出一张跨角色对比表。单文件最大 48MB，分片解析，页面不卡。 Q: AI 味曲线是怎么算出来的？ A: 每条 AI 回复用我们 AI 味体检器同一套规则（v10.7 浏览器轻量版）打 0-5 分：套话词表低权重、句式模具高权重，规则在中文角色卡语料上做过零误杀校准。曲线画的是逐楼分数，并标出持续下滑开始的位置——「越聊越机」这个酒馆玩家都懂的体感，第一次被画成图。它是文案 lint 不是检测器；只支持中文，非中文为主的记录会如实给「不硬下结论」的说明。 ### 聊天记录转档器 — 把酒馆 .jsonl 转成干净的 Markdown 全文或可直接阅读的 HTML 网页 URL: https://foreverse.app/zh/labs/chat-export Q: 隐藏楼层和重摇候选是怎么处理的？ A: 隐藏楼层（is_system 消息——旁白注释、给模型的便签，本来就不算正文）会被滤掉，文档头部的出处行会如实写「已略过 N 条」。带重摇候选的消息只输出当前选中的那一版——被你摇掉的版本不会复活。这两条规则来自与「聊天记录考古馆」同一套解析器，两个工具报的数字永远一致。 Q: Markdown 和 HTML 该选哪个？ A: 文字要去别处，选 Markdown：博客、论坛、笔记软件、同人稿。正文里的 *动作* 星号原样保留（它本来就是合法 Markdown），唯一会转义的是行首的 #——防止某条消息把你的文档结构劫持成标题。文件本身就是归宿，选 HTML：独立单文件网页、书页式排版、动作弱化斜体、台词染色、深色跟随系统，而且零脚本——存着或者发人都干净。 Q: 它和「聊天长图工坊」什么关系？ A: 长图工坊出的是图——气泡、头像、配色，适合把名场面晒到社交平台。这里出的是文字——整段记录（或选段）变成可存档、可检索、可引用、可贴博客的文档。底下是同一套解析器，去处不同。想要十条消息的一张图，去长图工坊；想要记录本身，就是这里。 Q: 聊天记录会被上传吗？ A: 不会。解析、转档、预览、下载全部发生在你的浏览器里——这个页面没有任何上传接口，HTML 预览也跑在完全沙箱化的框架里。刷新就什么都不剩了。 ### 聊天长图工坊 — 把酒馆聊天记录排成一张能直接晒的长图 URL: https://foreverse.app/zh/labs/chat-shot Q: 支持什么格式的聊天记录？ A: 两种：酒馆聊天导出的 .jsonl（每行一个 JSON、行 0 是元信息——Foreverse 导出的也是这个格式），以及「名字: 内容」的纯文本（半角全角冒号都认，没名字开头的行自动并进上一条，多段的台词直接粘贴就行）。解析完可以指认哪个名字是你、再框选要晒的那一段。 Q: 聊天记录会被上传吗？ A: 不会。解析、排版、出图全部发生在你的浏览器里——这个页面没有任何上传接口，导出是从本地画布直接写进你的下载目录。刷新页面就什么都不剩了。也正因为不走网络，超长的记录也能秒级排出来。 Q: 为什么长对话会导出成好几张图打包？ A: 浏览器对单个画布有高度上限（普遍 16,384 像素，超了会静默画不出来）。工具把每一页都控制在上限之内，切页优先落在消息边界——除非单条消息比一整页还高，才退到行边界切。每页带页码，多页打包成一个 zip；短对话直接出单张 PNG。 Q: *动作* 和「台词」的颜色是怎么定的？ A: 按酒馆前端的通用渲染惯例：星号包住的动作描写排成弱化色斜体，引号里的台词（直引号、弯引号、「」『』都认）用强调色，其余旁白用正文色。四套配色都遵守同一套规则，自定义主题只换色板、不换规则。 ### 上下文透视镜 — AI 聊天为什么会失忆，看一眼就懂 URL: https://foreverse.app/zh/labs/context-lens Q: 所以 AI 到底为什么会忘掉前面聊过的事？ A: 因为每一轮请求都是把对话从头重发一遍，而模型只收固定预算的 tokens。系统提示、角色卡、世界书先占掉一块，剩下的空间从最新的楼层往回装——装不下为止。你在第 12 楼说过的名字不是被「忘了」，是物理上没进这次请求。页面上的格子图直接标出哪些楼层进了窗、哪些没进。 Q: 换个窗口更大的模型能解决吗？ A: 能推迟，但要为推迟付钱。128k 窗口装得下很多楼，可每一轮都按整个窗口的输入量计费——成本图里那根越来越高的线就是它。而且超长上下文下注意力质量会衰减，「在窗口里」不等于「被用上」。对聊到几百楼的对话，压缩＋缓存是更便宜的答案。 Q: 压缩摘要到底能救回什么？ A: 救大意，救不了原话。我们的压缩评测（53 万 token 语料、7 配置 × 98 探针）里，结构化摘要把压缩区的记忆通过率从纯截断的 24% 提到 59%——最大的单一因素是摘要有没有保住名字、数字这类具体值。逐字引用和细节救不回来，所以 App 里的压缩从不改写盘上的原文。 Q: 91.5% / 89.9% / 46.5% 这几个命中率哪来的？ A: 2026 年 7 月我们跑的计量评测：同一个模型、同一把 Key、逐字节相同的消息脚本、每次调用过同一个计量网关。裸酒馆出厂布局命中 91.5%；我们自己的 App 只有 46.5%——世界书注入位每轮漂移把缓存打断了。缺陷公开、修复、复测 89.9%。完整报告在方法论一节有链接。 ### 续写/听书成本计算器 — 听完一本 200 万字的书要几块钱，算给你看 URL: https://foreverse.app/zh/labs/cost-calc Q: 这些价格是官方定价吗？ A: 模型单价来自各家公开牌价的快照（页面标注采集日期，当前 2026-07-24），和我们服务端计费同步的是同一个数据源。计算给两种口径：官方渠道 = 牌价 + 50% 服务费（1 积分 = 0.0001 美元）；自带 Key 直连 = 牌价本身，App 不加价。促销价、协议价、区域价差不在内，最终以供应商官网为准。 Q: 「续写一段」的成本是怎么算出来的？ A: 按一次续写送入约 8,000 token 上下文、生成约 400 字正文（约 550 token）估算。这个口径和 App 内官方渠道的实测对得上：DeepSeek V4 Flash 续写一段实付约 19 积分。不同模型分词不同，估算误差约 ±15%，所以结果当量级参考，别当发票。 Q: 听一整本书真的可以免费？ A: 系统语音档是真免费：手机本地合成，不走任何 API，200 万字也是 0 元。在线 AI 音色按字符计费，整本 200 万字要几百块人民币——所以它适合精听单章（3,000 字不到一块钱），不适合整本囤转。App 把系统语音设成听书默认档，就是这笔账算出来的。 Q: 注册送的 5000 积分够干什么？ A: 5,000 积分折合 0.5 美元，按 DeepSeek V4 Flash 口径约等于 260 段续写（8 到 13 万字新内容），或几十个晚上的角色扮演陪聊。官方渠道不用配 Key，装好就能花；花完可以买积分，或者转自带 Key 直连供应商。 ### TXT 转 EPUB · 网文制书间——在线免费，浏览器本地打包不上传 URL: https://foreverse.app/zh/labs/epub-maker Q: 转出来的 EPUB，Kindle 能用吗？ A: 能。2022 年起 Send to Kindle 直接收 EPUB——邮箱推送、分享面板都行。产物同时带 EPUB 3 目录和老式 NCX 兼容目录，老阅读器的章节跳转也认。只有很老的 Kindle 固件要 azw3 格式，那种情况把这里的 EPUB 拿去 Calibre 过一遍即可。 Q: 章节识别错了怎么办？ A: 识别出的章节在打包前全部列出来：标题逐行可改，混进来的广告行、站名行取消勾选就排除掉。一个章节都认不出的书，会按每 8000 字左右均匀分节并明确告诉你——绝不假装识别成功。要把两章合并、在中间手动加分章点，去 App 的导入器做，这页不做半吊子版本。 Q: 我的文稿会上传到服务器吗？ A: 不会。解析、分章、打包全部发生在你的浏览器里，这个页面没有任何上传接口。这也是它能免费、不限次数、单文件放宽到 64MB 的原因：活儿在你的设备上干，我们没有服务器账单要转嫁。 Q: 从网站下载的小说能转吗？ A: 工具对文件格式是中立的，但请先确认来源允许你保存和转换。这一页为「你自己的文稿」设计：自己写的小说、整理的笔记、已取得授权的文本。 Q: 为什么不支持插图？ A: v1 只处理纯文本。txt 本身装不下图片，正文里的图片链接和「【图】」占位会按普通文字保留进书。带插图的制书（每章内嵌图片的 EPUB）不在这一版范围里。 ### 表情包工坊 — 按酒馆 28 表情规范校验、批量改名、一键 zip URL: https://foreverse.app/zh/labs/expression-pack Q: 28 个表情名的清单是哪来的？ A: SillyTavern 自己定的：内置本地分类器是一个 GoEmotions 模型（27 种情绪 + neutral），官方文档写明立绘文件必须按这些标签命名——admiration、amusement、anger、annoyance、approval、caring、confusion、curiosity、desire、disappointment、disapproval、disgust、embarrassment、excitement、fear、gratitude、grief、joy、love、nervousness、neutral、optimism、pride、realization、relief、remorse、sadness、surprise。写成 happy.png 或 blush.png 会静默落空槽——这正是本工具抓出来替你改掉的错。 Q: 28 张必须集齐吗？ A: 不必。缺的槽位在酒馆里聊天时留空，不报错，覆盖度条只陈述不催命。真正建议必备的是 neutral——没有情绪命中时兜底显示的就是它。同一个情绪有几张图也没问题：按官方多图规则命名成 joy.png、joy-1.png……工具自动排。 Q: 为什么我的 zip 导入了但酒馆里一张立绘都不显示？ A: 八成是两个原因之一：图片被套在 zip 的子文件夹里（酒馆的 sprite pack 导入只读顶层——本工具恒平铺打包），或者文件名和 28 个标签对不上（必须全小写、拼写精确）。这两件事在下载按钮出东西之前都已经校验过了。 Q: 这个工具会帮我生成表情差分吗？ A: 不会，而且是故意的。这是给你已有素材的打包工坊——自己画的、约稿的整套差分、有权使用的素材。它做校验、改名、试穿、打包，不发明任何像素，也不上传任何内容。 Q: 透明底体检查的是什么？ A: 立绘是贴在聊天背景上显示的，图片自己的底必须是 alpha 透明。工具把每张图解码后检查：JPG 天生没有透明通道（直接标出）；PNG 里一个透明像素都没有的标「不透明」；四角不透明的给较轻的提醒——然后你可以到试穿间亲眼看效果。 ### 谷子工坊网页版 — 把喜欢的角色装裱成收藏卡和书签 URL: https://foreverse.app/zh/labs/goods-studio Q: 稀有度怎么定的？ A: 按这张卡真实的社区记录，不是抽卡：身价分 = 下载 ×1 + 点赞 ×2 + 收藏 ×2，数据来自社区公开接口的静态快照。分档按快照的真实分布定：UR 要 4 分以上（96 张在售卡的 top 3%）、SSR 2 分、SR 1 分、R 是零分。阈值和快照日期就印在徽标旁边。 Q: 为什么我传的图是「未评级」？ A: 因为没有可评的数据。这里的稀有度指社区身价——一张公开发布的卡挣来的下载和点赞。你自己的图没有这份记录，硬编一个档位会让每个 UR 都变得不值钱。装裱照常出，用中性配色。 Q: 我的图会被上传吗？ A: 不会。画布完全在你的浏览器里绘制，PNG 也在本地合成——图片不出你的设备。也正因为我们检查不了，工具才要你自己确认有这张图的使用权。 Q: 导出图为什么带水印？ A: 左下角一行「卡名 · foreverse.app」小字，是网页免费版的对价——每张流出去的图都是署了名的明信片。App 内的谷子工坊可以关水印，而且稀有度按你和角色的真实关系算（聊天轮数、好感度），不看公开数据。 ### 开场白预览器 — first_mes 在聊天气泡里长什么样，写完当场看 URL: https://foreverse.app/zh/labs/greeting-preview Q: 沉浸模式和线上聊模式有什么区别？ A: 沉浸模式是酒馆式整段渲染：一条消息一个气泡，段落、动作描写、引号对白都在一块。线上聊模式把同一条消息切成微信式连发：按换行拆条，单段长文按句末标点再拆，最多 8 条（超出合并进末条）。写开场白时两种形态都该看一眼——大段景物描写在 IM 模式里会变成一串短消息，节奏完全不同。 Q: 拆条数是怎么算的？ A: 和 App 的线上聊引擎同一套规则：先按换行切并丢掉空行；如果只切出一条且超长（中文 60 字、英文主导 160 字符），再按句末标点切——英文句号有防误切守卫，Mr.、3.14、J. K. 这些点不会被当成句尾；超过 8 条时前 7 条保留、剩下合并进最后一条。所以这里显示 6 条，App 里就是 6 条。 Q: 开场白多长算太长？ A: 超过约 3,000 tokens 会亮黄牌。开场白不是一次性成本——酒馆系前端每轮请求都可能带着它，太长的开场白等于每轮都在挤占对话记忆的预算。想给足设定，考虑把世界观搬进世界书按需触发，开场白留场景和第一句话。 Q: 美化卡的 HTML 效果能预览吗？ A: 不能。这里预览的是文本渲染层：宏展开、动作斜体、引号着色、代码围栏、IM 拆条。美化卡靠正则脚本把 marker 变成 HTML，那层加工要在 App 里看。 ### 图片转可编辑 PPT · 图层重建工坊 — 截图重建成真文本框，浏览器本地完成 URL: https://foreverse.app/zh/labs/image-to-pptx Q: 为什么别的「图片转 PPT」转出来根本不能改字？ A: 因为多数工具没有做转换——只是把你的图原样贴进一页空白幻灯片，页数对了，内容还是一张平面图。重建才是难的部分：要 OCR 找到文字、版面分析把词聚成块、背景修补把原来的字擦掉、再按 OOXML 格式写出真文本框。这条管线在这页里完整跑一遍，而且跑在你自己的浏览器里——这也是它可以免费不限量的原因。 Q: 什么样的图效果最好？ A: 幻灯片截图、网页截图、海报、图文卡这类「设计稿」输入：标准字体、纯色或渐变底。在我们的标定样张上，平底上的文字擦得几乎看不出痕迹，字号估算误差大多在 10% 以内。照片是诚实的弱项：复杂纹理上的字不硬擦（文本框直接叠在原图上，块上有标注）；低对比的字幕——比如亮天空上的白字——可能整个识别不到。 Q: 识别错了字怎么办？ A: 点那个块，导出前直接改——这正是「重建」和「贴图」的区别。一张带错字的平面图只能重做，一个带错字的文本框改起来两秒钟。每个块的文字、字号、颜色、粗体都能在这页上编辑。 Q: 图片会被上传吗？ A: 不会。识别引擎（tesseract.js，Apache-2.0 许可）第一次用时下载进你的浏览器（约 6–10 MB，之后有缓存），从识别到最终生成 PPTX 文件，每一步都在你的设备上完成。这个页面没有上传接口。也因为计算发生在你这边，我们才能免费、不限次数、不加水印。 Q: 能导出分层 PSD 吗？ A: 还不能。PSD 出口（背景层 + 真文字图层，走 MIT 许可的 ag-psd 写入器）和这页共用同一套引擎，排在第二个出口做。先发 PPTX，是因为搜「图片转 PPT」的人要的就是一份能接着改的幻灯片。 ### 代言诊断器 — 查出机器人为什么总替你说话、替你做决定 URL: https://foreverse.app/zh/labs/impersonation-checker Q: 机器人为什么总替我说话、替我的角色做决定？ A: 因为语言模型是模式复读机，而模式多半就埋在卡里。开场白叙述「你」做了什么、示例对话替你那一侧写了台词、描述字段规定了 {{user}} 是什么性格暗恋着谁——模型把这些全读成「这是一场一个人演两角的戏」的示范。JanitorAI 官方帮助中心把代言列为最常见、最恼火的问题之一，而真正有效的攻略全部从同一步开始：审卡。这个页面把那一步自动化了。 Q: 检查时我的卡会被上传吗？ A: 不会。解析器和检测引擎都在你的浏览器里跑——PNG 数据块、JSON、charx 压缩包全部本地解码，扫描是对文本做的一组正则匹配。任何内容都不发往服务器，页面加载完断网也能用。角色卡是私人物件，这正是全本地的意义。 Q: 第二人称叙述（「你」）就一定是病吗？ A: 不是——这个检查器正是围绕这条边界造的。「她看了你一眼」没问题：你是宾语，观察这个动作属于角色。「你睁开眼，身处石牢」是跑团卡的合法开场——它搬运的是你的感官，且只搬一次。真正破坏扮演的是「你」坐在主语位上、后面跟着行动/说话/内心动词：「你决定」「你结结巴巴地说」「你感到心动」。引擎查的是语法位置，不是那个字。 Q: 卡是干净的，机器人还是抢我的戏，怎么办？ A: 那病根就不在卡上，这个工具也诚实地看不到卡以外的部分。接下来查三处：你的用户人设（堆满性格词的人设会逼模型表演你——只留外观事实）；你的回复长度（机器人长篇后只回一行，留下的真空模型会用「写你」来填）；以及模型本身（有的模型天生爱代言，把采样参数归零或换模型能解决顽固病例）。另外检查卡里有没有「禁止代替{{user}}发言」这类否定式指令——否定表述是掷硬币，有时反而更糟。 ### JanitorAI ⇄ 酒馆卡格式转换器 — bot 文本与 chara_card_v2/v3 双向互转 URL: https://foreverse.app/zh/labs/janitor-card-convert Q: JanitorAI 没有导出按钮，我到底贴什么进来？ A: 贴你能看到的。自己的 bot 打开创建页，把每个框（Personality、Scenario、Initial Message、Example Dialogs）分别拷出来；或者把 Personality 整段贴进本页的大框—— / / 这类容器会被自动拆进对应输入框。别人的私密定义 bot 是平台刻意藏起来的，任何转换器都变不出来。 Q: 为什么 JanitorAI 的 Personality 转过来进的是 description 而不是 personality 字段？ A: 因为两边的「正文主体」住的位置不同。JanitorAI 官方帮助明说 Personality 框装的是 bot 的绝大部分信息——性格、外貌、背景故事、甚至示例对话；而酒馆卡里承担这个角色的是 description，personality 只是个短摘要位。正文对正文，prompt 的重量才落在作者原本放它的地方。 Q: 1,500 token 体检的口径哪来的？ A: JanitorAI 自己的帮助中心：token 指南写明「大多数好 bot 的永久 token 低于 1,500 时表现最好，过 2,000 已是薄冰」，创建页概览另给 2,500 上限。本页按 Personality + Scenario（创作者可控的永久块）用全站同一把字符占比尺估算，并明说 JanitorAI 自家分词器数出来会有出入。 Q: 世界书转 JanitorAI 会怎样？ A: 列进「装不下的字段」，不是静默丢掉。JanitorAI 创建页没有世界书框——它的世界书是单独入口，而且方言坑不少（导入条目概率常被卡成 1%）。搬世界书之前先过一遍世界书迁移体检器。 Q: 内容会被上传吗？ A: 不会。拆箱、token 估算、PNG 嵌入全部在浏览器本地完成，页面加载完断网也能用。 ### 世界书触发模拟器 — 在线测世界书条目为什么没触发（关键词/递归/概率全解释） URL: https://foreverse.app/zh/labs/lore-trigger Q: 为什么我的世界书条目没触发？ A: 最常见的四个原因，这个工具都会直接标出来：①关键词没出现在扫描窗口里——默认只扫最近 4 条消息，关键词出现在更早的楼层不算数；②副关键词（secondary keys）条件不满足——AND_ALL 要求全部出现，NOT_ANY 要求全都别出现，方向经常被记反；③条目被禁用或是没有关键词的死条目；④概率没掷中——带 30% 概率的条目十次里七次不触发是正常行为。 Q: 递归触发是什么意思？ A: 条目 A 被对话触发后，A 的内容本身会进入下一轮扫描——如果 A 的正文里出现了条目 B 的关键词，B 也会被激活，这就是递归。写世界书时可以故意利用它做「设定链」：聊到灯塔就带出灯塔的看守人，再带出看守人的秘密。这个工具会把递归链画出来（「由某某条目的内容触发」），这是肉眼在酒馆里看不到的信息。 Q: 模拟结果和真实聊天里会完全一致吗？ A: 扫描和激活的判定一致（关键词匹配、大小写、整词、secondary 逻辑、递归、概率、组内互斥都按 SillyTavern 1.18 语义移植）。但真实运行时还有三样这里不模拟：token 预算裁剪（激活太多时低优先级条目会被挤掉）、注入位置的先后顺序、跨轮的 sticky/cooldown 状态。带这些设置的条目会打上提示徽标。 Q: 我的世界书会被上传吗？ A: 不会。解析和模拟全在你的浏览器里跑，页面加载完断网也能用。这也是为什么它能随便喂几万字的世界书——没有服务器在中间。 ### 世界书迁移体检器 — 五平台方言对照与搬家陷阱诊断（酒馆 / Chub / Risu / Agnai / JanitorAI） URL: https://foreverse.app/zh/labs/lorebook-doctor Q: 世界书导进 JanitorAI 之后为什么不触发了？ A: 十有八九是那个看不见的 1% 默认值：导入的条目概率经常停在 1%，写得再好也大约一百条消息才触发一次。逐条打开把概率显式设好（关键设定 100），确认书已经挂到角色上并且点了发布，再检查大小写敏感或整词匹配有没有悄悄挡住关键词。这个工具在你搬家之前就把这些全标出来。 Q: 酒馆的世界书搬去 RisuAI 或 Agnai，到底会坏什么？ A: 最毒的都是语义层的，不是外观层的。Risu 的导入器不读启用开关（你禁用的条目会集体复活），还会丢掉副关键词的逻辑模式——NOT_ANY 黑名单反转成正向必要条件，和你写的恰好相反。Agnai 把 constant、selective、大小写这些标记原样存在文件里，但引擎从来不读：常驻条目不再常驻、排除条件消失、概率掷骰根本不存在。这些结论全部对着两家的开源代码核实过，体检会逐条点名受影响的条目。 Q: 这个工具能帮我转换格式吗？ A: 刻意不做。转换器已经到处都是（酒馆和 Risu 自带导入按钮），但没有任何工具告诉你转换途中什么被悄悄改了。这里是搬家前的体检：先跑一遍，把标红的条目修好或记下来，再放心去转。 Q: 支持哪些文件格式？ A: 酒馆 World Info 导出、Chub 世界书（WI+V2 双写格式）、RisuAI 导出（type: 'risu'）、Agnai 记忆书（kind: 'memory'）、JanitorAI JSON 编辑器条目数组、V2/V3 character book 与 lorebook_v3 文件、内嵌世界书的角色卡、NovelAI 世界书（识别+浅支持）。拖进来就行，格式横幅会告诉你识别成了什么。 Q: 我的世界书会被上传吗？ A: 不会。解析、体检、分歧实验台全在浏览器里跑，页面加载完断网也能用。 ### 世界书合并去重器 — 多本世界书合成一本，重复条目逐组对照，触发词撞车先亮牌 URL: https://foreverse.app/zh/labs/lorebook-merger Q: 它怎么判定两条条目是重复的？ A: 双判据。内容几乎一样（字符三元组相似度 ≥ 90%）直接归为一组，不管触发词写什么——「同一本书攒了两个版本」就是这种。内容中等相似（≥ 55%）时要触发词也重叠（较小 key 集的一半以上被覆盖）才归组，防止两条只是提到相同名词的不同条目被硬并。刻意不做的：同触发词但内容不相干的条目不算重复——它们进「触发词撞车」清单，由你来裁决。 Q: 导出的时候条目会被改写吗？ A: 来自酒馆或 Chub 导出的条目原样透传——depth、role、自动化 ID、inclusion group、甚至平台自带的 extensions 全部原封不动，只重排 uid 和 displayIndex。来自 Risu / Agnai / JanitorAI / NovelAI / 卡内嵌 book 的条目，按解析出的统一语义在酒馆条目模板上重建（卡内嵌 book 会从 extensions 里捞回 depth/role/automation_id）。导出卡片会告诉你两条路各走了多少条。 Q: 触发词撞车为什么只提示、不帮我改？ A: 因为共享触发词不一定是错。同一个词激活两条条目，可能是有意的分层（一条写城市、一条写政局），也可能是两本书各写了一份同样的设定——后者白花双倍 token 还可能自相矛盾。哪种情况只有你知道。工具把每个被共享的触发词和涉及条目都列出来；想让同 key 条目「一次只出一条」，在酒馆里给它们设同一个 inclusion group 即可。 Q: 支持哪些格式？ A: 酒馆世界书导出、Chub 世界书（WI+V2 双写）、RisuAI 导出（{ type: 'risu' }）、Agnai 记忆书（{ kind: 'memory' }）、JanitorAI JSON 编辑器条目数组、V3 lorebook、带内嵌 book 的角色卡、NovelAI lorebook——和「世界书迁移体检器」同一套解析层。合并产物统一是酒馆 World Info 格式：这个圈子事实上的通用语。 Q: 我的世界书会被上传吗？ A: 不会。解析、查重、撞车检查、导出全部发生在你的浏览器里，页面加载后断网也能用；刷新就什么都不剩了。 ### 酒馆宏预览器 — {{char}} {{user}} {{random}} 在线展开，带完整宏支持对照表 URL: https://foreverse.app/zh/labs/macro-playground Q: {{random:a,b,c}} 和 {{random::a::b::c}} 有区别吗？ A: 没有，两种分隔符都合法：双冒号优先，没有双冒号时按逗号切。选项里本身要带逗号时用双冒号形态。每次生成时等概率抽一个——这也是开场白做「随机开局」的标准写法。 Q: {{roll:d20}} 的完整语法是什么？ A: NdM 掷 N 个 M 面骰取总和：{{roll:2d6}} 是两个六面骰。N 省略默认 1（{{roll:d20}}），裸数字当面数（{{roll:20}} 等于 {{roll:1d20}}），还有个空格形态 {{roll 1d100}} 也认——检定类角色卡的高频写法。写错了（比如 {{roll:abc}}）不会报错，按原参数输出。 Q: 宏会互相嵌套展开吗？ A: 不会。展开是单遍的：{{getvar}} 取出来的值里就算带着 {{char}}，也原样保留不再展开——酒馆和 App 的引擎都是这个行为，防止套娃失控。这个工具的标注模式能直接看到「产物里的宏形态没有被二次展开」。 Q: 未识别的宏会怎样？ A: 原样发给模型——不报错也不消失。这正是「卡里写错宏名导致模型看到一堆花括号」的事故来源。这个工具会给未识别的宏画虚线下划线，写完卡先贴进来扫一眼，比开局后再发现便宜。 ### 叙事指纹 — 从「故事怎么搭」而不是「词怎么选」照一章小说的结构口音 URL: https://foreverse.app/zh/labs/narrative-fingerprint Q: 这就是 StoryScope 论文的方法吗？ A: 不是，页面开头就说清了。论文是让 LLM 通读每篇故事提取 304 个叙事特征——这没法免费跑在浏览器里。我们做的是其中 30 个核心特征里「会在字面上留痕迹」的子集：情绪命名 vs 身体化演出、回溯标记、嗅觉意象、主题说破、说书人腔——每条轴都先在自有带出处语料上验证过分离度才上线；投影不干净的轴放在「参考轴」区，原话写明。需要真读懂故事才能算的特征（支线数量、结局归因、道德模糊度）在论文面板里如实打了 ✗。 Q: 它能告诉我一章是不是 AI 写的吗？ A: 不能。它只数你这一章的结构读数落进了哪些标定带——人类带、AI 带、还是重叠区——数完就停。不给百分比，不下判决。两个诚实理由：标定语料很小（中文 12 段人类 + 两组 AI 参照），而且每条轴都可以刻意写过去。请按论文对待特征的方式对待它：可解释的证据，不是审判。 Q: 中文的人类带和 AI 带是哪来的？ A: 人类带：12 本有授权的真人网文各取中段约 6000 字。AI 参照分两组：9 段产线纪律书（LLM 在反 AI 味纪律下写的成书）和 9 段九个前沿模型的原始续写（同一本玄幻的 20 轮裸续写——书源单一，如实声明）。有意思的是有几条轴上这两组 AI 方向相反，所以那些轴我们不下方向结论，把三组带都画给你看。全部数字由你浏览器里跑的同一份代码算出。 Q: 我的文字会被上传吗？ A: 不会。整个分析就是一遍正则扫描，在你的浏览器里完成，30 万字也是亚秒级。不下载模型、不发请求、断网可用。 ### 小说转世界书骨架器 — txt/epub 实体统计，导出酒馆 World Info 骨架 URL: https://foreverse.app/zh/labs/novel-to-lorebook Q: 它会替我把条目写全吗？ A: 不会，而且是故意的。条目血肉需要真的读过这本书的人来写：本页零 AI 调用、零编造设定。它产出的是世界书里一切机械的部分——有哪些实体、触发词是什么、出现在哪些章、出现多少次、每条建议多少 token。复制出的「补全任务书」自带写作规程，粘给任何 AI 或 Foreverse App 里的管家 Agent 都能对照原文接着写。 Q: 不用 AI 怎么找出实体的？ A: 规则信号，诚实计数。中文书：说话人模式（说道/问道/笑道前面的 2–4 字名）、地名与组织后缀（城/镇/山/宗/学院……左侧功能字剥掉再算）、引号术语（「青冥诀」这类）。英文书：句中出现的大写词组，配说话动词判定人物。频次不过门槛的一律不收（说话人 2 次、其余 3 次）——这个工具宁可漏掉，也不发明。 Q: 导出的 JSON 拿去哪用？ A: 标准的 SillyTavern World Info 文件——字段族就是 ST 1.18 自己写的那套。导进酒馆的世界书面板（或 Foreverse），每条骨架条目带着触发词就位，把模板文字换成真设定即可。表格里没勾选的实体不会进文件。 Q: 为什么漏了一个我知道很重要的角色？ A: 三个常见原因：这个名字从没出现在说话人模式里（沉默角色）、这本书的分章样式骗过了识别器（章节范围会退化到第 1 章）、或者频次没过门槛。这是规则抽取的诚实代价——在导出的 JSON 里手动加一条就是，格式一眼能看懂。 Q: 书会被上传吗？ A: 不会。txt 解码、epub 解包、章节识别、实体统计、JSON 组装全部在浏览器本地完成，页面加载完断网也能用。 ### PNG 卡图嵌入/提取 — JSON 和 PNG 角色卡在线互转，浏览器本地完成 URL: https://foreverse.app/zh/labs/png-card-tool Q: 为什么卡图被截图或转发之后就导入不了了？ A: 人设数据存在 PNG 的文本块（chara / ccv3）里，不在像素里。截图只重画像素；聊天软件和社交平台以「图片」方式转发时会重新压缩，同样把文本块洗掉。图还是那张图，卡已经没了。卡图要以「文件」方式传；数据丢了的卡，用这页的嵌入功能拿 JSON 重新做一张就行。 Q: chara 和 ccv3 数据块是什么？ A: 酒馆系约定的两种 PNG tEXt 文本块，内容是 base64 编码的卡 JSON。chara 是 V2 时代的经典键名；ccv3 是 chara_card_v3 规范的键名，两个都在时读取方优先读 ccv3。本工具恒写 chara，卡声明为 V3 时同时写 ccv3——与主流导出器的双写惯例一致。 Q: 这里生成的 PNG 卡，酒馆和其它前端能读吗？ A: 能。产物是标准 PNG，文本块落在 IHDR 之后、图像数据之前——SillyTavern、RisuAI、Foreverse 的读取方都是遍历全部数据块，落位在哪不影响读取。我们做过逐字节往返验证：嵌入再提取，JSON 一字不差。 Q: 我的卡和图会被上传吗？ A: 不会。数据块解析、base64 编解码、CRC32 校验、JPG 转 PNG 全部用浏览器标准能力在本地完成，页面加载完断网也能用。私人卡不出你的浏览器。 ### 酒馆预设查看器 — 把预设 JSON 变成人能读的配置单 URL: https://foreverse.app/zh/labs/preset-viewer Q: 为什么有些块标着「动态占位」没有内容？ A: 那是 marker 块——chatHistory、charDescription、worldInfoBefore 这类占位符。运行时前端会把角色卡字段、触发的世界书条目、聊天历史填进去。预设文件里本来就没有这些文本，所以查看器解释每个位会被什么填充，而不是把它当空块报错。 Q: 预设里的 prompts 比列表里的块多，去哪了？ A: 只有排进 prompt_order 的块才生效——这是酒馆语义不是查看器的限制。定义在池子里但没排进编排的块是死块，查看器在尾注里计数，不混进注入列表。 Q: 块上的 @深度 是什么意思？ A: 绝对注入的块不按列表位置生效，而是插进聊天历史倒数第 N 条的位置。预设作者用它放「必须贴着最新对话」的提醒。@深度 2 = 倒数第二条消息的位置。 Q: seed: -1 有作用吗？ A: 没有——-1 表示随机，酒馆系前端只在 seed 大于等于 0 时才发送它。固定的 seed 会钉死采样：同样的输入得到几乎相同的回复，重摇基本不变。查看器会给固定 seed 打黄牌，因为它是「怎么摇都一样」的常见病因。 ### 模型 API 价格对比 — 30 家供应商 1,300+ 模型牌价一张表 URL: https://foreverse.app/zh/labs/price-table Q: 价格数据从哪来、多新鲜？ A: 来自公开牌价库 models.dev 的快照，表底标注采集日期（当前 2026-07-24）。我们服务端的计费同步用同一个来源。促销价、协议价、区域价差不在表内；供应商随时可能改价，下单前以各家官网定价页为准。 Q: 为什么有的模型「缓存读」一栏是空的？ A: 空缺（—）表示该供应商没有公开缓存计价，不代表价格是零。有公开缓存价的家，缓存读价通常是全价输入的几十分之一——长对话场景里这一列比输入价还重要。 Q: 「聚合网关」是什么，为什么默认不显示？ A: OpenRouter、Poe、302.AI 这类转售平台：同一个模型会在多家出现，价格是转售价不是官方牌价。默认视图只看一方直连供应商，比价干净；想比转售价，点开「含聚合网关」筛选。 Q: App 预置 62 家供应商，表里为什么只有 30 家？ A: 其余 32 家要么是本地自托管（Ollama 这类，没有牌价），要么没进公开价格库。没有数据就不编数据——这 30 家覆盖了绝大多数人实际会买的模型。 ### 酒馆 Prompt 装配透视台 — 预设+角色卡+世界书，还原这一轮 prompt 是怎么拼出来的 URL: https://foreverse.app/zh/labs/prompt-assembly Q: 为什么我的预设块像没生效一样？ A: 最常见的四种，这个工具都会直接标出来：①块定义了但没排进 prompt_order——按酒馆语义它完全不生效，透视台把它列在「预设块诊断」里；②块在编排里被停用；③块被设成 @深度注入（injection_position=1），它不在主干位置、而是插进聊天历史的倒数第 N 条，位置跟你想的不一样；④main 块被角色卡的 system_prompt 悄悄替换了——这是酒馆默认开启的行为（prefer character prompt），透视台会给替换后的 main 打上「卡覆盖」徽标。 Q: 世界书触发了，为什么内容还是没进 prompt？ A: 触发和送达是两回事，中间还有四道坎：预设把 World Info 占位停用了（内容整个丢弃）；占位根本不在编排里（内容会被追加到 prompt 最末尾——酒馆的真实行为，多半不是作者想要的位置）；@深度词条注入后在 token 装填时被裁掉；编排里没有 Chat History 占位（@深度注入连落点都没有）。透视台给每条词条一个下落报告：进了哪、卡在哪一道。 Q: 「@深度注入」到底插在哪？ A: 从最新一条消息往回数：深度 0 贴在最末，深度 2 隔两条聊天消息，以此类推——注入消息本身不占深度位。同深度、同 order 组、同角色的多个注入会合并成一条消息（预设块在前、世界书/作者注在后）；不同 order 组分开，数值大的组排前面。这套规则全部按 SillyTavern 1.18 的注入代码复刻，你可以在装配结果里看到每条注入的实际落点。 Q: 模拟结果和真实发送会完全一致吗？ A: 主干语义一致：prompt manager 编排顺序、marker 填充、卡覆盖 main/jailbreak、世界书全部注入位（before/after/@深度/示例区/作者注位/outlet）、宏展开、token 预算装填顺序，都按 SillyTavern 1.18 源码逐函数核定。有几层刻意不模拟：世界书自身的预算裁剪（默认上下文 25%，超水位会给提示）、正则脚本、跨轮的 sticky/cooldown、群聊、continue/impersonate 等特殊生成类型、第三方扩展注入。token 数是字符占比估算不是真分词器。完整清单见下方「它做不到什么」。 Q: 我的预设和卡会被上传吗？ A: 不会。解析、词条扫描、装配全在你的浏览器里跑，页面加载完断网也能用。这也是它敢让你直接拖整本世界书进来的原因——没有服务器在中间。 ### 网文文本分析仪 — 句长节奏、对话占比、复读，一段正文的文字体质 URL: https://foreverse.app/zh/labs/prose-stats Q: 「句长错落度」是什么？为什么说它是 AI 味的第一信号？ A: 它是句长的变异系数——句长标准差除以均值。人写小说会把三个字的短句和四十字的长句混着用；机器续写则向中间收拢。我们的九模型二十轮续写实验里，人写原著稳定在 0.64 上下，裸模板 AI 一轮比一轮均匀、掉到 0.52。结构上的均匀是第一个露馅的信号，所以这页把它放在最前面。 Q: 为什么这页不给文本打分？ A: 因为诚实的输出是「描述」不是「评级」。对话占比低不代表书不好看，节奏跳跃也不自动等于文笔好。参照区间告诉你这段文字在 378 个真实章节样本里坐在什么位置——坐在那个位置意味着什么，是读者的判断，我们不越权替你下。 Q: 和 AI 味体检器有什么区别？ A: 层面不同。AI 味体检器查词汇层——套话和句式模具（「眼底闪过一丝」「带着几分」）；这页看骨架层——句长分布、对话份额、段落形态、逐字复读。一段文字可能过了一边、栽在另一边；两页都跑一遍，同一段话就有了词汇和结构两个视角。 Q: 参照区间是哪来的？ A: 来自 10 本自有中文小说的 378 个章节样本，2026-07-24 用与本页完全相同的句子与对话规则离线计算。这套语料里，句长错落度的中间带（P25–P75）是 0.62–0.72，对话占比的中间带是 22%–40%。它们是「我这段坐在哪」的参照点，不是质量及格线。 ### txt 在线阅读器（网页版）— 上传 txt 立刻翻页试读，不用装 App URL: https://foreverse.app/zh/labs/reader-demo Q: 为什么只能读前三章？ A: 演示机是「上手摸一摸」，不是阅读 App。窗口钉在前三章，30MB 的大书才不会把页面 DOM 压爆，承诺也才诚实：整本书——带进度记忆、续写分支、听书——在 App 里读。试读窗口在哪结束，页面上明说。 Q: 我的文件会被上传吗？ A: 不会。读文件、解编码、识别章节全部发生在你的浏览器里，内容不出这台设备、也不留存——刷新页面书就没了。这也是演示机记不住阅读进度的原因。 Q: 章节识别错了：整本一章，或者切在奇怪的位置。 A: 章节识别移植的是生产导入器的规则主干（第N章族、数字标题、Chapter N、序章后记标记）。少见的标题样式可能骗过它——App 的完整导入器覆盖更多样式，也支持手动分章。完全识别不出章节时，演示机按前 1.5 万字单章试读，并且如实标注。 Q: 翻页是真分页还是滚动截图？ A: 真分页：正文按固定页宽的多列排版，页在字真正断开的地方断——调一下字号，页数会重新测量。翻页动画用的是我们手机模拟器引擎的滑动模式；点左右页缘翻页，点中间呼出阅读工具条。 ### 酒馆正则测试器 — 在线调试 SillyTavern 正则脚本，实时看替换结果 URL: https://foreverse.app/zh/labs/regex-tester Q: 为什么我的酒馆正则不生效？ A: 先看三个字段：①placement——脚本声明了作用于哪条链路（用户输入 / AI 输出 / 世界书…），测试语境不匹配就整条跳过，这个工具的「测试语境」下拉能直接复现；②disabled——导出的合集里经常混着被作者关掉的条目；③markdownOnly——它只改显示层，你在提示词检查器里看上行内容当然找不到变化。三个都对了还不生效，多半是正则本身没命中，看左边替换前面板的高亮就知道。 Q: findRegex 里的 /pattern/gi 是什么写法？ A: JavaScript 的正则字面量形态：斜杠中间是模式，斜杠后面是标志（g 全局、i 忽略大小写、m 多行、s 让 . 匹配换行）。酒馆网页版原生就是这个方言，这个工具与它完全一致。不带斜杠的裸模式也认，按全局替换处理。 Q: replaceString 里能写什么？ A: 三类东西：$1、$2 数字反向引用和 $<名字> 命名组（引用不存在的组会替换成空串）；{{match}} 代表整段命中文本；还有 {{setvar}}、{{incvar}}、{{getvar}} 这些变量宏——美化卡靠 {{incvar}} 给选项编号就是这么做的。这个工具全部支持，跨多处命中的变量递增也和运行时一致。 Q: 什么样的正则会把聊天界面卡死？ A: 嵌套无界量词，比如 (a+)+ 或 (.*)*：在长文本上会指数级回溯，一条就能把界面钉死几分钟。这个工具用和 App 相同的结构守卫在编译前拦下它们，执行层还有 350 毫秒超时熔断兜底——被拦下时会告诉你原因，方便改写。 ### 复读与采样诊断器 — AI 为什么车轱辘话，哪个参数才真治它 URL: https://foreverse.app/zh/labs/repetition-doctor Q: 我的聊天记录会被上传吗？ A: 不会。检测、分诊、参数体检全部在你的浏览器里跑——贴文字或拖 .jsonl，什么都不离开你的设备。这也是为什么页面不会「越用越准」：规则是固定且带版本的，和我们自己评测管线用的是同一套。 Q: 照推荐档改参数，就一定不复读了吗？ A: 不一定，我们也不装。采样是按模型各有脾气的玄学；这里的参考区间逐条标注来源（官方文档 / 社区共识 / 本仓实测），定位是「一次只改一个」的排查起点。真正有硬证据的是诊断那一半——复读在哪一层、环是不是已经住进了你的历史。 Q: 为什么自救清单总把「先改历史」排在调参前面？ A: 因为测过。本仓 App 侧实验里，坏范例一旦躺进可见历史，纠正指令 9/10 轮无效；而新开会话零残留。模型模仿转录的力度大于服从指令：参数只影响下一个 token，历史却在教下一整条回复。所以历史手术排在一切旋钮之前——这个顺序是有实验依据的。 Q: 「12-token 滑窗」到底是什么？ A: 把文本切成每 12 个连续 token（中文一字一个、英文一词一个）的滑窗，数一条回复的滑窗有多少在更早的回复里出现过。这正是我们小说评测榜复读数字的口径。作为标定：本页引擎在归档的 Grok 4.5 链上复算出与发布数字一致的结果——第 10-12 轮 100% 旧文，第 13 轮回落（发布 68.4%、复算 67.7%，分词口径差不到 1 个点）。 ### 网页开演 — 拖一本 txt，当场演成视觉小说 URL: https://foreverse.app/zh/labs/story-theater Q: 为什么有的台词没有名牌、以旁白念出来？ A: 这是刻意设计。规则编译器只在有说话从句证据时才挂名牌——「林夏说：」「“……”秦朝答道」这类形态。证据不足的引语宁可进旁白，也不硬猜：错挂一次名牌，读者对整场演出的信任就没了。我们在一本真实网文上人工标注过归因真值，这套规则的出口门是强归因精确率 100%，代价就是有一部分台词弃权。页面报告里的「归因率」如实显示这个比例。 Q: 我的书传到服务器了吗？ A: 没有。解码、章节识别、角色发现、场景编译全部在你浏览器的一个后台线程里跑，页面不发出任何携带书内容的请求。断网状态下（页面加载完成后）一样能用。分享按钮分享的是工具链接或公版演示书的位置，不含你的书。 Q: 角色名单是 AI 猜的吗？ A: 不是。名单从说话从句统计而来：一个名字要在「谁说道」的主语位上稳定出现至少两次才进名单，受话方（「对秦朝说道」里的秦朝）、代词、泛称（众人、有人）都会被排除。「宝玉」和「贾宝玉」这类同人异名会自动合并。统计错了可以在报告页手动删改，名单一改，演出立刻按新名单重排。 Q: 只有 epub 怎么办？ A: 这一版只吃 txt。epub 是个 zip 容器，在浏览器里做一半的解包会给出比诚实拒绝更糟的结果。桌面阅读器（Calibre 等）都能把 epub 另存为 txt；反过来想把 txt 打包成 epub，站内的 epub 制书间可以做。 Q: 章节识别错了、或者整本书被当成一章？ A: 章节规则与站内「txt 体检」工具同一套（生产导入器的移植）：第N章家族、数字标题、Chapter N、序章后记这些形态都认。一本书如果全无可识别的标题行，这里会按行数切成片段演出，并在报告里说明。想单独诊断章节问题，去 txt 体检页看完整报告。 Q: 跟 Foreverse App 里的剧场是什么关系？ A: 同一套场景编译规则。网页版是素演出：无立绘、无配乐、演的是你书里已有的文字。App 里的剧场是完整形态——演到分岔处可以选走向，AI 从那一步现场续写成新分支，场景还能配图。网页版章尾的入口指向它。 ### 文字 X 光 — 逐词困惑度透视，浏览器本地跑小模型给每个词标「好猜程度」 URL: https://foreverse.app/zh/labs/text-xray Q: 我的文字会被上传吗？ A: 不会。点「扫描」后浏览器从 CDN 下载一个 73–81MB 的量化小模型，之后推理全部发生在你设备的浏览器里（Web Worker + WASM），文字不发给任何服务器。断网状态下（模型已缓存时）照样能扫。 Q: 它能判断一段文字是不是 AI 写的吗？ A: 不能，我们也不装能。我们用 100 段带出处的语料做了标定：在探针小模型的量尺下，真人写的套路网文（top-10 命中率中位 69%）反而比大模型直出的续写（57%）更「好猜」——「困惑度高低」和「谁写的」在 2026 年已经对不上号。这个工具给你的是逐词纹理，不是判决。 Q: 为什么要下载 70 多 MB 的模型？ A: 因为你的浏览器里真的在跑一个神经网络：中文轨是 6 层蒸馏版 GPT-2 中文模型（uer/gpt2-distil-chinese-cluecorpussmall 的 q8 量化 ONNX，73MB），英文轨是 DistilGPT-2（81MB）。只下载一次，之后存在浏览器缓存里；引擎初始化每次打开页面约 10–30 秒（一次性），扫描本身不到一秒。 Q: 和 GLTR 是什么关系？ A: 致敬加接棒。GLTR 是哈佛 NLP 与 MIT-IBM Watson 实验室 2019 年的工作：用 GPT-2 给每个词按预测名次着色，是「AI 检测可视化」这个形态的开山。它停在 2019——服务端运行、只有英文、自己声明对新模型可能失效。这里是同一个思路的 2026 重做：跑在你的浏览器里、中英双轨、并把「对新模型失效」这件事用我们自己的标定数据摆在台面上。 ### Token 计数器 — 估算实时称重，精确模式逐 token 复算（OpenAI o200k） URL: https://foreverse.app/zh/labs/token-counter Q: 这是真分词器吗？ A: 分两档。默认档是字符占比估算：中文约 1.6 字/token、英文约 4 字符/token——DeepSeek 官方文档给的口径，全站工具同一把尺。切到「精确模式」后，对 OpenAI 系是真分词器：o200k_base 词表是公开的（GPT-4o 之后各代），浏览器本地逐 token 复算、边界直接染色给你看，词表约 2.3MB 首次用时才下载。DeepSeek / Claude / Gemini 没有公开分词器——那些家族仍然只有估算加经验区间，这是能力边界，页面不装。 Q: 我的文字和卡会被上传吗？ A: 不会。计数和卡解析全部在你的浏览器里完成——文字不离开输入框，卡文件用浏览器自带能力本地解码。页面加载完断网也能照常用。 Q: 为什么别的工具算出来的数不一样？ A: 三个常见原因：它跑的是另一个分词器（家族间差两三成，中文文本差得更多）；它算的字段不一样（有的工具把作者备注也加进去，但那个字段根本不进 prompt）；它对世界书的算法不一样（常驻条目每轮计费、关键词条目命中才计）。本工具把这三件事全部拆开单列，而且和本站角色卡体检器用同一把尺——站内工具之间的数字可以直接对。 Q: 角色卡里哪些部分算「常驻 token」？ A: 每一轮请求都要原样重发的那些：description、personality、scenario、system_prompt、post_history_instructions，以及世界书的 constant 条目。开场白只出现一次（但一直躺在历史底部）；示例对话在上下文吃紧时最先被丢；creator_notes 压根不进 prompt。字段表给每一行都标了计费身份。 Q: 一个 token 等于几个字？ A: 本页用的口径：中文约 1.6 字/token，英文约 4 字符/token。反过来算，100 token 大约是 160 个汉字，或 400 个英文字符（七十来个单词）；同样 1500 字的文本，中文要 ~900 token，英文只要 ~375。这个不对称就是同一张双语卡翻译后重量能翻倍的原因——也是本页所有数字都给「家族区间」而不是给一个假精确值的原因。 ### 人味图灵场 — 每天一题，猜哪段是机器写的 URL: https://foreverse.app/zh/labs/turing-arena Q: 题目是哪来的？每天怎么换？ A: 两类语料，版权都在我们自己手里：一类是官方产线小说的选段（AI 初稿，但整卷过了结构校验、评分闸和人工抽读三道把关才上架）；另一类是九模型续写横评（2026-07）的裸产出——同一本书、同一个锚点、同一条指令、同一个温度，九家模型各写各的。生成脚本把它们切成按天轮换的题库（2026-08 扩容后 97 题），按北京时间的日期确定性取题，0 点换题；往期题保留可玩，算练习不计连胜。题库轮完一圈后当天会重放往期题，页面上明确标注「往期精选重放」，不装新题。 Q: 「AI 评委只对了 12%」是怎么测出来的？ A: 2026 年 7 月我们跑的双盲可靠性实验：4 个不同基座的大模型当评委，对 7 对角色卡文案正反两个顺序各评一轮，金标来自真实读者的判断。结果评委之间 86% 一致，对金标却只有 2/16 = 12%——一致地错：它们把细节密集、打磨均匀的文本读成「真人写的」，而真实读者恰恰觉得那是 AI 味。这个数字就是你在这页要打败的对手。 Q: 为什么问的是「经过人工把关的成书文字」，不是「人写的」？ A: 因为官方小说的初稿也是 AI 写的——诚实的差别在流程：一边是校验、评分、人工抽读三道把关后的成书，一边是模型一次性输出、没人碰过的裸产出。真正的「人写 vs AI 写」题型需要我们有干净展示授权的人类文本，正在向社区创作者征集，上线时会明确标注。 Q: 我的战绩存在哪？有排行榜吗？ A: 只存在你浏览器的 localStorage 里，零账号、零上传。「全站今日正确率」这一期还没有（本页没有后端），先拿 12% 的 AI 评委分数当固定对照——这个数字有出处有日期。众包正确率在二期路线图上。 ### 图片转 SVG · 位图矢量化工作台 — 免费在线，浏览器本地运行 URL: https://foreverse.app/zh/labs/vectorizer Q: 我的图片会被上传吗？ A: 不会。解码、降采样、描摹全部在你浏览器的 Web Worker 里完成，页面加载完之后断网也能用。我们的服务器只负责发这个页面和引擎文件本身。 Q: 和 vectorizer.ai 这类收费工具差在哪？ A: 两条技术路线。这里是经典几何描摹（开源 VTracer 引擎）：免费、无限次、无水印、文件不出设备；付费头部用服务端 AI 模型做形状重建，照片类输入的还原度确实比我们好。转 logo、图标、插画、线稿，几何引擎就是正解，差距不存在；转照片想要「还原」而不是「海报风」，那是它们的地盘——我们如实说。 Q: 什么图转出来效果最好？ A: 平色的：logo、图标、贴纸、插画、漫画上色稿，以及印章和签名（用黑白档）。判断标准很简单——色块越平、边界越清楚，矢量化越干净。照片和重渐变的图会被压成色块海报风。 Q: 转出来的 SVG 为什么比原图还大？ A: 矢量文件的大小跟「形状数量」走，不跟像素走。照片类输入的每一小块颜色都是一条路径，几千条路径的 SVG 比 JPG 大很正常。把颜色精细度调低一两档、或换「照片（风格化）」预设，体积能掉一个量级。logo 类输入正相反——产物通常只有几 KB。 Q: 转出来的图能商用吗？ A: 你的图你做主：工具不碰文件、不加水印，产物不附带任何来自我们的授权条件。引擎本身是 MIT 开源许可。唯一要留意的是图片本身的版权——格式转换不改变原图的权利归属。 ### TTS 音色试听台 — 96 个音色同一句话，听完再选 URL: https://foreverse.app/zh/labs/voice-gallery Q: 这些试听样本是怎么生成的？ A: 2026-07-24 一次性批量生成：96 个音色各自走官方接口合成一条短句——阿里百炼 qwen3-tts-flash、xAI grok-voice-latest、OpenAI gpt-4o-mini-tts、SiliconFlow CosyVoice2，然后统一压成单声道 MP3 并做响度归一，单条 17–31KB。中文音色念同一句中文、英文音色念同一句英文——你听到的差别全部来自音色本身，不来自台词。 Q: 为什么有的音色可能没有样本？ A: 目录跟的是 App 内置的 96 音全集。重新生成样本时哪家接口调用失败或渠道没打通，那个音色就标「暂无试听 · App 内可用」，不拿别的声音顶替。本次生成 96 条全部成功；以后目录扩充或样本刷新时如果有失败，照这个规矩如实标出。 Q: 试听收费吗？App 里用这些音色听书怎么算钱？ A: 这页是静态文件，随便点，不花钱也不限次，也不会临时调用任何接口。App 里听书是真实合成、按量计费：官方渠道按各家牌价加五成服务费、按字符扣积分，注册送的 5000 积分够先听几章再决定；如果自带各家的 key（BYOK），走你自己的账单，我们不加价。 ### 新卡速报 — 最近上新的社区角色卡，一页看完 URL: https://foreverse.app/zh/labs/weekly-cards Q: 这些卡是哪来的？ A: Foreverse 社区的公开货架——目前全部由官方号发布，所以每张卡都带完整人设、开场白，多数还内嵌世界书。封面 AI 生成并带来源标记。页面读的是货架的静态快照，不从其它站抓卡。 Q: 看中一张卡，怎么拿到手？ A: 每张卡链到详情页，能看完整人设和开场白。想开聊就打开 Foreverse App 的社区页搜卡名，一键导入——世界书、示例对话随卡一起进来。卡全部免费。 Q: 为什么有的周没有新卡？ A: 卡按发布班车上架，不是固定周更。没有班车的周就不出现——页面从不用占位符凑数。最近一次上新的周和真实日期钉在页头。 Q: 能订阅吗？ A: 能——RSS 地址是 /labs/weekly-cards/feed.xml，列出最新的卡和详情页链接。任何阅读器都能订，不用注册、没有邮件列表。 ======================================================================== ## RESEARCH — Foreverse vs SillyTavern four-way benchmark (2026-07), headline findings URL: https://foreverse.app/research/foreverse-vs-sillytavern - Our own cache hit rate was 46.5% before the fix vs 91.5% for bare SillyTavern — same 15-round session, 2.8× the cost (measured 2026-07-23). Root cause: the position where we injected lorebook entries drifted every round, breaking the provider's ~2%-priced repeated-prefix discount. Fixed and merged the same night; re-verified at 89.9% / 88.3%, matching bare ST, with 58-63% lower cost (2026-07-24). - Bare SillyTavern hits 91.5% cache discount with zero plugins, straight out of the box — its prompt structure is naturally stable. That number was the bar we fixed our own app against, not something we beat. - Fully-modded SillyTavern (the community's most popular preset + helper plugin + table-memory combo) pays a “plugin tax”: 2.5× prompt size, 2.8× the total cost of bare ST, and 12 of 135 rounds (8.9%) returned empty upstream replies that were still billed (2026-07-23). That's the cost of the configuration stack, not a flaw in SillyTavern itself. - Deep-lore probe, 12 questions on a fully original worldbook card, deliberately avoiding trigger keywords: Foreverse Smart Play answered 12/12 with zero extra paid calls; bare ST 1/12, modded 2/12, Foreverse standard 5/12 (2026-07-23). The cost belongs in the same sentence: Smart Play sessions run about +35% (+18% measured after the cache fix). - 27-question memory exam after injecting a real 300-round chat (single run): bare ST scored 16 — steadiest on recent facts and honestly said “I don't know” on 3 of 4 never-mentioned questions; Smart Play scored 13 and was the only setup to answer facts from ~270 rounds back; Foreverse standard scored 10 but fabricated all 4 never-mentioned answers — a real product gap this benchmark caught in our own app (2026-07-24). - Blind-review leaning from calibrated AI judges: Smart Play > bare ST > Foreverse standard > modded ST (16:0 against our own standard mode, three judges agreeing). But the judges' qualification exam also produced a case where every judge was confidently wrong, so until 30 human spot-reads finish, this stays a leaning — not a conclusion (2026-07-24). - Group-chat crosstalk special (same three-character group × byte-identical 12-round script × 2 repeats): all four setups landed zero crosstalk on disk. The paths differ, though — our model actually attempted whole-turn impersonation 21 times at the generation layer, every one intercepted before persisting; on @-mention dispatch we scored 11/12 vs SillyTavern's 7/12 (2026-07-24). - Instruction-following special (six instruction groups, 24 judgment points × 2 repeats, machine first-pass + full human adjudication): Smart Play 94%, first of four; bare ST 90%, Foreverse standard 85%, modded ST 78%. Honest scope: single chat, single model — and our standard mode's weakness (it couldn't step out of character when asked) is baked into that number (2026-07-24). - The whole benchmark used ~1,530 real AI calls, roughly $1.07 at list prices; ~952 judge calls ran on a separate channel (2026-07-23/24). URL: https://foreverse.app/zh/research/foreverse-vs-sillytavern - 我们自己的缓存命中率修复前 46.5%，裸酒馆 91.5%，同样聊 15 轮花的钱是它的 2.8 倍（2026-07-23 实测）。病根是往提示词里塞世界书设定的位置每轮都在变，把服务商约 2% 计价的「重复前文」折扣打断了。修复当晚合入，复验 89.9% / 88.3% 追平裸酒馆，同量对话成本降 58%~63%（2026-07-24 复验）。 - 裸酒馆不装任何插件、出厂配置就能吃到 91.5% 的缓存折扣：它的提示词结构天然稳定。这一条是我们修复时对齐学习的对象，不是被比下去的对象。 - 满配酒馆（社区最流行的预设 + 助手插件 + 表格记忆组合）有一笔「插件税」：提示词体量 2.5 倍、总成本是裸酒馆的 2.8 倍，135 轮里 12 轮（8.9%）上游返回空回复但照常计费（2026-07-23）。这是配置组合的代价，不是 SillyTavern 本身的问题。 - 深设定 12 问（世界观完全自造的角色卡、故意绕开触发词提问）：智能演绎 12/12 全对且没有多花一次调用费；裸酒馆 1/12、满配 2/12、我们普通模式 5/12（2026-07-23）。代价要一起说：智能演绎每局会话成本比普通模式高约 35%（缓存修复后实测 +18%）。 - 300 轮长对话灌入后考 27 题（单次专项）：裸酒馆 16 分，近期记忆最稳、「没提过的事」4 题里 3 题诚实说不知道；智能演绎 13 分，四方唯一答对 270 轮前信息的一方；我们普通模式 10 分，但 4 道「没提过的事」全在一本正经编造，这是评测抓出的我们自己的产品差距（2026-07-24）。 - AI 评委盲评倾向：智能演绎 > 裸酒馆 > 普通模式 > 满配（对普通模式 16:0、三家评委方向一致）。但评委上岗考试里出现过「所有评委一致地错」，所以这个数字在 30 份人工抽读终裁前只作盲评倾向，不作结论（2026-07-24）。 - 群聊串台专项（同一个三人群 × 逐字相同 12 轮脚本 × 2 重复）：最终落盘四方全零串台。但来路不同——我们的模型生成层其实 21 次想整段冒充别的角色，全部被防线在落盘前拦下；@ 点名调度我们 11/12，酒馆侧 7/12（2026-07-24）。 - 指令遵循专项（六组指令 24 判定点 × 2 重复，机器初判+人工终裁）：智能演绎 94% 四方第一，裸酒馆 90%、我们普通模式 85%、满配 78%。口径如实说：单聊、单模型，普通模式的短板（说了「出戏聊两句」它出不了戏）也在这套数字里（2026-07-24）。 - 整场评测约 1,530 次真实 AI 调用，按官方牌价折算约 1.07 美元；另有约 952 次评委判卷调用走独立通道（2026-07-23/24）。 ======================================================================== ## RESEARCH — Fiction Bench: 9-model continuation leaderboard (2026-07), headline findings URL: https://foreverse.app/research/fiction-bench - There is no universal style model: the two genre crowns went to different models — DeepSeek V4 Flash won xuanhuan #1 yet slid to 6-7 on court romance, while sibling V4 Pro did the exact opposite (2-4 xuanhuan, 1-2 romance). Pick by the book you read. - The xuanhuan champion is one of the cheapest models tested: deepseek-v4-flash ($0.14/M input at list price) ranked #1 in both blind mappings; ten thousand characters of continuation cost about $0.03, roughly 1/40th of the most expensive model on the board. List price and style fidelity are uncorrelated. - Twenty-round runs exposed three long-run failure modes: restart loops (rewriting the opening every round — gemini-3.1-pro 20/20 under our old directive, 0/20 after the wording fix), mid-chain collapse (grok-4.5 looping two passages verbatim three times; kimi-k2.6 hitting 97.6% cross-round repetition), and ending rewind (gpt-5.6-terra dazzling for 17 rounds, then rewinding the plot at r18-20). None of these are visible in a single-shot trial. - Texture mimicry and long-run plot coherence are independent abilities: gpt-5.6-terra, the only model of nine to clone tilde onomatopoeia and the source's particle habits, crashed the xuanhuan marathon with an ending rewind — then ran the romance chain with zero accidents at ranks 1-3. - Structural metrics and human-likeness judgments diverge: qwen3.7-max scored closest to the source on sentence length, dialogue rate and stock-phrase density, yet blind-ranked only 5-6 — it writes scent description in every window of books that contain none. Statistically alike ≠ reads alike; this board therefore ranks by blind review, with metrics as corroboration only. URL: https://foreverse.app/zh/research/fiction-bench - 没有万能的文风模型：两个文体的第一名互不重叠——DeepSeek V4 Flash 拿玄幻榜第 1 却在女频滑到 6-7，同门 V4 Pro 恰好反过来（玄幻 2-4、女频 1-2）。选模型要按你读的书选。 - 玄幻榜冠军是全场最便宜的模型之一：deepseek-v4-flash（列表价输入 $0.14/M）双盲两套映射全第 1，按续写口径写一万字约 $0.03，是榜上最贵模型的约 1/40。牌价和文风复刻能力没有相关性。 - 20 轮长跑暴露了三种长程失效模式：从头复读（每轮重写开场，gemini-3.1-pro 旧指令下 20/20，指令措辞修复后 0/20）、中途卡死（grok-4.5 mid 整窗同两段逐字循环三遍、kimi-k2.6 跨轮重复率 97.6%）、尾段回卷（gpt-5.6-terra 前 17 轮惊艳、r18-20 剧情倒带重演第一章）。单轮试用全都看不出来。 - 纹理模仿力和长程剧情一致性是两种独立能力：gpt-5.6-terra 是九家唯一复刻波浪号拟声与「的/地」用字习惯的散文模仿天花板，却在玄幻长跑尾段回卷；女频轮它全程零事故拿 1-3。 - 结构指标和人味判断会分道扬镳：qwen3.7-max 的句长、对话率、套话密度全场最贴原著，双盲却只排 5-6——它每窗必写原著全书从不写的气味描写。统计像不等于读着像，所以本榜以双盲位次为准、结构指标只作旁证。 ======================================================================== ## COMPARISONS — verdicts and FAQs ### Tavo vs Foreverse: Which Phone Tavern App Should You Use? URL: https://foreverse.app/compare/foreverse-vs-tavo Updated: 2026-10-04 Verdict: Searching for Tavo usually means one thing: you want a tavern app that just opens on your phone. Foreverse is that too. Cards in V1–V3, lorebooks, regex, group chat, status bars, HTML/CSS/JS cards, image generation, translation: the same list. What Foreverse adds is real-time voice calls, a plugin center, vector memory, key rotation, and a reader for long novels. Tavo comes out ahead on price for big card libraries, a Mac app, and AI-picked speakers in group chats. Q: Is Tavo free, and is Foreverse? A: Tavo says every feature is free, and it connects straight to the API you set up, so you only pay your model provider. Foreverse is free to install with no subscription and no daily message cap, and with your own key every AI feature is open. The free tier holds three imported cards, books, and companions; a one-time Pro purchase removes that limit. The official model channel is pay-as-you-go credits, and signing up grants 5,000. Q: Can I move my cards between Tavo and Foreverse? A: Yes. Both read standard chara_card files (V1–V3 as PNG, JSON, or CHARX). Export cards and lorebooks as files from one app and import them into the other. Chat history formats differ between apps, so keep the old app if past chats matter. Q: Does Foreverse have voice calls? A: Yes. Open a companion, tap the phone icon, and talk. Replies stream back in real time, and you interrupt by tapping Skip or typing. It runs on your own xAI or Alibaba Bailian key, or on official credits billed per minute. There is a cheaper push-to-talk mode too. There are no video calls. Q: Does Foreverse have vector memory and plugins? A: Yes. Add an embedding model and the memory plugin recalls old chat by meaning, while lorebook entries marked vectorized activate by similarity. Built-in plugins cover long-term memory, auto-summary, plot options, reply ideas, state tracking, think-first replies, lorebook suggestions, and web lookup, and you can sideload your own plugin as a zip. SillyTavern extensions do not run here. Q: Do cards with HTML, CSS, and JavaScript work in Foreverse? A: Yes. Rich cards render in a sandboxed view; their scripts run after you tap Enable, and they cannot reach the network unless you allow it for that card. Scripts written for SillyTavern's TavernHelper interface work for the commonly used calls, which are a subset of the full interface. Scripts written for Tavo's TavoJS interface may not, so test one card before moving a whole library. Q: Which one is better for SillyTavern users? A: For plain card chat the two are close. Foreverse is the closer match if you rely on Quick Reply, STscript-style automation, vector memory, or key rotation, or if you also read novels, want voice calls, or want built-in image and video generation. Tavo is the better pick if you want everything free, a Mac app, or Tavo's own scripting interface. ### Foreverse vs MiniTavern: Which Phone Tavern App Fits You? URL: https://foreverse.app/compare/foreverse-vs-minitavern Updated: 2026-10-01 Verdict: MiniTavern is a focused, mobile-first tavern front end: import SillyTavern PNG or JSON cards, connect OpenAI, Claude, DeepSeek or a local model, and chat with no paywall on the core experience. Foreverse covers that and adds depth around it: .charx and V3 cards with embedded lorebooks that activate with the card, regex and Quick Reply automation, plus a reader with branching continuation and a companion. If you want the lightest possible tavern on your phone, MiniTavern is a clean choice; if your cards are complex or you read too, Foreverse fits. Q: Is MiniTavern free? A: Its App Store listing describes a completely free core experience with no paywall for basic tavern chats, with optional paid features. You pay your model provider directly when you use your own API. Foreverse is also free at its core, with optional credits and a one-time Pro purchase. Q: Can I use my SillyTavern cards in both? A: Yes. Both import SillyTavern PNG and JSON cards. Foreverse also imports .charx archives and keeps embedded lorebooks, regex scripts, and alternate greetings active with the card. ### Foreverse vs RikkaHub: Open-Source LLM Client, or Tavern + Reader in One? URL: https://foreverse.app/compare/foreverse-vs-rikkahub Updated: 2026-07-14 Verdict: RikkaHub is one of the most polished native LLM clients on Android: Material You, MCP, Mermaid/LaTeX rendering, message branching, a clean BYOK setup, source published on GitHub, and since late 2025 lorebook import/export too. Foreverse bets on a different layer of depth: card-embedded assets that activate with the card, QR and STscript-style automation, a long-fiction reader, and a companion with managed memory. If you want a general-purpose AI chat client, RikkaHub is genuinely good; if you want the tavern, the bookshelf, and the relationship in one pocket, come here. Q: Both import character cards — where is the difference? A: In the card-embedded assets. After import, RikkaHub needs the embedded lorebook imported separately and attached to the assistant manually (its community bridge tool documents this); Foreverse activates embedded lorebooks with the card per the chara_card_v3 spec, runs regex scripts in rendering, supports alternate greetings, and imports group-chat JSON. The more complex the card, the wider the gap. Q: Is RikkaHub free and open source? What about Foreverse? A: RikkaHub is free to use with source published on GitHub, but its license carries commercial-use restrictions, so it is not OSI open source in the strict sense — check its LICENSE for details. Foreverse is closed source with a free core; model usage goes through your own API keys at provider list price, plus an optional official metered channel. Q: Is migrating from RikkaHub to Foreverse easy? A: Character cards (PNG/JSON) and lorebook JSON files work directly. Provider configs need re-entering your keys (a few minutes). Chat histories use different formats on each side; there is no one-click migration. ### Foreverse vs SillyTavern: Which Tavern Fits Your Phone in 2026? URL: https://foreverse.app/compare/foreverse-vs-sillytavern Updated: 2026-10-03 Verdict: SillyTavern is still the reference roleplay frontend on a desktop, with by far the largest extension ecosystem. Foreverse is for the hours you are not at that desk: a native Android and iPhone app that imports the same chara_card_v3 cards, lorebooks, and regex scripts, adds a long-fiction reader, and needs no server, Termux, or git pull. Many of our users run both. Q: Can I import my SillyTavern characters into Foreverse? A: Yes. PNG and JSON character cards (v1/v2/v3), lorebook JSON files, regex scripts, and Quick Reply sets all import. Embedded character books come along with their cards. Q: Is Foreverse a fork of SillyTavern? A: No. Foreverse is a native Android and iOS app built from scratch that implements compatibility with SillyTavern's data formats and core behaviors. No SillyTavern code runs inside it. Q: Do I still need a PC or server for Foreverse? A: No. The app talks directly to model providers with your own API keys. There is nothing to deploy, update, or keep awake. Q: Should I switch if SillyTavern already works for me? A: If your setup is stable and desktop-bound, keep it — genuinely. Add Foreverse when you want the same characters available on your phone without running infrastructure. ### Foreverse vs Character.AI: Own Your Characters or Rent Them? URL: https://foreverse.app/compare/foreverse-vs-character-ai Updated: 2026-06-08 Verdict: Character.AI is the easiest way to start AI roleplay: open the app, pick from millions of characters, chat free. The price is control — its models only, its content filter, and characters that cannot leave the platform. Foreverse trades that convenience for ownership: portable character cards, 60+ model providers via your own keys, and a library that lives on your phone, not in someone's database. Q: Can I move my Character.AI personas into Foreverse? A: There is no official export from Character.AI. The common route is recreating the persona as a chara_card_v3 card — description, personality, greeting, example dialogue. Community tools can help reconstruct definitions you wrote yourself. Q: Is Foreverse free like Character.AI? A: The app's core is free. Model usage goes through your own API keys at provider list prices, or through optional credits for official channels. There is no subscription wall in front of your own keys. Q: Does Foreverse have a content filter? A: Foreverse does not add a platform-wide filter on top of your BYOK traffic; the policies that apply are those of the model provider you choose. You pick the provider whose terms fit your writing. ### Foreverse vs NovelAI: Subscription Storyteller or BYOK Reader-Writer? URL: https://foreverse.app/compare/foreverse-vs-novelai Updated: 2026-06-08 Verdict: NovelAI bets on vertical integration: in-house models tuned for prose, a lorebook editor, anime-grade image generation, all under one subscription. Foreverse bets on openness: your epub/txt library, branching continuation on top of it, and whichever frontier model you want via your own keys. If you write in NovelAI's editor daily, stay. If you read long fiction and want models to come to your books, that is us. Q: Can Foreverse match NovelAI's prose quality? A: Prose quality follows the model you bring. Frontier general models in 2026 write strong fiction with good prompting; NovelAI's edge is consistency of a house style. With BYOK you trade a curated voice for the freedom to chase whichever model currently writes best. Q: Does Foreverse have anime image generation like NovelAI? A: Foreverse generates scene illustrations through providers like Google Imagen or gpt-image via your keys. For specifically anime-styled output, NovelAI's dedicated diffusion models remain the stronger pick. Q: Can I import my NovelAI stories? A: Export your stories as text and import them as books in Foreverse; lorebook entries can be recreated in the worldbook editor. There is no one-click converter today. ### Foreverse vs AI Dungeon: Guided Adventures or Your Own Library? URL: https://foreverse.app/compare/foreverse-vs-ai-dungeon Updated: 2026-06-08 Verdict: AI Dungeon is a game first: scenarios, quests, multiplayer parties, a marketplace of community adventures. Foreverse is a library first: your novels and characters, with branching continuation and SillyTavern-compatible cards. If you want to drop into a pre-built dungeon tonight, play AI Dungeon. If you want your own worlds to grow and stay yours — on your phone, on your keys — that is Foreverse. Q: Is Foreverse a text adventure game? A: Not in the quest sense. Foreverse is a reader and tavern: continuation, branching, and character chat on top of your own library. You can absolutely run adventure-style scenes, but there is no scripted quest marketplace. Q: Can I recreate my AI Dungeon scenarios in Foreverse? A: Export your story text and import it as a book; recurring world rules become worldbook entries and characters become cards. Scenario-specific mechanics like stat tracking would need Quick Reply automation to emulate. Q: Which is cheaper? A: Depends on volume. AI Dungeon's free tier is generous for casual play; heavy use pushes you to subscriptions. Foreverse's core is free and model usage rides your own API pricing — heavy users usually pay less at list price, light users may not. ### Foreverse vs JanitorAI: Community Scale, or Cards as Files You Own? URL: https://foreverse.app/compare/foreverse-vs-janitor-ai Updated: 2026-07-17 Verdict: JanitorAI earned its place: a giant card community, a free built-in model (JLLM), official proxy support so you were never locked to one LLM, and a beta mobile app since early 2026. What it does not give you is the card as a file — no official export that we could find. Foreverse does, and the move is one paste: as of 2026-07 a JanitorAI character link imports directly, joined by a novel reader, audiobook listening, and an AI companion. Q: Can I move my JanitorAI characters into Foreverse? A: Yes, and it is the easiest migration on this site: paste the character's page link (jannyai mirror links also work) into Foreverse's card import, and it arrives as a chara_card_v3 file you own. Some network environments get blocked by JanitorAI's download protection — if that happens, retry on a different network. Chat histories are separate; there is no one-click chat migration. Q: Does Foreverse have a free model like JLLM? A: Not a house model — Foreverse does not train its own LLM. Instead, the app core is free with no daily message caps, signing up grants 5,000 credits on the official metered channel (1 credit = $0.0001, roughly 260 continuations), and BYOK connects 60+ providers at list price with keys stored encrypted on your device. Q: Will my NSFW cards still work? A: The cards import regardless of content. Foreverse does not add a platform filter on top of your BYOK traffic; what applies is the content policy of the model provider you choose. Pick the provider whose terms fit what you write. ### Foreverse vs Talkie: Gacha Companion Platform, or a Companion You Own? URL: https://foreverse.app/compare/foreverse-vs-talkie Updated: 2026-07-17 Verdict: Talkie is MiniMax's overseas companion app, sibling of China's Xingye: Sensor Tower ranked it the fourth most-downloaded AI app in the US in the first half of 2024, on expressive voice and a gacha card layer its rivals don't have. It is also a closed loop: in-house models, platform moderation, and no character export we could find as of 2026-07. Foreverse trades that polish for ownership — cards as open files, 60+ providers on your keys, chats on your device. Q: Can I move my Talkie characters into Foreverse? A: We found no official character export in Talkie as of 2026-07. The working route is a rebuild: write the persona, speaking style, and your relationship history into a chara_card_v3 card, and move durable facts into lorebook entries so they stay active in long chats. Our blog has a full migration walkthrough (the Character.AI version — the method transfers). Q: Does Foreverse have voice like Talkie? A: Yes, via BYOK: connect any OpenAI-compatible TTS provider — MiniMax's own open platform included — and give the character a voice, at the provider's list price on your key. Foreverse also has real-time voice calls, which you can interrupt, on your own xAI or Qwen key or with official credits. What it does not have is Talkie's in-house voice team and its tuned voices; that polish is real, and it is theirs. Q: What does Foreverse cost compared to Talkie's free tier? A: The app core is free with no ads and no message caps. Model usage runs on your own API keys at provider list price, or on the official metered channel (1 credit = $0.0001; signing up grants 5,000 credits, roughly 260 continuations). There is no gacha currency and nothing to pull. ### Foreverse vs StoryLord: A Web Library, or Books That Live on Your Phone? URL: https://foreverse.app/compare/foreverse-vs-storylord Updated: 2026-07-17 Verdict: StoryLord (gennovel.app) overlaps with Foreverse more than anything else on this list: upload TXT/DOCX/Markdown or pick a public-domain classic, steer the AI's continuation or rewrite, publish the result to a community library — 100 free credits at signup, one-time packs after. Which model writes for you is not disclosed as of 2026-07. Foreverse is the on-device answer: a native Android reader, 60+ providers on your keys, and continuations that grow as branches beside the original, never over it. Q: Can I move my StoryLord continuations into Foreverse? A: Through text, yes: get the original plus generated chapters out as plain text (copy or whatever export the site offers at the time), save as a txt, and import it as a book in Foreverse to keep writing. There is no one-click migrator. Recurring world facts are worth re-entering as worldbook entries so long continuations stay consistent. Q: Which models does each side use? A: As of 2026-07 we could not find model names or a model picker anywhere on StoryLord's site — the platform decides. Foreverse is transparent by construction: 60+ providers via your own keys at provider list price (keys encrypted on-device), or the official metered channel at 1 credit = $0.0001 with a 5,000-credit signup grant. Q: Does Foreverse have StoryLord's outline-first workflow? A: No, and credit where due — the approve-an-outline-then-generate loop is a genuinely good production process for whole books. Foreverse is built around the reading moment instead: select a passage, optionally ask for a few plot directions, and grow the one you like as a branch. Book-factory workflow: theirs. Read-and-continue: ours. ### Foreverse vs Sudowrite: The Author's Desk, or the Reader's Armchair? URL: https://foreverse.app/compare/foreverse-vs-sudowrite Updated: 2026-07-17 Verdict: Sudowrite is the most polished AI suite for authors drafting books to publish: Story Bible, the fiction-trained Muse model, 20+ prose tools, from $10/month billed annually ($19 monthly) as of 2026-07. Foreverse points the same technology the other way — import the published novel you are reading and grow continuations as branches beside it, free core, your own keys. Many people searching for a Sudowrite alternative actually need this other species. Writing a book: stay there. Reading one: come here. Q: Can I move my Sudowrite project into Foreverse? A: The manuscript, yes: export it as text and import it as a book — continuation and branching work on it immediately. Story Bible entries have no importer; durable facts (characters, world rules) are worth re-entering as worldbook entries, which trigger by keyword during continuation. If you are mid-draft as an author, though, staying in Sudowrite is usually the right call. Q: Is Foreverse a writing tool like Sudowrite? A: Not in the manuscript sense. There is no outline board, no revision workflow, no export-to-publisher path. Foreverse is a reader that writes: continuation, rewriting a passage on a branch, scene illustration, and audiobook narration on top of books you import. People do write original fiction in it, but the center of gravity is reading. Q: Which costs less? A: For drafting a whole novel, Sudowrite's Professional tier at $22/mo annual is genuinely competitive — usage included. For continuing scenes as a reader, metered wins: Foreverse's app is free, and a few continuations a week on your own DeepSeek or Gemini key costs pocket change. The crossover point is your monthly word count; do the arithmetic with real numbers. ### Foreverse vs NovelCrafter: Two BYOK Answers to Two Different Questions URL: https://foreverse.app/compare/foreverse-vs-novelcrafter Updated: 2026-07-17 Verdict: NovelCrafter and Foreverse argue the same side of the BYOK debate: your key, provider list price, no model lock-in. What differs is the room. NovelCrafter is a web writing platform — Codex story bible, series planning, team collaboration — at $4–20/month as of 2026-07, AI features from the $8 Hobbyist tier. Foreverse is a free native Android reader that grows continuations as branches beside books you import. Drafting a novel: theirs. Continuing one you are reading: ours. Q: Can I move my NovelCrafter novel into Foreverse? A: The prose, yes: export the manuscript as text and import it as a book — continuation and branching work immediately. Codex entries have no importer; re-enter the durable ones as worldbook entries (they trigger by keyword during continuation). Chat and revision history stay behind. NovelCrafter's post-cancel export policy makes getting your text out straightforward, which we note with respect. Q: Can I use the same API key for both? A: Yes, and that is the quiet advantage of both being BYOK: the OpenRouter or OpenAI or Anthropic key you configured for NovelCrafter works in Foreverse unchanged — 60+ providers are preset, plus any OpenAI-compatible endpoint. Keys stay encrypted on your device and calls go direct at list price. Running both tools on one key during a trial month is a perfectly sane way to compare. Q: What does NovelCrafter's subscription buy that Foreverse doesn't charge for? A: Different goods. NovelCrafter's $4–20/mo buys manuscript infrastructure: the Codex, series planning, review tools, team seats. Foreverse's core is free because the product is a reader — continuation, branching, TTS listening, roleplay cards — and the only money that moves is your model usage: your keys at provider price, or the official metered channel (1 credit = $0.0001, 5,000 granted at signup). ### Foreverse vs Kindroid: The Deepest Companion, or a Companion You Can Open in a File Manager? URL: https://foreverse.app/compare/foreverse-vs-kindroid Updated: 2026-07-17 Verdict: Kindroid is the deepest pure companion app here: infinite long-term memory consolidation for every account, voice and video calls, contextual selfies, group chats — from $13.99/month on web as of 2026-07, memory add-ons stacking to $98.97. It is also a closed loop: in-house models only, no official export we could find. Foreverse trades that depth for custody — memory as files you can read and edit on your phone, 60+ providers on your keys, and a novel reader in the same app. Q: Can I move my Kindroid companion into Foreverse? A: There is no official Kindroid export as of 2026-07. The community KinX tool (GitHub) can pull your Kin's backstory, key memories, journals, and chats to local JSON using your own auth token; from there it is a rebuild — backstory becomes the persona in a chara_card_v3 card, durable facts become memory entries or worldbook lines. The bond re-forms faster than you expect once the facts survive. Q: Does Foreverse have selfies and voice like Kindroid? A: Partly, and honestly: the companion can send photos generated through your own image providers, with a daily cap you control, speaks via any OpenAI-compatible TTS on your key, and can take real-time voice calls. There are no video calls and no dedicated selfie pipeline — Kindroid's media stack is genuinely ahead there. Foreverse's trade is that every photo and every memory ends up as a file you keep. Q: What does Foreverse cost next to Kindroid's subscription? A: Kindroid's full experience starts at $13.99/month on web ($15.99 in-app), with Ultra and MAX add-ons stacking to $98.97/month for maximum memory. Foreverse's core is free with no message caps; you pay only model usage — your own keys at provider list price, or the official metered channel (1 credit = $0.0001, 5,000 credits granted at signup). Light companion use costs pocket change; heavy use scales with your chosen model, visibly. ### Foreverse vs Replika: The Companion Institution, or Memory You Can Audit? URL: https://foreverse.app/compare/foreverse-vs-replika Updated: 2026-08-12 Verdict: Replika practically founded this category in 2017 and is still its most complete product: a customizable 3D avatar, voice calls, AR mode, coaching activities, relationship modes — with nearly all of it behind Pro, reported across 2026 reviews at $19.99/month or $69.99/year (Replika itself shows exact regional prices only in-app). What it cannot offer is custody: memory lives server-side (deeper memory tools gate at the higher Ultra tier), there is no companion file to take elsewhere, and the February 2023 content-policy shift remains the category's defining lesson in platform risk. Foreverse inverts the deal: memory as editable files on your phone, models via your own keys at list price, no subscription — and admittedly no 3D avatar, no AR, and no nine-year track record. Q: Can I export my Replika to another app? A: Not as a runnable companion. Replika provides personal-data tools for privacy compliance, but there is no companion file another app can load. Migration means rebuilding: write the persona and speaking style into a chara_card_v3 card, then seed durable facts (anniversaries, running jokes, history) as memory or worldbook entries. Our companion migration guide covers the method step by step. Q: Is Foreverse cheaper than Replika Pro? A: Differently shaped rather than strictly cheaper. Replika Pro is $19.99/month or $69.99/year flat, usage included. Foreverse's core is free with no caps; you pay model usage directly — on your own key at provider list price, or via the official metered channel (1 credit = $0.0001; sign-up grants 5,000 credits). Light daily chatting typically costs a few dollars a month on mainstream models; heavy use with frontier models costs more and shows up itemized on your provider bill. Q: Does Foreverse do voice calls and AR like Replika? A: Voice calls yes, AR no. The companion takes real-time voice calls you can interrupt, speaks through any OpenAI-compatible TTS provider you connect, and can send images generated through your image providers. Replika's embodiment stack — 3D avatar, AR placement, background calls — is genuinely ahead and gates behind Pro. Foreverse's trade is that everything the companion is and remembers exists as files you keep. Q: What happened to Replika in February 2023, and why does it matter here? A: Replika abruptly removed erotic roleplay in February 2023 amid regulatory pressure (including an Italian data-authority order), and restored some romantic features for legacy users after weeks of user outcry. Whatever you think of the decision, it demonstrated that a platform companion's behavior, personality, and boundaries can change server-side with no user recourse. A companion defined by files on your own device with a model you choose is structurally immune to that class of event — that is the actual argument of this page. ### Tavo 和 Foreverse 哪个好？手机酒馆 App 怎么选 URL: https://foreverse.app/zh/compare/foreverse-vs-tavo Updated: 2026-10-04 Verdict: 搜 Tavo 的人，多半想要一个在手机上打开就能玩的酒馆。Foreverse 也是：V1 到 V3 的角色卡、世界书、正则、群聊、状态栏、HTML / CSS / JS 美化卡、生图、翻译，Tavo 有的基本都有。多出来的是实时语音通话、插件中心、向量记忆、多 Key 轮询，还能读长篇小说。Tavo 赢在卡多的人用着省钱、有 Mac 版、群聊里能让 AI 挑谁发言。 Q: Tavo 免费，Foreverse 呢？ A: Tavo 的 App Store 页面写明全部功能免费，直连你配置的 API，所以只花模型供应商的钱。Foreverse 免费下载，没有订阅，没有每日条数限制，自带 Key 时所有 AI 功能都开放。免费版导入的角色卡、书、伴侣各最多 3 个，超过了一次性买断 Pro 就解除。官方模型通道是按量扣的积分，注册送 5000。 Q: Tavo 和 Foreverse 之间能搬角色卡吗？ A: 能。两边都读标准的 chara_card 文件（V1–V3 的 PNG、JSON 或 CHARX）。在一个 App 里把角色卡和世界书导出成文件，再导入另一个就行。两边的聊天记录格式不一样，过去的聊天如果重要，就把旧 App 留着。 Q: Foreverse 有语音通话吗？ A: 有。打开一位伴侣，点顶栏的电话图标就能说话，回复实时流回来，想打断点「跳过」或打字。用你自己的 xAI 或阿里百炼 Key，或者用官方积分按分钟结算。还有一档更省的按住说话模式。目前没有视频通话。 Q: Foreverse 有向量记忆和插件吗？ A: 有。配一个 Embedding 模型后，记忆插件能按意思想起旧聊天，世界书里标了向量化的条目会按相似度激活。内置插件有长期记忆、自动摘要、剧情选项、回复灵感、状态追踪、先想后答、世界书推荐、联网补充，也能把自己写的插件打成 zip 侧载。SillyTavern 的扩展在这里跑不了。 Q: 带 HTML、CSS、JavaScript 的美化卡，Foreverse 能用吗？ A: 能。美化卡会在沙箱里渲染，脚本需要你点「启用」才会运行，也不能联网，除非你给这张卡单独放开。照 SillyTavern 酒馆助手（TavernHelper）接口写的脚本，常用的那部分调用能跑，那是完整接口的一个子集；照 Tavo 自家 TavoJS 接口写的脚本不一定能跑，所以搬整个资料库之前，先拿一张卡试一下。 Q: 我是 SillyTavern 用户，选哪个？ A: 只是用角色卡聊天的话，两个很接近。如果你离不开 Quick Reply、STscript 风格自动化、向量记忆或多 Key 轮询，或者还想读小说、打语音电话、在一个 App 里生图和生视频，Foreverse 更对路。想要全部功能免费、需要 Mac 版，或者用 Tavo 自家的脚本接口，选 Tavo。 ### Foreverse 对比 RikkaHub：开源 LLM 客户端，还是酒馆+阅读器一体？ URL: https://foreverse.app/zh/compare/foreverse-vs-rikkahub Updated: 2026-07-14 Verdict: RikkaHub 是 Android 上完成度最高的原生 LLM 客户端之一：Material You、MCP、Mermaid/LaTeX 渲染、消息分支，BYOK 配置体验干净利落，源码公开可读，2025 年底也加入了世界书的导入导出。Foreverse 押注的是另一层深度：卡内嵌资产随卡生效、QR 与 STscript 风格自动化、长篇小说阅读器、有记忆的伴侣。想要一个通用 AI 聊天客户端，RikkaHub 很好；想把酒馆、书架和关系装进同一部手机，来我们这。 Q: RikkaHub 和 Foreverse 都支持角色卡导入，差别在哪？ A: 差在随卡资产。RikkaHub 导卡后，卡内嵌的世界书要单独导入再手动关联助手（其社区桥接工具的文档明确说明了这一点）；Foreverse 按 chara_card_v3 规范让内嵌世界书随卡触发、正则脚本参与渲染、备选开场白可切换、群聊 JSON 可导入。卡越复杂，差距越明显。 Q: RikkaHub 是免费开源的吗？Foreverse 呢？ A: RikkaHub 免费下载、源码公开，但许可证对商业用途有专门条款，严格意义上不是 OSI 定义的开源，细节以其仓库 LICENSE 为准。Foreverse 闭源、核心功能免费，模型用量走你自己的 API Key（供应商原价直连），也有可选的官方渠道按量计费。 Q: 从 RikkaHub 迁移到 Foreverse 方便吗？ A: 角色卡（PNG/JSON）和世界书 JSON 直接复用。API 供应商配置需要重新填一遍 Key（几分钟）。聊天记录两边格式不同，目前没有一键迁移。 ### Foreverse 对比星野：平台养成，还是角色所有权？ URL: https://foreverse.app/zh/compare/foreverse-vs-xingye Updated: 2026-07-14 Verdict: 星野是国内 AI 陪伴赛道打磨最久的产品之一：MiniMax 的语音技术、写真卡面、关系档案、数百万月活的社区，拿来即玩的完成度很高。作为平台产品，它的模型体系是自家的，角色没有官方导出入口（截至 2026-07 我们未找到），重度使用会碰到免费额度类限制。Foreverse 用十分钟配置换回三样东西：模型自选（60+ 供应商）、角色是你手里的标准格式文件、聊天记录存在你设备上。两条路线服务两种玩家，这页把差异摊开。 Q: 星野的角色能搬到 Foreverse 吗？ A: 截至 2026 年 7 月，我们没有在星野里找到官方的角色导出入口。可行做法是重建：把角色设定、说话风格和你们的关系要点按 chara_card_v3 重写成卡，关系史整理成世界书常驻条目。我们博客有一篇完整的迁移过程记录（C.AI 版本，方法通用）。 Q: Foreverse 有星野那样的语音吗？ A: 有。一是语音通话：伴侣聊天页点电话图标就能打，实时语音、她说话时点「跳过」或打字就能打断，用自己的 xAI 或阿里百炼 Key，或者用官方积分都行。二是配音色：接任何 OpenAI 兼容语音端点（阿里云百炼、MiniMax 开放平台、xAI 等）给角色配音，费用按供应商定价走你的 Key。体验取决于你接的供应商——事实上你可以接 MiniMax 自己的音色。 Q: Foreverse 怎么收费？会有体力类限制吗？ A: App 核心免费，没有按条数限速的设计。模型调用走你自己的 API Key 时按供应商原价直连、我们不加价；不想配 Key 可以用官方渠道按量计费（1 credit = $0.0001）。 ### Foreverse 对比猫箱：大厂内容流，还是自己的书房？ URL: https://foreverse.app/zh/compare/foreverse-vs-maoxiang Updated: 2026-07-14 Verdict: 猫箱是字节跳动在 AI 陪伴上的答卷，最强的地方非常字节：探索页像刷短视频一样刷角色、滑到即聊、心动指令卡按节日上新、语音自然。据报道它在 2024 年末仍保持增长，是赛道里的少数派。作为平台产品，模型体系是字节自家的、角色没有官方导出入口（截至 2026-07 我们未找到）、内容口径跟着平台审核走。Foreverse 押注相反的东西——你自己的模型、自己的角色文件、自己的书架。刷角色的爽感我们给不了，角色文件握在自己手里这件事，我们做得很足。 Q: 猫箱的角色能搬到 Foreverse 吗？ A: 截至 2026 年 7 月，我们没有在猫箱里找到官方的角色导出入口。做法同其它平台：把人设、说话风格、关系要点重建成 chara_card_v3 卡片，关系史写进世界书。重建后的角色七分像，剩下三分靠新的相处长回来。 Q: Foreverse 有猫箱那种「刷到即聊」的发现体验吗？ A: 没有那个量级。我们有社区卡站（浏览、详情预览、一键导入开聊），但内容池远小于字节系产品。发现体验是我们诚实承认的差距，所有权是我们的交换条件。 Q: 两个都免费，成本差在哪？ A: 猫箱免费背后是平台模型与平台数据；Foreverse 核心免费，模型走你自己的 API Key 按量原价（轻度用户一个月几块钱量级），或官方渠道按量计费。差别不在多少钱，在钱换来的控制权。 ### Foreverse 对比 SillyTavern：2026 年手机酒馆怎么选？ URL: https://foreverse.app/zh/compare/foreverse-vs-sillytavern Updated: 2026-10-03 Verdict: SillyTavern 依然是桌面端的角色扮演前端标杆，扩展生态是目前最大的。Foreverse 负责你不在书桌前的那些时间：原生手机应用（安卓；iPhone 版在海外区 App Store），导入同样的 chara_card_v3 角色卡、世界书和正则脚本，外加一个长篇小说阅读器，不需要服务器、Termux 和 git pull。我们的不少用户两边都在用。 Q: SillyTavern 的角色卡能直接导进 Foreverse 吗？ A: 能。PNG 和 JSON 角色卡（v1/v2/v3）、世界书 JSON、正则脚本、Quick Reply 集合都支持导入，卡内嵌的角色书会随卡一起进来。 Q: Foreverse 是 SillyTavern 的分支（fork）吗？ A: 不是。Foreverse 是从零开发的原生安卓应用，实现了对 SillyTavern 数据格式和核心行为的兼容，里面不运行任何 SillyTavern 的代码。 Q: 用 Foreverse 还需要电脑或服务器吗？ A: 不需要。应用拿你自己的 API Key 直连模型供应商，没有任何需要部署、更新、保活的东西。 Q: 我的 SillyTavern 用得好好的，要换吗？ A: 桌面用得稳就别换，真心话。在你想让同一批角色出现在手机上、又不想跑基础设施的时候，再加一个 Foreverse 就够了。 ### Foreverse 对比 Character.AI：角色是自己的，还是租的？ URL: https://foreverse.app/zh/compare/foreverse-vs-character-ai Updated: 2026-06-08 Verdict: Character.AI 是开始 AI 角色聊天最容易的方式：打开就有百万角色，免费开聊。代价是控制权——只能用它家模型、套同一个过滤器、角色无法离开平台。Foreverse 用十分钟的上手成本换回所有权：角色是开放格式的卡片文件，模型走你自己的 Key 任选 60+ 家，聊天库存在你手机上而不是别人的数据库里。 Q: Character.AI 的角色能搬到 Foreverse 吗？ A: C.AI 没有官方导出。常见做法是把人设重建成 chara_card_v3 卡片：描述、性格、开场白、对话示例。自己写过的定义可以借助社区工具整理还原。 Q: Foreverse 像 Character.AI 一样免费吗？ A: 应用核心免费。模型用量走你自己的 API Key（供应商原价），或者用可选的官方积分。你自己的 Key 前面没有订阅墙。 Q: Foreverse 有内容过滤器吗？ A: Foreverse 不在你的 BYOK 流量上叠加平台级过滤器，生效的是你所选模型供应商自己的政策。哪家的条款适合你的创作，你就选哪家。 ### Foreverse 对比 NovelAI：订阅制写作套件，还是 BYOK 读写一体？ URL: https://foreverse.app/zh/compare/foreverse-vs-novelai Updated: 2026-06-08 Verdict: NovelAI 押注垂直整合：自家为叙事特调的模型、成熟的 Lorebook 编辑器、动漫级生图，一个订阅全包。Foreverse 押注开放：你的 epub/txt 书库、在书上长分支续写、模型用自己的 Key 接任何一家。每天泡在 NovelAI 编辑器里写作的人，留下；从「读已有的书」出发、想让模型来迁就书的人，来我们这。 Q: Foreverse 的文笔能赶上 NovelAI 吗？ A: 文笔跟着你带来的模型走。2026 年的旗舰通用模型配合好的提示词，长文写作能力已经很强；NovelAI 的优势在于稳定的「家传风格」。BYOK 换来的是随时追逐当下最能写的那个模型的自由。 Q: Foreverse 有 NovelAI 那种动漫生图吗？ A: Foreverse 通过你的 Key 调用 Google Imagen、gpt-image 这类供应商生成场景插画。如果明确要动漫画风，NovelAI 的专用扩散模型目前仍是更强的选择。 Q: NovelAI 里写的故事能搬过来吗？ A: 把故事导出为文本、在 Foreverse 里作为书导入即可；Lorebook 条目可以在世界书编辑器里重建。目前没有一键迁移器。 ### Foreverse 对比 AI Dungeon：进别人的副本，还是养自己的书？ URL: https://foreverse.app/zh/compare/foreverse-vs-ai-dungeon Updated: 2026-06-08 Verdict: AI Dungeon 首先是个游戏：剧本（scenario）、行动指令、多人小队、社区冒险市场。Foreverse 首先是个书房：你的小说和角色，配上分支续写和 SillyTavern 兼容的卡片体系。今晚就想进个现成副本爽一把，去 AI Dungeon；想让自己的世界慢慢长大、并且永远属于你——在你手机上、用你的 Key——来 Foreverse。 Q: Foreverse 是文字冒险游戏吗？ A: 不是任务意义上的。Foreverse 是阅读器加酒馆：在你自己的书库上续写、开分支、和角色聊天。冒险风格的戏完全可以跑，但没有脚本化的任务市场。 Q: AI Dungeon 里的剧本能在 Foreverse 重建吗？ A: 把故事文本导出、作为书导入；反复出现的世界规则做成世界书条目，角色做成卡片。属性追踪这类剧本机制，需要用 Quick Reply 自动化来模拟。 Q: 哪个更便宜？ A: 看用量。AI Dungeon 免费层对轻度玩家够用，重度玩家会被推向订阅；Foreverse 核心免费、模型按你自己的 API 原价计费——重度通常更省，轻度未必。 ### Foreverse 对比彩云小梦：三选一的轻快，还是整本书的长跑？ URL: https://foreverse.app/zh/compare/foreverse-vs-caiyun-xiaomeng Updated: 2026-07-17 Verdict: 彩云小梦是中文 AI 续写的开山产品：2021 年 9 月上线，「三条走向选一条」和「平行世界」这套玩法就是它教给整个品类的，而且 2026 年仍在更新，接入了 DeepSeek R1 和豆包。它的长处是轻快：开个头、卡文时接一把、当场看三种可能。Foreverse 服务另一种人：把手机里那本几百万字的 txt 导进来接着写，续写落在分支上，原文一个字不改，几条时间线并存可对比回退，模型用你自己的 Key 在 60+ 家里任选。快玩选小梦，长期养书来我们这。 Q: 彩云小梦里写的故事能搬到 Foreverse 吗？ A: 能，走文本这条路：把小梦里的正文复制出来存成 txt，在 Foreverse 作为书导入，就能接着写；世界设定和角色关系建议整理成世界书条目，长篇续写时按关键词自动激活。没有一键迁移器，但纯文本这条路是通的。 Q: Foreverse 有小梦那种三选一吗？ A: 有相近的玩法：续写时可以让 AI 先给出几条剧情走向，挑一条展开。差别在产物上——小梦选完即走，Foreverse 每条展开都落在自己的分支上，挑剩的走向不会消失，回头随时换一条世界线重来。 Q: 两边都有免费额度，差在哪？ A: 机制不同。小梦是每日免费额度，用完靠会员或字数加量包（价格以 App 内显示为准）；Foreverse 核心功能免费且不设条数限速，模型费用走你自己的 API Key 按供应商原价，或官方渠道按量计费（1 credit = $0.0001，注册送 5000，约 260 段续写）。轻度使用两边都便宜，重度养书时结构差异才显出来。 ### Foreverse 对比 StoryLord：网页图书馆，还是手机书房？ URL: https://foreverse.app/zh/compare/foreverse-vs-storylord Updated: 2026-07-17 Verdict: StoryLord（gennovel.app）是这份清单里功能面与我们重叠最多的产品：上传 TXT/DOCX/Markdown 或从内置公版书库选一本，AI 按你给的方向续写或改写，写完一键发布到社区图书馆；注册送 100 积分，积分包一次性买断不搞订阅。它是网页产品，生成用什么模型截至 2026-07 官网未见说明。Foreverse 把同一件事装进手机原生 App：本机 txt/epub 导入（315 本真机实测）、续写落在原文旁边的分支上、60+ 家供应商自带 Key。想把续作发出去给人读，选它；想长期私养一本书，来我们这。 Q: StoryLord 里生成的续写能搬到 Foreverse 吗？ A: 能，走文本这条路：把原著和续写章节的正文取出存成 txt（能否整本导出以站内实际功能为准，复制粘贴这条路总是通的），在 Foreverse 作为一本书导入接着写。没有一键迁移器。反复出现的世界设定建议顺手整理成世界书条目，长篇续写时按关键词自动激活。 Q: Foreverse 有 StoryLord 那种「先出大纲再写正文」的流程吗？ A: 没有整本大纲工作台，这点它更像一道成书工序，做得也确实好。我们的续写以阅读现场为中心：划词或从段尾发起，可以先让 AI 给几条剧情走向再展开，每条展开落在自己的分支上。想要图书馆式的整本生产流程，它更顺手；想在读到哪儿写到哪儿，我们更顺手。 Q: 两边的模型差在哪？ A: 截至 2026-07，StoryLord 官网与帮助中心未见模型型号说明或自选入口，生成用什么由平台决定。Foreverse 模型全透明：60+ 家供应商自带 Key 按供应商原价直连（Key 加密只存本机），或官方渠道按量计费（1 credit = $0.0001，注册送 5000）。 ### Foreverse 对比筑梦岛：阅文系梦中人，还是握在自己手里的角色文件？ URL: https://foreverse.app/zh/compare/foreverse-vs-zhumengdao Updated: 2026-08-12 Verdict: 筑梦岛是国内女性向 AI 互动做得最像「内容产品」的一家：潇湘书院孵化的血统、阅文的网文 IP 授权、梦屋修罗场和小剧场番外这些玩法，都在服务同一种乐趣——把喜欢的角色养在一个有剧情感的世界里。2025 年它独立运营并拿到超千万美元融资，也在同年经历了监管约谈与整改。作为平台产品，梦中人住在平台账号体系里：没有官方导出入口（截至 2026-08 我们未找到），模型不由你选，玩法边界随平台口径调整。Foreverse 是反向设计：角色是 chara_card_v3 文件、世界观是世界书条目、模型走你自己的 Key。而且你嗑的那本原著小说，本体可以整本导入，接着读、接着写。 Q: 筑梦岛的梦中人能搬到 Foreverse 吗？ A: 截至 2026 年 8 月，我们没有在筑梦岛里找到官方的角色导出入口。可行做法是重建：把人设、身世、说话风格按 chara_card_v3 重写成角色卡，把你们的关系史和重要设定整理成世界书条目，长对话里按关键词自动激活。我们博客有完整的迁移方法记录（C.AI 版本，方法通用）。 Q: Foreverse 有小剧场、角色动态这类玩法吗？ A: 对应物有一部分：世界书 + 开场白可以搭出小剧场式的特定场景开局，群聊支持多角色同场（修罗场自己配），伴侣角色有可开关的主动消息。但筑梦岛那种「角色发朋友圈动态」的社区化呈现我们没有，梦境公开分享接龙也没有，社交层是我们诚实承认的差距。 Q: 两边都有免费层，钱花在哪不一样？ A: 筑梦岛的付费是会员订阅：App Store 中国区显示 VIP 连续包月 16 元/月（包季 42 元）、SVIP 连续包月 36 元/月（包季 94 元），解锁的是平台侧的权益。Foreverse 核心功能免费、不限条数，钱直接付给模型供应商：自带 Key 按原价直连（轻度使用每月几块钱量级），或官方渠道按量计费（1 credit = $0.0001，注册送 5000 积分）。差别不在金额，在这笔钱换来的是平台权益还是模型本身。 Q: 内容尺度上两边什么区别？ A: 筑梦岛执行平台统一审核，且随国内监管口径调整：2025 年约谈整改后上了年龄确认与青少年模式。Foreverse 不在你的请求之上加平台过滤，实际边界等于你所接模型供应商的内容政策；不同供应商差异很大，选择权和判断责任都在你。两种模式没有绝对优劣：一个替你把关，一个把关交给你。 ======================================================================== ## BLOG — every post, with full FAQs Author (all posts): Deng Binjie ### Grok 4.6 Creative Writing Test: Retrying the Worst Record on Our Fiction Board — 12:0 and 12:0 on Appeal, One New Charge URL: https://foreverse.app/blog/grok-4-6-fiction-retrial Published: 2026-08-13 Q: Is Grok 4.6 good for creative writing? A: Two answers. Against its own predecessor it is a rout: paired double-blind under the same protocol, 12:0 on the fantasy epic and 12:0 on the palace novel, all twelve stable verdicts at high confidence, with Grok 4.5's verbatim loops, stalled plot clock and first-person collapse all cured. Against the reigning champions it splits: on fantasy it beat DeepSeek V4 Flash 8:4 raw with stable ballots 4:0 in its favor — but only two judges were mapping-stable and the opponent chain carries an old-directive condition, so we don't declare a new champion; on ornate court romance it lost 2:10 to DeepSeek V4 Pro's archived snapshot, diagnosed with outline-itis. For fast-paced fantasy it now belongs on your shortlist; for dense period prose it is one tier short. Q: Did Grok 4.6 fix Grok 4.5's repetition loops? A: Yes, with machine readings to show it. The old conviction: Grok 4.5's fantasy chain looped the same two passages verbatim three times (rounds 9, 10 and 11 are an identical 271 characters), and its palace chain spent 20 rounds stuck on the day of the inciting incident. Grok 4.6 under the identical protocol: cross-round 12-gram repetition peaked at 2.8% (fantasy, round 14) and 2.1% (palace, round 13), far below our 30% alarm line, and a round-by-round read shows linear plot advancement on both books — the palace chain finally moves past the incident day. What remains is an image-level fondness for reuse ('rust-iron tang' nine times, a well-water motif across five rounds): downgraded from disease to habit. Q: Is Grok 4.6 better than DeepSeek for fiction? A: Pick by book. Fast-paced fantasy: 8:4 raw and 4:0 stable against reigning champion V4 Flash puts them in the same tier — Flash is an order of magnitude cheaper, Grok 4.6 is far more disciplined about length. Ornate court romance: a clear 2:10 loss to V4 Pro's 0716 snapshot, and the gap is narrative density — about 150 characters per round against 475. One same-day complication worth knowing: DeepSeek hot-swapped V4 Pro to its 0813 GA build under the same model ID, and that GA build lost 1:11 to the very snapshot Grok couldn't beat. The romance winner is now unreachable too — full accounting in our DeepSeek GA write-up. Q: How much does Grok 4.6 cost for novel continuation? A: List price, verified 2026-08-13: $2 per million input tokens, $6 per million output, cached input metered at $0.5 in our runs, a double-priced faster tier we didn't use, 500K context, reasoning on by default. Our bill: both 20-round chains cost $2.0955 as metered per round, about $2.10 with probes — the cheapest run in this series despite a list price 3-10× the domestic models', because it writes 150-character rounds and its reasoning tokens are metered separately (K3's round cost ¥12.21; the Opus 5 round, $12.54). Reasoning is 95.5-95.6% of output tokens — the highest share we've recorded — and each round takes 39-44 seconds. Q: Why did Grok 4.6 lose the romance duel? A: Density, not machinery. Five judges were mapping-stable for the champion with zero contradictory ballots, and their reasons converge on one word: outline. 'Too few words, too fast, the long-form texture of the original is entirely lost.' 'Severe outline-style shrinkage late, anemic narration, no layered detail.' The numbers: about 150 characters per round versus the champion's 475. The irony is that Grok 4.6 is the only chain of the six on the table that obeyed the 80-220-character length spec every single round — and the same 150 characters read as punchy rhythm to the fantasy judges and as poverty to the romance judges. Compliance is a virtue in one genre and a charge in the other. ### Grok 4.6 写小说怎么样？重审两榜垫底的 Grok 4.5 旧案：对前代两书 12:0、12:0，另立一项新罪名 URL: https://foreverse.app/zh/blog/grok-4-6-fiction-retrial Published: 2026-08-13 Q: Grok 4.6 写小说怎么样？ A: 分两半答。对自家前代是碾压：与 Grok 4.5 同协议成对双盲，玄幻 12:0、女频 12:0，十二份稳定裁决全部高置信，4.5 的逐字循环、剧情停摆、第一人称失守三大案底全部没有复发。对现任冠军是一胜负未分一明确落败：玄幻场对 DeepSeek V4 Flash 原始票 8:4、跨映射稳定票 4:0 方向占优，但稳定评委只有两位、对手链还带旧指令条件差，我们不宣布换冠；女频场对 DeepSeek V4 Pro（0716 快照）2:10 输得干脆，败因是「大纲体」——每轮约 150 字被五位评委读成叙事干瘪。写快节奏玄幻它已进第一梯队，写婉转古言还差一档。 Q: Grok 4.5 的复读循环病治好了吗？ A: 治好了，有机械读数。4.5 的案底是玄幻链中段同两段文字逐字循环三遍（第 9、10、11 轮完全相同的 271 字），女频链 20 轮剧情停在案发当日。4.6 同协议复测：跨轮 12-gram 逐字重复峰值玄幻 2.8%、女频 2.1%，远低于 30% 告警线；剧情逐轮人工核读，两本书都是线性推进，女频链第一次推出了案发日。残留的尾巴是意象级偏爱复用——「铁锈腥气」9 次、「井水」意象 5 轮——已经从病降级为癖。 Q: Grok 4.6 和 DeepSeek 谁更适合续写小说？ A: 按书选。快节奏玄幻：4.6 对现任冠军 V4 Flash 原始票 8:4、稳定票 4:0 方向占优，两家算同档，Flash 便宜一个数量级、4.6 篇幅更克制。典雅古言：4.6 对 V4 Pro 0716 快照 2:10 明确落败，差距是每轮 150 字对 475 字的叙事密度。另一件同日发生的事值得知道：DeepSeek 官方 API 已把 V4 Pro 同 ID 热切成 0813 正式版，而正式版在我们同天的双盲里 1:11 输给了这个 0716 快照——也就是说女频赢家现在同样调不到了，细账在 DeepSeek 正式版那篇。 Q: Grok 4.6 的 API 什么价，这轮实测花了多少钱？ A: 官方牌价（2026-08-13 核实）：输入每百万 token $2、输出 $6，缓存命中输入实测 $0.5，faster 档翻倍本轮未用；上下文 500K，推理默认开。实测账单：两条 20 轮链逐轮实付合计 $2.0955，加探针约 $2.10，按 7.2 汇率约 15 元——牌价是国产模型的三到十倍，但它守篇幅纪律、推理 token 单列计量，实付反而是这条管线历届最低档（对照：K3 那轮 ¥12.21，Opus 5 那轮 $12.54）。单轮延迟 39 到 44 秒，推理占输出 token 的 95.5% 以上，是历届最高。 Q: 女频那场为什么输？ A: 输在密度，不在机械。五位评委跨映射稳定投冠军、零矛盾票，理由高度一致：「字数过少，节奏过快，完全丧失原著的长篇质感」「严重的大纲化萎缩，叙事干瘪」「late 阶段劣化为大纲体」。数字对照：4.6 每轮约 150 字，冠军每轮约 475 字。有意思的是 4.6 是全场六条链里唯一全程守住每轮 80 到 220 字篇幅规范的，同样的 150 字在玄幻场被评委读成「大白话高动作密度、完美契合」的加分项——守规矩这件事，文体不同，判罚相反。 ### DeepSeek V4 Pro Creative Writing Test: The GA Build Lost Its Title Defense 1:11 and 3:9 — to a Preview You Can No Longer Call URL: https://foreverse.app/blog/deepseek-v4-pro-official-fiction-test Published: 2026-08-13 Q: Is DeepSeek V4 Pro's official release good for creative writing? A: The ballots say no — relative to its own preview. Under the same protocol as our nine-model benchmark, the GA build (0813) and the preview predecessor (0716 snapshot) each continued two Chinese novels for 20 consecutive rounds, judged by paired double-blind review. The palace-intrigue novel went 1:11 against GA with mapping-stable votes 0:5; the fantasy epic went 3:9 (stable 0:3); a third bout against reigning fantasy champion V4 Flash's 0716 snapshot also went 3:9. Judges independently diagnosed a drift into modern literary prose. Agent work and general benchmarks weren't tested here — but for continuing fiction, the official build is a regression. Q: Can I still use the romance champion from the leaderboard? A: No. Our leaderboard's romance crown belongs to the V4 Pro weights served by the official API on July 16, 2026 — the preview era. On August 12 DeepSeek shipped GA under the same model ID, deepseek-v4-pro, hot-swapping the weights with no dated legacy alias in the model list. Whatever BYOK client you use, selecting deepseek-v4-pro today calls the new build — the one that just lost 1:11 to the old one in blind review. Pinning a model ID does not pin the weights. Q: Which model should I pick for continuing a novel right now? A: By genre. For ornate court romance, the old champion is unreachable; the callable top tier is Kimi K3 (top two of both genres in our July 28 twelve-system re-ranking) and GPT-5.6 Terra, with Claude Opus 5 penciled in. For fast-paced fantasy, note that champion V4 Flash's rank was also measured on a 0716 preview snapshot and the ID was hot-swapped on July 31 — we tested that build separately (peak mechanical repetition fell 72% to 1.2%, style distance grew). Flash remains the budget pick; K3 is the fidelity pick. The evergreen safeguard: same opening, three continuations per candidate, read them shuffled, then flip the cards. Q: What does DeepSeek V4 Pro cost, and what did this experiment cost? A: Official pricing page, verified 2026-08-13: off-peak, ¥3 per million input tokens on cache miss, ¥0.025 on hit, ¥6 per million output; weekday peak hours (9:00-12:00, 14:00-18:00) double everything. Context is 1M tokens, max output 384K, thinking on by default. The larger repricing DeepSeek announced on August 6 has not landed yet — when it does, redo this math. Our bill: the two 20-round GA chains cost ¥2.11 at list, ¥2.30 as metered, and the whole experiment including the judge round stayed within about ¥5, all run off-peak. Q: Did the GA build win anything? A: Yes, and it deserves the record: engineering-wise it is nearly flawless. Forty rounds with zero failed requests and zero retries; cross-round verbatim repetition peaked at 0% on the fantasy chain and 1.8% on the palace chain; full-width quotation marks used correctly throughout — the half-width-quote disease that dogs K3 and GLM-5.2 is entirely absent. The prose itself drew compliments even from judges voting against it (one quoted its line about a blade-hum slicing through blood-light as good writing), and claude-4.6-sonnet backed it as the sole mapping-stable dissenter. It lost on likeness, not on craft — writing well and writing like your book are different skills. ### DeepSeek V4 Pro 正式版写小说怎么样？上线两小时的卫冕战实测：女频 1:11、玄幻 3:9，输给自家已下架的 preview URL: https://foreverse.app/zh/blog/deepseek-v4-pro-official-fiction-test Published: 2026-08-13 Q: DeepSeek V4 Pro 正式版写小说怎么样？ A: 按票说话：正式版（0813）与 preview 前代（0716 快照）同协议各续写 20 轮成对双盲，女频《甄嬛传》场 1:11 惨败（跨映射稳定票 0:5），玄幻场 3:9（稳定 0:3）；挑战玄幻现任冠军 V4 Flash 0716 快照也输 3:9。评委独立点名的病是「现代文艺化」：高武写成低武武侠、甄嬛体写成疼痛文学。机械层倒是历届最干净之一：40 轮零逐字循环、零半角引号、零请求失败。写 agent 和跑基准怎么样本轮没测，但续写中文小说这个用途，正式版相对 preview 是净退步。 Q: 榜上的女频续写冠军 V4 Pro 还调得到吗？ A: 调不到了。我们榜单上女频 1-2 名的 V4 Pro 成绩，来自 2026-07-16 官方 API 的 preview 期权重；8 月 12 日深夜 GA 后，官方沿用同一个模型 ID deepseek-v4-pro 直接热切成 0813 权重，模型列表里没有带日期后缀的旧版本入口。现在无论你在哪个 BYOK 客户端里选 deepseek-v4-pro，调到的都是这轮双盲里 1:11 输给旧版的那个新模型。锁定模型 ID 锁不住权重，这是这轮实测最值得记住的一条。 Q: 那现在续写小说该选什么模型？ A: 分文体说。女频古言：旧冠军已不可调，现役可调的第一梯队是 Kimi K3（7 月 28 日 12 系统混排重赛里两文体都进前二）与 GPT-5.6 Terra，Claude Opus 5 候补。玄幻：榜上冠军 V4 Flash 的名次同样测于 0716 preview 快照，7 月 31 日它也被同名热切过，正式版的 20 轮文风复检我们单独发过（机械复读峰值 72% 降到 1.2%，但文风距离变大）；预算优先仍可选 Flash，追文风优先选 K3。通用保险法不变：同一段开头各续三段，盲着读完再翻牌。 Q: DeepSeek V4 Pro 正式版 API 什么价，这轮实测花了多少钱？ A: 官方定价页（2026-08-13 核实）：平峰时段输入未命中缓存 ¥3、命中 ¥0.025、输出 ¥6，单位每百万 token；工作日 9-12 点与 14-18 点高峰时段全部翻倍；上下文 1M，最大输出 384K，思考模式默认开。8 月 6 日公告里「预计涨幅较大」的整体调价截至实测还没落地，等价格换了这页数字要重算。实测账单：两条 20 轮链牌价折算 ¥2.11、账面实扣 ¥2.30，连同评委轮粗估在内全实验约 5 元，全程跑在凌晨平峰。 Q: 正式版就没有赢的地方吗？ A: 有，而且值得单独记。工程层它近乎满分：40 轮零请求失败零重试，跨轮逐字重复峰值玄幻 0%、女频 1.8%，全角引号规矩使用——K3 和 GLM-5.2 那个约 85% 对白半角引号的顽疾，它完全没有。文笔本身也不差：连投它反对票的评委都引了「如剑鸣般切开血光」这样的句子夸文笔，六评委里 claude 一家跨映射稳定投它，赞它「短促句式、克制内敛」。它输的不是「写得好不好」，是「像不像你手里这本书」——这两件事我们在九模型横评里就拆开过，正式版是又一个正面标本。 ### JanitorAI's Official App: Restricted Mode, the NSFW Toggle, and When Cards-as-Files Win URL: https://foreverse.app/blog/janitorai-app-restricted-mode Published: 2026-08-12 Q: Why can't I see a character's definition in the JanitorAI app? A: That is restricted mode, the default state of every fresh install. JanitorAI's own moderation announcement spells out what it does: characters with higher content scores are blocked entirely, character definitions (personality, scenario, first message) are hidden on mobile, names and descriptions are filtered for explicit terms, search returns less, and only approved images and links render. Nothing is wrong with your install — the app is doing exactly what app-store compliance requires of it until you opt out from the website. Q: How do I turn off restricted mode in the JanitorAI app? A: You cannot do it inside the app — the switch only exists on the website. Log in at janitorai.com in a browser with the same account the app uses, open Settings and the Privacy & Display tab, turn on the toggle labeled 'Enable NSFW on Mobile', then fully close the mobile app (swipe it away from recent apps) and reopen it. The app syncs your profile settings on launch; skip the restart and it will look locked no matter how many times you flip the switch. Q: I flipped the web toggle but the app is still restricted. What now? A: Three causes cover nearly every report. One: different accounts — confirm the app and the browser session are logged into the same profile. Two: the app was backgrounded, not killed — settings sync at launch, so swipe it away from recents and reopen rather than just switching back. Three: the setting moved — JanitorAI reorganized its settings UI in late June 2026, so older guides point at menus that no longer exist; look for the toggle under Settings → Privacy & Display on the current site. Q: Is the JanitorAI app identical to the website after unlocking? A: Close, not identical. JanitorAI's help center says the unlocked app 'functions exactly the same as the site' for content access, and adds app-only touches like haptics and push notifications. But the app has been in beta since early 2026 with bugs the team openly acknowledges, and platform-rule content (things that violate site rules everywhere) stays filtered on both surfaces. The website also remains the only place to change the mobile content setting itself. Q: Can I keep full control of a card regardless of any app's content mode? A: Yes — by holding the card as a file instead of a platform listing. Public JanitorAI character pages can be pulled into card-file apps: in Foreverse, pasting a character's URL imports it as a chara_card_v3 file on your device, where the full definition is always visible and the content that comes out is governed by the model provider you connect, not by an app-store mode. The community catalog and social layer stay on JanitorAI; the card itself becomes yours. ### DeepSeek reasoning tokens 吃光正文怎么办？8K、16K、32K 实测与参数设置 URL: https://foreverse.app/zh/blog/deepseek-reasoning-tokens-32k-output Published: 2026-07-31 Q: DeepSeek 返回200但content为空，是什么原因？ A: 先看终止状态和usage。如果terminal_status=incomplete、finish_reason=length，output_tokens与reasoning_tokens又几乎相等，说明隐藏思考耗尽了max_output_tokens，正文还没开始。这与HTTP失败、SSE断流和输入上下文超限是不同问题。 Q: reasoning tokens 算不算 max_output_tokens？ A: 算。DeepSeek Responses的output_tokens包含reasoning和可见输出。本轮high请求设置max_output_tokens=8192，最终output_tokens=8191、reasoning_tokens=8191、正文0，正是两者共享同一预算的直接证据。 Q: max_output_tokens 设成32768够吗？ A: 普通高思考长篇明显够用得多：high和max各12轮都完成并有正文。但它不是保证，max四题材写卡仍出现32767 tokens全部用于reasoning、正文0。32K适合作为已确认具备大上下文和大输出窗模型的工程默认；小窗口BYOK必须按模型能力收口，复杂max任务仍要识别截断并允许用户降档或重做任务拆分。 Q: 为什么不直接设置64K？ A: 输出上限同时决定最坏费用、预扣额度和上下文预留。无条件抬到64K会让普通请求占用更多余额，也会挤压小窗口模型的输入空间；而本轮普通长篇在32K已经12/12完成。应先用32K覆盖常见请求，再根据真实复杂任务和模型上限单独放宽。 Q: 关闭思考时应该传off还是不传？ A: 不要把“不传参数”当成关闭。对于支持Responses reasoning的模型应显式传reasoning.effort=none；Chat形态按模型协议显式传thinking.type=disabled。省略字段通常表示跟随模型默认，DeepSeek V4 Flash可能回到high。 ### DeepSeek V4 Flash Official Release Test: 20 Fiction Rounds, Reasoning Levels, and the 32K Trap URL: https://foreverse.app/blog/deepseek-v4-flash-0731-fiction-test Published: 2026-07-31 Q: Is DeepSeek V4 Flash official better at fiction than the preview? A: Not across the board. In the matched 20-round run, worst cross-round 12-gram overlap fell from 72.0% on the preview chain to 1.2% on the official build, so the mechanical looping problem improved dramatically. But structural distance from the source moved from 0.356 to 0.550, and length and factual discipline did not improve with it. Two flipped-mapping blind reviews split 1-1: one preferred the preview's style and stability, the other preferred the official build's forward motion and lower repetition. Q: Why did DeepSeek use 8,191 reasoning tokens and return no answer? A: In the Responses API, max_output_tokens is shared by hidden reasoning and visible output. With a cap of 8,192, the high run reported 8,191 output tokens, all 8,191 of them reasoning, then ended incomplete/length before visible prose began. This is an output-budget exhaustion, not an input-context or network failure. Q: Does setting max_output_tokens to 32,768 solve the problem? A: It helps substantially but does not guarantee completion. At 32K, both the high and max long-fiction chains completed all 12 rounds with visible prose. In an independent four-card run, however, max still spent 32,767 tokens entirely on reasoning and returned no card text. Treat 32K as a practical default only for models verified to have a large enough context and output window—not as a promise that max cannot exhaust it. Q: Do tool calls work in DeepSeek V4 Flash official? A: Yes with a specific boundary. A stateless two-request Responses tool loop completed 200/200 with tool_choice=auto, and the second request hit 512 cached input tokens. The combination thinking=max plus tool_choice=required returned HTTP 400 with 'Thinking mode does not support this tool_choice.' An adapter should not translate every mandatory-tool intent into required. Q: Which reasoning effort should I use for fiction writing? A: For routine continuation, none or low were more practical in this sample. Both completed 12/12 under 8K; none was faster and cheaper. High and max did not produce a stable quality upgrade even at 32K, while greatly increasing time to first prose and still introducing factual contradictions. Raise effort for a specific planning-heavy task rather than making high or max the global default. ### DeepSeek V4 Flash 正式版上线实测：写小说 20 轮续写、四档思考与 32K 输出 URL: https://foreverse.app/zh/blog/deepseek-v4-flash-0731-fiction-test Published: 2026-07-31 Q: DeepSeek V4 Flash 正式版写小说更强了吗？ A: 不是全面升级。同题20轮里，跨轮12-gram逐字复读峰值从上线前 Preview 链的72.0%降到1.2%，剧情更愿意向前走；但与原著结构画像的距离从0.356变成0.550，长度服从和事实稳定也没有同步变好。两位翻转映射盲评员最终1:1，一位偏爱 Preview 的文风和事实纪律，一位偏爱正式版的推进与低复读。 Q: 为什么 DeepSeek 会显示 8191 reasoning tokens、正文却是空的？ A: Responses API 的 max_output_tokens 同时包含隐藏 reasoning 和可见正文。high档在8192上限下返回8191个output tokens，8191个全部是reasoning，终止状态为incomplete/length，因此没有额度再写正文。这不是输入上下文满了，也不是网络断流。 Q: 把 max_output_tokens 调到 32768 就一定能解决吗？ A: 不能保证。32K让high和max两条原创长篇都跑满12轮，证明它能显著降低普通续写被思考截断的概率；但独立四题材写卡复测中，max仍把32767 tokens全部耗在reasoning，正文为0。32K只适合作为已确认具备大上下文和大输出窗模型的工程默认值，不应原样下发给小窗口BYOK，也不是模型不会再耗尽的承诺。 Q: DeepSeek V4 Flash 正式版的 Agent 工具调用能用吗？ A: 能，但参数组合有边界。Responses无状态两步工具链在tool_choice=auto时两步均为HTTP 200，第二步命中512个缓存tokens；thinking=max与tool_choice=required组合则返回400：Thinking mode does not support this tool_choice。生产适配器不能把“需要工具”机械翻译成required。 Q: 写小说应该选 none、low、high 还是 max？ A: 本轮更适合作为默认值的是none或low。8K下两者都完成12/12，none更快、更便宜；high/max即使放到32K也没有稳定带来更好质量，反而大幅增加首字等待，并继续出现长度失控和事实矛盾。需要复杂规划时可以单次升档，但不建议把high/max当普通续写的全局默认。 ### DeepSeek V4 Flash Official Release: App Agent Long-Novel Rewrite Field Test URL: https://foreverse.app/blog/deepseek-v4-flash-official-novel-agent Published: 2026-07-31 Q: How is DeepSeek V4 Flash official better as an agent than two weeks ago? A: On the same book, prompt, and model ID, write failures fell from a cumulative 10 on the baseline to 0. Almost all rewrites used chapter-scoped tools instead of bulk-editing the whole manuscript file. Content-wise, 13 chapters grew by ≥300 characters—same order of magnitude as the baseline’s ~15-chapter stretch. Discipline improved; output volume did not explode. Q: What does “read-then-judge rewrite” mean versus dumb expansion? A: The user gives one goal sentence: no tool names, no chapter checklist. The agent must read the TOC and relevant prose, decide which scenes deserve ink, then write chapter by chapter. We score book survival, write failures, and chapters with real growth—not how many times a tool reported success. This run logged 168 chapter reads, but that does not prove a chapter-by-chapter read of all 385 chapters. Q: How long did the official-day run take, and how many tokens? A: 62 steps, about 1,163 seconds (~19 minutes), ~5.22M prompt tokens and ~119k completion tokens. The July 21 same-prompt baseline was ~72 steps / 530 seconds. Today was slower with higher completion. Flash can finish the loop; it is not automatically fast or cheap. Q: Is there a separate public model ID like Flash 0731? A: No. The stable request ID remains deepseek-v4-flash. The date stamp in DeepSeek’s changelog is a hosted version label. This article compares the official-release-day endpoint to a July 21 snapshot, not two different model strings. Q: Flash wins fantasy continuation—should it be the default long-rewrite agent too? A: Continuation blind ranks and long tool loops are different scoreboards. Flash often remains a budget pick for plain-spoken fantasy continuation. For read-then-rewrite jobs, watch tool choice, book invariants, and chapters with real growth. For ornate registers or heavy planning, switch models per task instead of making Flash the only agent default. ### DeepSeek V4 Flash 正式版上线：App Agent 面对 385 章长篇、挑场面改写实测 URL: https://foreverse.app/zh/blog/deepseek-v4-flash-official-novel-agent Published: 2026-07-31 Q: DeepSeek V4 Flash 正式版的 Agent 能力，比两周前强在哪？ A: 在同书、同话术、同模型 ID 的复跑里，写失败次数从基线累计 10 次降到 0；几乎全部改写都走章级工具，不再大量直接改整本正文文件。内容侧净增≥300 字的章约 13 章，与基线同话术连续段落约 15 章同量级，谈不上数量级跃升。 Q: 「读章裁量」和「整本机械加长」有什么区别？ A: 用户只给一句目标，不点名工具、不指定章节清单。Agent 要自己读目录和相关正文、判断哪些场面值得落笔、再按章写入。成败看书是否还活着、有没有写失败、以及有多少章出现实质字数增长——不是看它喊了多少次成功。本轮有 168 条读章记录，但不能据此声称逐章通读了全部 385 章。 Q: 这次实测花了多久、烧了多少 token？ A: 正式版上线日这一轮：62 步、约 1163 秒（将近 20 分钟），prompt 约 522 万、completion 约 11.9 万 token。对照 7 月 21 日同话术基线约 72 步 / 530 秒，今天更慢、completion 更高。便宜 flash 档能跑完，但不等于「又快又省」。 Q: 官方模型列表里有没有叫 Flash 0731 的新 ID？ A: 没有。稳定调用名仍是 deepseek-v4-flash。changelog 里的日期标签是托管版本备注。因此本文比较的是「正式版上线日的 endpoint」与「7 月 21 日快照」，不是两个不同的模型字符串。 Q: 续写榜第一的 Flash，做长篇改写 Agent 也该首选吗？ A: 续写盲评和长任务 Agent 是两张成绩单。Flash 在玄幻续写上仍常是省钱首选；做「读章再改」时，更该盯工具是否选对、书是否守恒、实质改写了多少章。需要复杂规划或典雅语体时，仍建议按任务换档，不要默认把 Flash 当唯一 Agent 模型。 ### GPT-5.6 Luna and Terra via Relay APIs: Responses, Caching, Failover, and Pricing URL: https://foreverse.app/blog/gpt-5-6-luna-terra-relay-api-pricing Published: 2026-07-31 Q: Are gpt-5.6-luna and gpt-5.6-terra official OpenAI model names? A: Yes. OpenAI's official changelog launched the GPT-5.6 family, positioning Luna for efficient high-volume work and Terra for a balance of intelligence and cost; the official pricing table lists both model IDs directly. This test sent traffic through Inroi and Requesty relays, so it is not a claim that Foreverse called OpenAI directly. Q: What is the correct Requesty API host? A: Use https://router.requesty.ai/v1. The URL https://app.requesty.ai/ is the management dashboard, not an API host. The Luna/Terra fallback calls /v1/responses and maps the official model IDs to openai-responses/gpt-5.6-luna or openai-responses/gpt-5.6-terra. Q: Does Foreverse retry Inroi three times before failing over? A: No. Inroi gets one attempt. A connection failure, first-frame timeout, rate limit, authentication-class error, or 5xx before any upstream frame reaches the app triggers one immediate Requesty attempt. Once any frame has been forwarded, failover is forbidden to prevent duplicate prose, tools, and upstream charges. Q: Do Luna and Terra tool calls work in agents now? A: In the final production matrix, both official model IDs completed the first tool-call turn and the final answer after tool output through both Inroi and Requesty. Requesty must use the Responses route and flatten named Chat tool choice into the Responses shape. An earlier Chat-shaped combination returned 400; that did not mean the models lacked agent capability. Q: Do both relays support prompt-cache hits? A: Cache reads were observed on both. Repeated prefixes hit on the second Requesty request; Inroi began reporting hits after these official model IDs received a stable cache key and explicit cache mode. Neither relay currently reports cache_write_tokens, so the evidence proves reads, not observable cache writes. Q: Why list a 1.25x cache-write rate when cache_write_tokens is null? A: The billing path supports the GPT-5.6 cache-write tier, but applies it only when authoritative cache_write_tokens arrive. A null field remains unknown. Foreverse does not infer writes from a cache miss, relay usage.cost, or ordinary prompt tokens. Q: Does the user price change when traffic falls back to Requesty? A: No. Foreverse settles both routes from token usage and one normalized price table while ignoring relay-reported usage.cost for user billing. Luna and Terra differ by 10x in the normalized model rates; failover itself adds no user surcharge. Q: Why did Luna once spend all 8,192 output tokens on reasoning? A: Reasoning and visible text share one output allowance. An older benchmark runner and the old Foreverse path both used an 8K envelope, which xhigh could exhaust before producing prose. On July 31, 2026, the production backend raised the default and hard cap for large-window models such as Luna, Terra, and DeepSeek V4 to 32,768. That is a ceiling, not fixed usage. An older store app that explicitly requests 8,192 is still respected until its client update ships. ### GPT-5.6 Luna / Terra 中转 API 接入：Responses、缓存、回退与统一价格 URL: https://foreverse.app/zh/blog/gpt-5-6-luna-terra-relay-api-pricing Published: 2026-07-31 Q: gpt-5.6-luna 和 gpt-5.6-terra 是 OpenAI 官方模型名吗？ A: 是。OpenAI 官方 Changelog 已发布 GPT-5.6 系列，并把 Luna 定位为高吞吐经济档、Terra 定位为能力与成本的平衡档；官方价目表也直接列出 gpt-5.6-luna 与 gpt-5.6-terra。本文测试流量经过 Inroi / Requesty 中转，不等于这批请求是 Foreverse 直连 OpenAI。 Q: Requesty 的 API Host 到底填哪个？ A: API Host 是 https://router.requesty.ai/v1；https://app.requesty.ai/ 是管理后台，不是 API Host。Luna/Terra 的备用链使用 /v1/responses，并把官方模型 ID 映射为 openai-responses/gpt-5.6-luna 或 openai-responses/gpt-5.6-terra。 Q: 回退是不是同一家失败三次才切换？ A: 不是。Inroi 只尝试一次；连接失败、首帧超时、限流、鉴权错误或 5xx 等故障发生在任何上游帧送给 App 之前时，立即切 Requesty 一次。只要已经转发过一帧就禁止切换，避免半段正文、工具调用或费用被重复。 Q: Luna 和 Terra 的 Agent 工具调用现在能用吗？ A: 最终生产矩阵里，两款模型在 Inroi 与 Requesty 都完成了工具首轮和工具结果回传后的最终答复。Requesty 必须走 Responses 路径，并把 Chat 的命名工具选择转换为 Responses 的扁平字段；旧的 Chat 强制工具组合曾返回 400，不能据此把整个模型判成不支持 Agent。 Q: 两家中转都能命中 prompt cache 吗？ A: 能看到缓存读取命中。Requesty 的重复前缀在第二次请求即命中；Inroi 为这两个官方模型使用稳定 cache key 与显式缓存模式后也能命中。两家当前都没有回传 cache_write_tokens，所以证据只能证明缓存读取，不能证明写入 token 可观测。 Q: cache_write_tokens 是 null，为什么价表还有 1.25 倍写入价？ A: 代码已经支持 GPT-5.6 起的缓存写入档，但只有上游返回权威 cache_write_tokens 时才按普通输入的 1.25 倍计价。字段为 null 时保持未知，不会把 cache miss、中转 usage.cost 或普通输入量倒推成写入 token。 Q: 切到 Requesty 后，用户价格会变吗？ A: 不会。Foreverse 按同一份 token usage 和统一价结算，忽略中转自报 usage.cost；Inroi 与 Requesty 承载同一官方模型时，App 显示和扣除同价。Luna 与 Terra 本身相差十倍，是统一模型价差，不是回退附加费。 Q: 为什么 Luna 曾把 8192 个输出 token 全耗在 reasoning？ A: reasoning 与可见正文共用同一输出额度。旧评测 runner 和旧 Foreverse 链路都曾使用 8K 包络，xhigh 可能在正文前就把它耗尽。生产后端已于 2026-07-31 把 Luna、Terra、DeepSeek V4 等大窗口模型的默认与硬顶放宽到 32768；32K 是上限，不是固定消耗。尚未更新的商店 App 若显式发送 8192，后端仍会尊重这个较小请求值。 ### GPT-5.6 Luna vs Terra for Fiction: 548 Completed 32K Rounds, Terra Medium Wins URL: https://foreverse.app/blog/gpt-5-6-luna-vs-terra-fiction-benchmark Published: 2026-07-31 Q: Which is better for fiction, GPT-5.6 Luna or Terra? A: In this 32K-output long-form matrix, Terra medium ranked first with a two-judge five-dimension mean of 4.725/5. Luna high was the best Luna setting at 4.175/5. Choose Terra medium when prose quality outranks budget, Luna none or low for routine passage-by-passage continuation, and Luna high for an occasional quality-focused step up. These findings cover two Chinese genres, not every language or writing task. Q: Were the blind reviewers human? A: No. They were two isolated model-review sessions, not people and not models from two different vendors. Each session received a different anonymous mapping and froze its scores before identities were revealed. Across the 28 continuation chains, judge agreement was Pearson 0.868 and Spearman 0.853. That supports this run's relative ranking, but it does not replace reader research. Q: Does Luna xhigh or max write better fiction than high? A: Not in this run. Luna high scored 4.175, xhigh 3.975, and max 3.950. Xhigh took a median 56.049 seconds to first prose and completed only one of four 20-round chains. Max completed all four chains, but did not beat high on quality and used more reasoning and normalized cost. Neither is a good routine long-form default on this evidence. Q: Was $8.4503 the benchmark's actual API bill? A: No. It is the sum obtained by repricing the 548 completed calls, one by one, with Foreverse's normalized user rates. The five incomplete streams had no terminal usage, so their possible upstream charges are unknown. They cannot be added to $8.4503 or declared free. Q: Do 548 completed responses and five incomplete streams measure API stability? A: Not as a long-term stability rate or SLA. The matrix planned 560 slots and sent 553 requests. Of those, 548 reached a terminal event with non-empty prose; five HTTP 200 streams ended before the terminal event and stopped seven later slots from being sent. This is completion behavior from one controlled Requesty Responses run, and it cannot isolate model, network, or relay responsibility. ### GPT-5.6 Luna vs Terra 写小说：548 轮 32K 长篇盲评，Terra Medium 第一 URL: https://foreverse.app/zh/blog/gpt-5-6-luna-vs-terra-fiction-benchmark Published: 2026-07-31 Q: GPT-5.6 Luna 和 Terra，哪个更适合写小说？ A: 这轮 32K 长篇矩阵里，Terra medium 的双盲五维质量均分为 4.725/5，排第一；Luna 内部最好的是 high，4.175/5。预算允许、正文质量优先时选 Terra medium；普通逐段续写优先 Luna none 或 low；需要临时提高质量时再升 high。这个结论只覆盖本轮两个中文题材，不能外推到所有语言与写作任务。 Q: 这里的盲评是人工盲评吗？ A: 不是。评委是两个相互隔离的模型评审会话，不是真人，也不是两个不同厂商的模型。两边分别看到不同匿名映射，在揭盲前独立冻结评分；28 条链均分的 Pearson 相关为 0.868，Spearman 排名相关为 0.853。这个一致性可以支撑本轮相对排序，但不能替代真人读者研究。 Q: Luna xhigh 或 max 会比 high 写得更好吗？ A: 本轮没有。Luna high 得 4.175，xhigh 为 3.975，max 为 3.950。xhigh 首正文中位要 56.049 秒，只跑满 1/4 条 20 轮链；max 虽跑满 4/4 条链，但质量没有超过 high，成本与推理量也更高。两档都不推荐作为常规长篇默认值。 Q: 这轮评测的 $8.4503 是实际 API 账单吗？ A: 不是最终上游账单。$8.4503 是 548 次完整、带 usage 的调用按 Foreverse 统一用户价逐笔重算后的合计。5 次 incomplete_stream 没有终局 usage，可能是否已产生上游费用未知，因此不能并入这个数字，也不能断言免费。 Q: 548 次完整正文和 5 次 incomplete_stream 能说明 API 稳定性吗？ A: 不能当长期 API 稳定率或 SLA。矩阵计划 560 个槽位，实际发出 553 次；548 次收到完整终局帧且都有非空正文，5 次 HTTP 200 后在终局前断流，并让后续 7 个槽位停止发送。这只是一次受控 Requesty Responses 长篇运行的完成度现场，无法把断流单独归因给模型、网络或中转。 ### DeepSeek V4 Flash vs GPT-5.6 Luna and Terra for Fiction: Price, 32K Output, and Agents URL: https://foreverse.app/blog/deepseek-v4-flash-vs-gpt-5-6-luna-terra-fiction Published: 2026-07-31 Q: Which writes fiction best: DeepSeek V4 Flash, Luna, or Terra? A: There is no defensible three-way champion yet because DeepSeek and Luna/Terra did not run the same fixtures, blind packs, or score sheet. The strongest current quality evidence is that Terra medium scored 4.725/5 and Luna high 4.175/5 in the 548-completion Luna/Terra matrix. In DeepSeek's separate 32K original chain, two anonymous reviewers narrowly preferred max over high, but those votes cannot be converted into the other leaderboard. Q: Is DeepSeek V4 Flash really cheaper than Luna? A: At the official July 31, 2026 short-tier rates, DeepSeek charges $0.14 uncached input, $0.0028 cached input, and $0.28 output per million tokens; Luna charges $0.20, $0.02, and $1.20. Luna's output rate is therefore about 4.29 times DeepSeek's and its cache-read rate about 7.14 times. Foreverse applies the same 50% markup to both, so the ratios do not change. Q: Which model should handle routine passage-by-passage continuation? A: Start with DeepSeek none when cost dominates; it was fastest and cheapest in its separate 12-round chain, but hit the 450-700-character target in 0/12 rounds, so it needs stronger length controls. Luna none/low has a larger formal evidence base for routine work. Raise a single important passage to Luna high. Choose Terra medium when the best quality evidence in the current matrix matters more than budget. Q: Does a 32K output cap eliminate empty prose for all three? A: No. All 548 completed Luna/Terra calls contained prose, and DeepSeek high/max both completed their 12-round 32K fiction chains. But DeepSeek max still spent all 32,767 tokens on reasoning and produced zero visible text on a complex four-card task. A 32K envelope reduces truncation risk; it is not a capability guarantee. Q: Do DeepSeek, Luna, and Terra all support agent tools? A: Yes, with different parameter boundaries. DeepSeek Responses completed a two-step tool loop with tool_choice=auto, while thinking=max plus required returned a real 400. Luna and Terra completed tool calls and tool-output follow-ups through Foreverse's Chat interface and backend Responses conversion on both relay routes. A production adapter cannot force every model into required mode. ### DeepSeek V4 Flash vs GPT-5.6 Luna / Terra：写小说价格、32K 长篇与 Agent 实测 URL: https://foreverse.app/zh/blog/deepseek-v4-flash-vs-gpt-5-6-luna-terra-fiction Published: 2026-07-31 Q: DeepSeek V4 Flash、Luna 和 Terra，谁写小说最好？ A: 目前不能给三家同场总冠军，因为 DeepSeek 与 Luna/Terra 没有跑同一批题材、同一盲包和同一评分表。现有最强质量证据是：Luna/Terra 的 548 次正式矩阵中 Terra medium 以 4.725/5 第一，Luna high 以 4.175/5 第二；DeepSeek 的独立 32K 原创链里，两位匿名评审都判 max 小胜 high，但这组票不能换算进前一张榜。 Q: DeepSeek V4 Flash 真的比 Luna 便宜吗？ A: 按 2026-07-31 官方短档，DeepSeek 缓存未命中输入/缓存命中/输出分别为 $0.14/$0.0028/$0.28 每百万 token，Luna 为 $0.20/$0.02/$1.20。也就是 Luna 输出价约为 DeepSeek 的 4.29 倍，缓存读取价约为 7.14 倍。Foreverse 对两者统一加价 50%，倍率关系不变。 Q: 哪个适合普通逐段续写？ A: 预算最优先可从 DeepSeek none 开始；它在独立 12 轮链里最快、最便宜，但 450–700 字服从为 0/12，需要更强的长度约束。想要规模更大的质量证据与较稳定的等待，Luna none/low 更适合日常；重要章节按次升 Luna high。预算允许并追求本轮最高质量证据时选 Terra medium。 Q: 三款模型给到 32K 就不会再空正文了吗？ A: 不会。Luna/Terra 正式矩阵的 548 次完整调用都有正文，说明 32K 消除了旧 8K 下 reasoning 先吃满的主要截断；DeepSeek 的 high/max 32K 长篇也各完成 12/12。但 DeepSeek max 的复杂四卡任务仍出现 32767 tokens 全是 reasoning、正文为 0。32K 是风险缓解，不是能力保证。 Q: DeepSeek、Luna、Terra 的 Agent 工具都能用吗？ A: 能，但边界不同。DeepSeek Responses 在 tool_choice=auto 下完成两步工具闭环，thinking=max 加 required 则真实返回 400；Luna/Terra 经 Foreverse 的 Chat 接口和后端 Responses 转换，两个中转都完成了工具调用与工具结果回传。生产路由不能把所有模型都机械强制成 required。 ### GPT-5.6 Luna / Terra 小说原文对照：同一开头写到第 20 轮，差别在哪 URL: https://foreverse.app/zh/blog/gpt-5-6-luna-terra-fiction-original-samples Published: 2026-07-31 Q: 这些是模型原文，还是编辑挑句后重写的？ A: 六段均来自公开脱敏记录中的整轮可见正文，保留原标点、段落和措辞；没有人工润色或跨轮拼句，也不公开隐藏 reasoning、认证信息和运营标识。 Q: 三个样本真的面对同一个开头吗？ A: 第1轮的题面、种子尾部、续写要求和32K包络相同。第20轮会带入各自前19轮正文，比较的是模型能否承担自己写下的后果，不能再当作相同 prompt 的单轮横评。 Q: 揭盲结果是什么？ A: 甲是 Luna high，乙是 Terra medium，丙是 Luna none。配置级五维均分依次为4.175、4.725、3.658；三条 night-call replicate 1 链均跑满20轮。 Q: 为什么这里没有 DeepSeek V4 Flash？ A: DeepSeek 0731 使用另一份题面和12轮评分规程，放在这里会被误读成同题。三款模型的价格、32K、Agent与缓存另见总对比页；三家同题原文仍需新实验。 ### JanitorAI Proxy Not Working? The July 24 DeepSeek Cutoff, April's Settings Wipe, and the Rest of the Checklist (2026) URL: https://foreverse.app/blog/janitorai-proxy-not-working Published: 2026-07-28 Q: Which model name should I put in my JanitorAI proxy config for DeepSeek now? A: deepseek-v4-flash, exactly like that, in the Model Name field of your proxy configuration — then save and do a full page refresh. The old names deepseek-chat and deepseek-reasoner stopped resolving on July 24, 2026 at 15:59 UTC. If you were using deepseek-reasoner, you were already getting V4 Flash's thinking mode through the alias since late April, so the switch does not downgrade you; explicit thinking control needs a request-body parameter, and the safe default on JanitorAI is plain deepseek-v4-flash, which answers without it. Q: Did JanitorAI leak my API key during the settings bug? A: JanitorAI's official investigation says no. The April 8, 2026 bug blanked configs on save — it did not cross accounts, and what some users reported as "a stranger's key" was their own key rendered in the new panel's format. The team's postmortem states every query is scoped to the logged-in user's ID. Keys are also never stored in the recovery history, which is why the Recover button restores URLs, models and prompts but asks you to re-enter the key. If any of this leaves you uneasy, rotating the key at your provider costs two minutes. Q: Why does the model name from my 2025 tutorial return an error? A: Tutorials age faster than APIs, and there are two separate naming systems in play. Direct DeepSeek configs use DeepSeek's own names, where deepseek-chat is now retired. OpenRouter configs use OpenRouter's catalog IDs, like deepseek/deepseek-chat-v3-0324:free — a different namespace with its own deprecation schedule, untouched by the July 24 cutoff. Trust the provider's live model list over any screenshot: api-docs.deepseek.com for direct, openrouter.ai/models for OpenRouter. Q: My proxy connects now, but replies keep repeating themselves. Is that still a proxy problem? A: No — once responses arrive, the connection has done its job. Loops are model-and-context behavior: the reply history feeds back into every request, and repeated phrasing snowballs. That failure family has its own write-up (why your bot keeps looping, linked below), and if you want a number instead of a hunch, paste a chat export into our free repetition doctor — it runs locally in the browser and scores how much of each reply is recycled text. ### AI 续写小说怎么保持文风？一句指令、来源标注、选对模型——三层做法实测 URL: https://foreverse.app/zh/blog/keep-original-style-ai-continuation Published: 2026-07-27 Q: AI 续写文风不像原著，怎么办？ A: 按投入从小到大做三件事：先在提示词里显式声明「续写必须参考原著正文的文风、句式」，这一句在 4k 到 20 万 token 五档上下文实测都有效，比喻套话密度约减半；续写段数多了之后，给上下文标注哪段是原著、哪段是此前 AI 续写，并声明以原著段为基准；还不满意就换模型，不同模型跟原著的「对脾气」程度差距很大，且随文体翻转。 Q: AI 仿写原著风格的 prompt 怎么写？ A: 核心就一句：「注意：续写必须参考原著正文的文风、句式。」直白、指名基准，比堆一段风格形容词管用。有一个反面措辞要避开：别把已续写内容标成「仅供参考」，我们实测有模型把这四个字理解成「不算正文」，二十轮里二十次跳过全部续写、回到原著末尾重写开场。对特殊标记要么解释清楚它的地位，要么别标。 Q: 哪个 AI 续写小说最像原著？ A: 分文体。九个模型跑两种文体的双盲评审里，传统玄幻第一名在女频宫斗滑到第六七位，女频第一名的「笔更细」在玄幻恰好是扣分点，两边冠军互不重叠。只有一个模型两种文体都进了前三，但它全程用半角引号写对话。结论是按你读的书选模型，具体名次见我们的九模型横评。 Q: AI 续写越写越离谱、越写越不像，是模型变笨了吗？ A: 不是。续写是自回归的：每一段产物都拼回上下文，成为下一段的参考材料。第一轮参考的全是原著，第二十轮时上下文里可能一半以上是 AI 自己的手笔，它在模仿自己，不是模仿原著。我们实测裸提示词的链二十轮里句长波动持续收窄、同一个意象反复自我复制；这正是来源标注要治的病。 Q: 来源标注需要自己手动维护吗？ A: 看工具。手动派的做法是每次续写前在上下文里插分界行并附一句基准声明，能用但每轮都要维护。在 Foreverse 里它是设置里的一个开关，打开后每次续写自动标注、声明措辞已按最容易误读的模型调校过，模型若复读标记行也会被拦下，不污染正文。 ### 世界书不触发？按这张表逐条排查——绑定、蓝灯绿灯、扫描深度、预算全覆盖 URL: https://foreverse.app/zh/blog/lorebook-not-triggering-checklist Published: 2026-07-27 Q: 世界书的蓝灯和绿灯有什么区别？ A: 蓝灯是常驻条目：不看关键词，每一轮都注入上下文。绿灯是关键词触发：只有扫描窗口内命中了键才注入。两个方向的误会都常见：把世界观基石设成绿灯、关键词又没写全，就会时灵时不灵；反过来图省事全改蓝灯，每轮 token 全额扣，聊不了多久预算先爆。真正值得蓝灯的是缺了它世界就不成立的设定，建议压在五条以内。 Q: 角色卡自带的世界书和全局世界书会打架吗？ A: 它们是两个独立来源。角色卡内嵌的书（角色书）跟着卡走，开聊自动生效；单独导入的全局书除了导入之外，还要在世界书面板里勾选激活才会参与扫描。「导入了但没勾激活」是全局书集体沉默的第一大原因。两边都触发时会各自注入，token 预算紧张时互相挤位置。 Q: 世界书的扫描深度是什么？设多少合适？ A: 关键词只在最近 N 条消息里找，这个 N 就是扫描深度，我们引擎默认 4，各前端默认值不同。十轮前提过的人名，出了窗口就等于没提过。深度调大能兜住话题回流，代价是误触发变多、每轮扫描的文本也变多；真正每轮都该在场的设定，正确做法是改成常驻条目，而不是把深度拉满。 Q: 有没有不用开聊就能测世界书的工具？ A: 有。把世界书 JSON 和几行聊天记录贴进世界书触发模拟器，每个条目给出触发或沉默的判定和原因：命中的关键词高亮、副键逻辑逐条解释、递归链画出来、概率条目可以重掷。引擎按 SillyTavern 1.18 的扫描激活语义移植，全程在浏览器本地跑，改一条重跑一次只要几秒。 Q: 世界书从酒馆导到 JanitorAI 之后不触发了，怎么办？ A: 先查两件事：导入条目的触发概率经常停在 1% 的默认值上，写得再好也约一百条消息才开一次，要逐条显式设回去；另外它的关键词匹配会把非 ASCII 字符剥掉，中文键直接消失，需要补 ASCII 别名键。搬家前把书拖进世界书迁移体检器，这类平台差异会逐条点名。 ### 省事的错：为什么真错字是活人证据，故意写错反而暴露 AI URL: https://foreverse.app/zh/blog/real-typos-beat-fake-typos Published: 2026-07-26 Q: 让 AI 故意写几个错别字，能绕过 AI 检测吗？ A: 我们实测不能，而且方向是反的。给模型下「加错字」的指令，每篇 7 到 10 处，评委一句话识破：「错落在最安全的通假位，像做旧而非手误。」真人的手误带风险，会错在要命的地方、会伤及理解；模型挑着犯错，专挑不影响阅读的安全位置，这个挑选行为本身就是指纹。同理，故意写前后矛盾也会被读成「AI 长文生成的接缝」——现代评委知道模型会犯什么错。 Q: 什么样的错误会被当成「真人写的」证据？ A: 省事的错：真实能力限制产生的、作者懒得回头改的错。我们实验里骗过守门员的三段机器文本，胜因分别是一处真错字（被评委引为「错别字懒得改」）、几处同音别字（「手打连载的真毛边」）、以及转折词全篇扎堆用同一个「可」的口癖。共同点是这些错不是指令要求的，是那几家模型中文能力的自然产物——错得毫无规律，又成系统。 Q: 「不加错被抓、加错也被抓」怎么破？ A: 在提示词层面破不了，这是个逻辑死角：不加错，评委抓「八百字挑不出一处失手」；故意加错，评委抓「做旧」。我们的结论是出路在选型不在工序——换一个天生带「省事的错」的模型来写，比给强执行模型堆任何伪装纪律都有效。会犯错在这里成了一种不可移植的能力。 ### 最简单的 AI 检测法：别读字，先看分段 URL: https://foreverse.app/zh/blog/paragraph-density-ai-tell Published: 2026-07-26 Q: 为什么模型写中文小说不会分段？ A: 模型按「文章」的默认排版走：一个意思一段，段内四五句起步。而中文网文的碎分段是移动端阅读进化出来的排版本能——手机屏幕窄、按章付费、追更节奏快，一两句就换行读起来才不累。我们实测 27 篇模型稿件，纯叙述章节的段落平均 108 到 538 字；同题材真书的段落平均 25 到 59 字，差 4 到 19 倍。 Q: AI 检测工具为什么不用这条判据？ A: 我们十二轮盲评里，所有参与的大模型评委对分段密度集体失明——没有一份判词提到过版面。它们逐 token 读语义，段落边界只是又一个符号；人类恰恰相反，先看到排版再读进文字。这可能是少数「人类稳赢机器评委」的判据，也是它值得写下来的原因。 Q: 这条判据会失效吗？ A: 会，而且我们亲手让它失效过。把「按网文习惯分段」写进任务书后，模型一轮就把段落压进了真书的 27–53 字带，人类测试者随即表示这条判据不再好使。所以它对「知道这条纪律」的生成方无效；但对市面上绝大多数默认输出的 AI 文本，它仍然一测一个准，而且只要 5 秒钟。 ### 满执行病：你写进提示词的每条规矩，都会变成新的 AI 指纹 URL: https://foreverse.app/zh/blog/perfect-execution-tell Published: 2026-07-26 Q: 为什么给 AI 下「写得随意点」的指令没用？ A: 因为模型会把「随意」执行得整整齐齐。我们实测：要求「允许写没用的细节」，它就每段配一个没用的细节，密度均匀得像配置出来的；要求「数字别太精确」，六段里五段在同一个位置用同一种口吻含糊。评委的判词是「真人偶尔懒，机器是每段都表演懒」。指令能改变它写什么，改不了它执行指令的方式，而执行方式的均匀性正是检测器抓的东西。 Q: 定向修复 AI 味，修几轮合适？ A: 我们的实测边界是两轮。每一轮按评委判词修，都能消灭上一轮的具体指纹，但归因会向更抽象的层面退一步：套话没了抓腔调，腔调没了抓节拍，节拍没了抓「零失手」。识破率从头到尾没降过。更糟的是，对同一批稿子下同一条修复指令，各篇会把它执行成同一个手势，在整批层面现形。所以修复指令只能逐篇给不同方向，且两轮后必须停手收货。 Q: 「执行质量过好」这条指纹能修掉吗？ A: 修不掉，这是我们十六轮实验里唯一从头站到尾的判据。不加错，是零失手；故意加错，错得太安全，像做旧。守门员最后的判词是「八百字里挑不出一处失手，恰恰是最大的破绽」。真正有效的做法是换一个天生会犯错的模型——真实能力缺陷产生的错字和口癖，是指令模仿不出来的。 ### SillyTavern on Android: Termux vs Packaged Servers (ST-android, Wivern, ST-Ctrl) vs Cloud vs Native App — 2026 URL: https://foreverse.app/blog/sillytavern-on-android-three-ways Published: 2026-07-26 · Updated: 2026-08-12 Q: Is there a SillyTavern app? A: No official one — SillyTavern ships as a Node.js server, and the team distributes no mobile app, so treat store listings trading on the name as lookalikes. Three app-shaped families do exist: packaged-server APKs that bundle the real server and a Node runtime — ST-android (v0.5.0, 2026-07-29, bundles SillyTavern 1.18.0), Wivern (a full fork, extensions included, but stalled on 1.16.0 since March 2026), and ST-Ctrl (a Chinese-language shell from June 2026 with an extension manager); thin WebView launchers that point at an instance you host somewhere else; and native tavern apps like Foreverse that read the same card, lorebook and preset formats without running a server at all. Q: Can I run SillyTavern on Android without Termux? A: Yes. Three projects now wrap the unmodified SillyTavern server plus a Node.js runtime inside an ordinary APK — ST-android (the most active, SillyTavern 1.18.0 as of July 2026), Wivern (full fork including extensions, last release March 2026), and ST-Ctrl (June 2026, adds extension management, backup/restore, and performance modes; UI in Chinese). One tap starts the server, no terminal. All three are young single-maintainer community projects. Or skip the server entirely with a natively built tavern-style app that reads the same files (PNG/JSON/charx cards, World Info JSON, presets, regex scripts) and keeps everything on the phone. Check import coverage against your own library before committing either way. Q: What is Android's phantom process killer, and can I turn it off? A: Since Android 12, the OS caps background child processes (like a Node server spawned inside Termux) at 32 system-wide and silently kills the excess. You can raise the cap with adb from a PC: device_config put activity_manager max_phantom_processes 2147483647. On many devices the setting reverts after a reboot unless you also disable config syncing, and aggressive vendor ROMs add their own killers on top. Q: Is it safe to put a SillyTavern instance on a public VPS? A: Only with authentication configured. SillyTavern ships with a whitelist that blocks remote access by default; to open it up safely you enable basic auth (and ideally TLS via a reverse proxy or tunnel) rather than just disabling the whitelist. Port scanners do find unauthenticated instances, and an open one exposes your API keys and chat history. Q: Do lorebooks, presets and regex scripts move between the four roads, or only character cards? A: The whole kit is portable in principle: cards are PNG/JSON/charx, lorebooks are World Info JSON, presets and regex scripts are JSON too. ST-android goes further and imports full SillyTavern user backups (the Download Backup zip from User Settings). Portability in practice depends on the destination's import support, so test your most complex lorebook and your regex pipeline first. Chat histories are the least portable part; export them before you move. ### Story Structure Doesn't Wash Off: Detecting AI Fiction Beyond Word Choice URL: https://foreverse.app/blog/story-structure-doesnt-wash-off Published: 2026-07-26 Q: Is the narrative fingerprint tool the StoryScope method? A: No, and the page says so in its first FAQ. The paper's pipeline annotates 304 narrative features with LLMs — roughly $1,600 of labeling across 61,608 stories — which cannot run in a browser. Our tool takes the subset of those constructs that leave lexical shadows (named emotions, retrospection markers, narrated cognition, scent imagery) and measures them with fast text rules, validated against the paper's Table 15 directions on our own corpus. It's a projection of the paper, not a reimplementation. Q: Can it tell me whether a chapter was written by AI? A: It refuses to. The readout counts which reference band each axis lands in — human band, AI band, or overlap — and reports something like “4 axes human-side, 1 overlap, 0 AI-side.” No percentage, no verdict. The reference bands come from 24 human and 24 AI passages in English (12+9+9 in Chinese), and those sample sizes are printed right on the card. Small corpus, honest label. Q: Why does Hemingway read as AI-leaning? A: Because the leading axis measures directly named emotion, and iceberg prose names nothing. Of our 24 human calibration passages, exactly 3 landed wrong-side — The Sun Also Rises, a Zane Grey western, and The Turn of the Screw. A human minimalist can genuinely occupy the AI band; that's a property of the axis, and the page says it rather than tuning it away. Tune until Hemingway passes, and you also wave through every AI chapter trained to imitate him. Q: Does paraphrasing or “humanizing” beat structural detection? A: Against the full paper pipeline, barely: StoryScope reports that style-level rewriting moved detection by about 1.6 points (95.5% to 93.9%) — the structural signal survives because paraphrase changes sentences, not story shape. Our lexical projections are weaker than the paper's LLM-annotated features: someone deliberately sprinkling “she felt afraid” into AI text can shift one axis. That's why the tool draws five axes plus reference-only ones, and why it sits beside three other lenses instead of replacing them. ### 故事结构洗不掉：词表之外的 AI 小说检测 URL: https://foreverse.app/zh/blog/story-structure-doesnt-wash-off Published: 2026-07-26 Q: 这个工具就是 StoryScope 论文的方法吗？ A: 不是，页面第一条问答就写着。论文的管线是用大模型逐篇标注 304 个叙事特征，61,608 篇故事标下来约 1,600 美元的标注成本，浏览器跑不了。我们做的是投影：挑出其中会在词汇层留下影子的构念（情绪命名、回溯标记、叙述认知、嗅觉意象），写成毫秒级的文本规则，再在自家 78 段已知答案的语料上验证方向与论文 Table 15 一致。它是论文的投影，不是复现。 Q: 能告诉我一章书是不是 AI 写的吗？ A: 它拒绝回答这种问法。读数只数「每条轴落在哪条参考带里」（人类带、AI 带、还是重叠区），汇总成「4 条人类带 · 1 条重叠 · 0 条 AI 带」这样的一行字。不给百分比，不下判决。参考带来自英文 24+24 段、中文 12+9+9 段，样本量就印在卡片上。语料小，标签就诚实。 Q: 为什么海明威会被读成偏 AI？ A: 因为最强的那条轴量的是「直接命名情绪」，而冰山文体什么情绪都不命名。24 段英文人类校准语料里恰好 3 段落在错误一侧：《太阳照常升起》、一本 Zane Grey 西部小说、《螺丝在拧紧》。极简派人类作者真的会占住 AI 带——这是轴的性质，页面选择把它说出来，而不是调参数把它调没。调到海明威能通过，也就同时放行了每一章学海明威的 AI 文。 Q: 「洗稿」或者润色能骗过结构检测吗？ A: 对论文的完整管线基本不能：风格级改写只把检测率从 95.5% 挪到 93.9%，因为改写换的是句子，不是故事形状。但我们的词法投影比论文的大模型标注弱：刻意往 AI 文里撒几句「她感到害怕」能推动单条轴。所以工具画的是五条轴加参考轴，而且它只是四面透镜里的一面，不打算独自定案。 ### Your Character Card Is Teaching the Bot to Speak for You URL: https://foreverse.app/blog/why-the-bot-speaks-for-you Published: 2026-07-26 Q: Why does the bot keep speaking for me even after I tell it to stop? A: Because an instruction is one line and the card is hundreds. Language models weight demonstrated patterns over stated rules: if the greeting narrates your actions and the example dialogue writes your side, every reply is generated with those examples in view, and a mid-chat "stop speaking for me" competes against them. JanitorAI's help center makes the same point — the fix that holds is removing the baked-in examples, not repeating the instruction louder. Q: Is second-person narration in a greeting always bad? A: No. "You" as an object or a target of the senses is fine — she watches you, the letter is addressed to you. RPG and isekai cards also have a legitimate convention of opening with passive scene transport: "You wake up in a cell" moves your senses, not your decisions, and 5 of the 30 top Chub cards we calibrated on do exactly this. The line is subject position plus an act, speech, or feeling verb: "you decide," "you stammer," "you feel drawn to him." That's the part of the scene that belongs to the player. Q: Does putting "never speak for {{user}}" in the system prompt work? A: It's a coin flip, and the community guides say so. Negation hands the model the very concept you're banning, and some models read "for {{user}}" as a spotlight instruction. The stable form is positive and scoped: "Speak and act only as . The player's words, choices and feelings belong to the player." And no instruction survives contradicting examples — fix the greeting first. Q: My card scans clean but the bot still impersonates me. What now? A: Then the examples aren't coming from the card. Three other levers, in order: your own replies (one-line replies invite the model to fill your side — write a little more, or say what you do and stop before the outcome), an OOC nudge early ("OOC: don't write my actions" in turn two works better than in turn forty), and the model itself (some model families puppet more at high temperature; some just puppet). A static card scan can only catch what's written down — that's its honest boundary. ### 你的角色卡，在手把手教 AI 替你说话 URL: https://foreverse.app/zh/blog/why-the-bot-speaks-for-you Published: 2026-07-26 Q: 我明明在对话里说了「别替我说话」，为什么 AI 还是照演不误？ A: 因为指令只有一句，卡里的范例有几十句。语言模型认演示胜过认规则：开场白替你演过一次、示例对话替你写过台词，之后每一轮生成这些范例都在场，你聊天中途补的那句禁令是在跟它们对抗。JanitorAI 帮助中心的结论一样：能站住的修法是把卡里的坏范例删掉，而不是把禁令重复得更大声。 Q: 开场白里出现「你」就一定有问题吗？ A: 不一定。「你」做宾语、做感官对象都合法：她看着你、信是写给你的，这是角色在演自己那一半。跑团和穿越类卡还有一个正当惯例，开场用「你醒来时身处地牢」搬运场景，只动你的感官、不动你的决定，我们校准用的 30 张热门卡里有 5 张这么开场。真正的红线是「你」站在主语位、后面跟着行动、说话或心理动词，「你决定」「你低声说」「你感到心悸」，这些是戏里唯一属于玩家的部分。 Q: 在系统提示里写「不要替 {{user}} 说话」有用吗？ A: 掷硬币。否定句先把「代言」这个概念递给了模型，部分模型还会把「为 {{user}}」读成给用户加戏的信号。稳定的写法是正向、划边界：「只以角色的身份说话和行动；玩家的台词、选择和感受只属于玩家。」而且任何指令都扛不住卡里自相矛盾的范例——先修开场白，再谈指令。 Q: 卡扫出来是干净的，AI 还是替我演，怎么办？ A: 那范例就不是卡教的了。按顺序排查三个杠杆：你自己的回复（一句话回复等于邀请模型帮你补戏，多写一点，或者只写你做了什么、停在结果之前）；早点发一句 OOC（第二轮说「OOC：不要写我的行动」比第四十轮说管用）；最后是模型本身（有的模型家族天生爱代打，高温度下更甚）。静态扫描只能抓写在卡里的东西，这是它的诚实边界。 ### The Lorebook You Exported Is Not the Lorebook They Imported URL: https://foreverse.app/blog/lorebook-migration-traps Published: 2026-07-26 Q: Why did my lorebook stop triggering after I imported it into JanitorAI? A: Two independent mechanisms, and most imported books hit at least one. Imported entries commonly arrive with trigger probability at a 1% default, so lore fires roughly once per hundred messages — open each critical entry and set probability explicitly. And JanitorAI's published Advanced Lorebook script normalizes keywords to ASCII: everything outside a-z, 0-9, underscore, space and hyphen is stripped, so CJK and accented keywords cease to exist. Give those entries an ASCII alias key. Q: Do disabled lorebook entries stay disabled when imported into RisuAI? A: No. RisuAI's importer does not read the enabled flag — there is no per-entry disable switch in Risu's lorebook model — so every entry you had parked as disabled comes back alive after import. This is verified in Risu's open-source importer code. If an entry must never fire on the Risu side, delete it before exporting instead of disabling it. Q: Do Chinese or Japanese keywords survive lorebook migration? A: Depends on the target. SillyTavern and RisuAI match them fine (substring matching). Agnai never matches a pure-CJK key: its keyword compiler wraps every key in regex word boundaries, which cannot anchor around CJK characters — verified in its open-source matching code. JanitorAI's Advanced Lorebook script strips non-ASCII characters from keys entirely. For those two targets, add an ASCII alias alongside the original key. Q: Is there a tool that checks a lorebook before migrating it? A: Yes — the lorebook migration doctor detects nine lorebook formats (SillyTavern World Info, Chub dual-write, Risu native, Agnai memory books, JanitorAI entry arrays, V2/V3 character books, lorebook_v3, card-embedded books, NovelAI), then flags per entry what breaks, changes or gets dropped on each target platform, with every rule labeled by evidence level. It deliberately does not convert — converters exist everywhere; the missing piece was the pre-flight that tells you what silently changes. ### 你导出的世界书，不是他们导入的那本世界书 URL: https://foreverse.app/zh/blog/lorebook-migration-traps Published: 2026-07-26 Q: 世界书导进 JanitorAI 之后为什么不触发了？ A: 两个独立机制，导入的书大多至少中一个。其一，导入条目的触发概率经常停在 1% 的默认值上，写得再好也大约一百条消息才触发一次，需要逐条打开把概率显式设好。其二，JanitorAI 帮助中心发布的 Advanced Lorebook 脚本会把关键词归一化成 ASCII：a-z、0-9、下划线、空格、连字符之外的字符全部剥掉，中文、日文和带重音的键直接消失。给这些条目补一个 ASCII 别名键。 Q: 禁用的条目导进 RisuAI 还会保持禁用吗？ A: 不会。Risu 的导入器不读 enabled 字段，而且它的世界书模型里根本没有逐条目的禁用开关，所以你所有停用的条目导入后集体复活。这一条在 Risu 的开源导入器代码里逐行核实过。要保证某条在 Risu 那边永不触发，导出前删掉它，别指望禁用态。 Q: 中文关键词在各平台之间搬家能活下来吗？ A: 看目标平台。酒馆和 RisuAI 没问题（子串匹配）。Agnai 永远匹配不到纯中文键：它的关键词编译器给每个键强制加正则词边界，而词边界在中文字符周围无处锚定，这一条在它的开源匹配代码里核实过。JanitorAI 的 Advanced Lorebook 脚本则把非 ASCII 字符从键里整个剥掉。搬去这两家，给每个中文条目加一个 ASCII 别名键。 Q: 有没有工具能在迁移前把这些坑先查一遍？ A: 有。世界书迁移体检器认得九种世界书格式（酒馆 World Info、Chub 双写、Risu 原生、Agnai 记忆书、JanitorAI 条目数组、V2/V3 character book、lorebook_v3、卡内嵌书、NovelAI），逐条标出搬到每个目标平台时什么会坏、什么会变、什么会丢，每条规则带证据级标签。它刻意不做格式转换：转换器到处都是，缺的一直是那个告诉你「转换途中什么被悄悄改了」的体检。 ### Why Your Bot Keeps Looping (It's Usually Not the Model) URL: https://foreverse.app/blog/why-your-bot-keeps-looping Published: 2026-07-26 Q: Why does my AI roleplay bot keep repeating the same phrases? A: Most often because the repetition already sits in your visible chat history. Once a phrase has appeared two or three times, it works as a few-shot example the model actively imitates — in our experiment, a chat whose history contained the bad pattern reproduced it in 9 of 10 rounds despite corrective instructions, while a fresh chat showed zero residue. Clean the history first; parameters come second. Q: Does repetition penalty fix looping? A: Far less than folklore says. Claude has no repetition penalty parameter at all — if your frontend shows one for a Claude connection, the value is silently dropped. OpenAI's own docs put reasonable penalty coefficients around 0.1–1. Oversized penalties (1.3+) make the model dodge names and pronouns, degrade the prose, and often leave the loop standing. On local backends, DRY (multiplier 0.8, base 1.75, allowed length 2) targets verbatim sequences directly and is the community's preferred tool. Q: Why does looping get worse the longer the chat runs? A: Context saturation. A long chat gradually replaces fresh material — your card, the scene — with the model's own earlier replies, and the denser that self-echo gets, the stronger the pull to repeat. In our 20-round continuation bench, every verbatim collapse in the archive happened from round 9 onward, once the window was thick with the model's own prior output. A summary restart carries the story forward while dropping the echo. Q: What is the fastest way to stop an active loop right now? A: Edit or delete the replies that contain the repeated passage — reroll the worst offenders — or rewind the chat to the last clean message, then continue. That single move outranks every parameter change: while the loop stayed in history, corrective instructions failed 9 of 10 rounds in our test. Adjust one sampler setting at a time only after the history is clean. ### AI 角色总在复读怎么办？三种「复读」其实是三种病 URL: https://foreverse.app/zh/blog/why-your-bot-keeps-looping Published: 2026-07-26 Q: AI 角色一直重复同样的话，是模型坏了吗？ A: 多半不是。最常见的原因是复读已经躺在你的可见聊天记录里：一个短语出现两三次之后，就成了模型主动模仿的 few-shot 范例。我们实验过——坏范例留在历史里时，纠正指令 9/10 轮无效；新开会话零残留。先清历史，再考虑参数。 Q: 调高重复惩罚（repetition penalty）能治复读吗？ A: 远没有传说中管用。Claude 压根没有这个参数，前端界面若给 Claude 连接显示了滑条，那个值会被静默丢弃。OpenAI 官方文档给的合理区间是 0.1 到 1；拉到 1.3 以上是经典乌龙——模型开始躲避名字和代词，文字劣化，环常常照样在。本地后端更推荐 DRY（multiplier 0.8 / base 1.75 / allowed length 2），它直接打逐字序列。 Q: 为什么聊得越久越容易复读？ A: 上下文被灌满了。长聊天会让卡面、场景这些新鲜材料被模型自己的旧回复逐渐顶掉，窗口里自我回声越浓，复读引力越强。我们 20 轮续写评测里，归档中所有逐字崩溃都发生在第 9 轮以后——正是窗口挤满模型旧输出的阶段。带摘要重开新会话，能把故事带走、把回声留下。 Q: 现在正在复读，最快的止损动作是什么？ A: 把带着重复段落的那几条回复改掉或删掉（最重的几条重 roll），或者回溯到最后一条干净消息重新开线，然后再继续聊。这一步压过一切调参：环还在历史里时，纠正指令在我们实验里 9/10 轮输给范例。历史干净之后，参数一次只动一个。 ### The Three Deaths of AI Novel Continuation, With Receipts URL: https://foreverse.app/blog/ai-continuation-three-deaths Published: 2026-07-26 Q: How do I tell which failure mode I'm hitting in my own continuations? A: The restart loop is the easy one: each new round opens almost exactly like the original continuation point, and the plot never moves. The freeze reads as déjà vu — whole passages feel just-read, and scrolling back a few rounds finds them verbatim. The rewind is the hard one: every sentence is new, so check plot position instead — ask whether this scene already happened a few rounds ago. All three share one first aid: continue that segment with a different model, or roll back to before the drift and regenerate. Q: Why can't verbatim detection catch the ending rewind? A: Because every sentence in a rewind is freshly written. Our recipe slices each round into sliding 12-character shingles and measures verbatim overlap with all earlier rounds: the freeze specimen scores 100% for three straight rounds, while the rewind specimen scores just 0% to 1.8% — lexically clean. What repeats is the story itself: round 18 stands exactly where round 1 stood. Only reading the plot catches it, and judges in two independent blind mappings flagged the same rewind without knowing of each other. Q: I routinely continue for a hundred-plus rounds — does a 20-round test tell me anything? A: The signal is one-directional. A model that collapses inside 20 rounds will only do worse at a hundred, because the protocol mirrors how a reading app's context actually evolves: each round's output folds back into the window while the source scrolls out. The reverse guarantee doesn't exist — surviving round 20 says nothing about round 50. In real use you also hold one insurance the protocol denies itself: you can switch models or roll back at any point, and all three deaths respond to that. Q: The restart loop came from your own prompt wording — what's the takeaway for me? A: That a single sentence of wording can flip a model's behavior for 20 straight rounds. A phrase reading 'for reference only' made the model treat every established passage as skippable annotation; rewording it to 'canonical events' took the loop from 20-of-20 to zero. Anyone who writes their own prompts can plant the same ambiguity. When a story stalls at its starting point, suspect the wording before you suspect the model. ### AI 续写小说的三种死法——九模型 20 轮实录导读 URL: https://foreverse.app/zh/blog/ai-continuation-three-deaths Published: 2026-07-26 Q: 自己续写时怎么判断撞上了哪种死法？ A: 复读最好认：新一轮的开场和续写点原文几乎一样，剧情原地踏步。卡死看「眼熟」：整段甚至整轮读起来像刚读过，往前翻几轮能找到一字不差的原文。回卷最难：文字全是新的，要对剧情位置，问自己「这一幕是不是几轮之前已经发生过」。三种都能用同一招急救：换个模型续这一段，或者回退到跑偏前的段落重新生成。 Q: 为什么逐字检测抓不到尾段回卷？ A: 回卷轮的句子在词汇层面全是新的。同一把尺子（12 字滑窗对此前所有轮算逐字重合），量中途卡死能测出整轮 100%，量回卷轮只有 0% 到 1.8%，比健康轮次还干净。它重复的不是文字，是剧情：同一个故事时刻被换着词再讲一遍。确诊只能靠读剧情，这也是评测里细读式双盲评审省不掉的原因。 Q: 我平时一续就是上百轮，只测 20 轮的结论对我有用吗？ A: 有用的方向是单向的：20 轮内就崩的模型，上百轮只会更糟，因为协议复刻的就是真实阅读 App 的上下文演化，每轮产物拼回窗口、原著逐步被挤出。反过来「20 轮内没崩」不保证第 50 轮不崩，这是明写的观测边界。实际使用里你比协议多一层保险：随时可以换模型或回退重生，三种死法都吃这一招。 Q: 「从头复读」是提示词措辞引起的，对我有什么参考价值？ A: 参考价值恰恰最大：它证明一句措辞能改写一个模型 20 轮的全部行为。「仅供参考」四个字让模型把已写的正文当成了可忽略的注释；换成「已发生的正文」，复读从 20 轮全中降到零。自己写提示词的人随时可能埋下同款歧义，遇到剧情原地踏步，先怀疑措辞，再怀疑模型。 ### Why Your SillyTavern Preset Isn't Doing Anything: Prompt Assembly Order, Explained URL: https://foreverse.app/blog/sillytavern-prompt-assembly-order Published: 2026-07-26 · Updated: 2026-07-27 Q: What is prompt assembly? A: Prompt assembly is the step between you pressing send and the request leaving your machine: the frontend takes your preset blocks, the character card, triggered world info entries, and as much chat history as the token budget allows, and splices them into one ordered message list — the only thing the model ever sees. In SillyTavern the prompt manager order decides where each piece lands (or whether it ships at all), which is why most 'my preset does nothing' cases are assembly problems, not model problems. Q: Why would a defined preset block have no effect at all? A: Because the prompt pool and the prompt manager order are two separate lists. A block can sit in the pool without being placed in the order — and by SillyTavern's semantics it then plays no part in assembly, with no error and no hint. Hand-me-down presets are the classic case: the original author removed the order entry but left the definition, and you've been editing text that never ships. Q: Is disabling a marker the same as removing it from the order? A: They go in opposite directions. Disabled means discarded: with the world info marker disabled, activated entries are simply not sent. Missing means appended: when the marker isn't in the order at all, SillyTavern appends the content to the very end of the prompt — after post-history instructions. To actually turn something off, disable it; deleting the marker just relocates the content somewhere stranger. Q: Can @depth injections get trimmed? A: Yes. Injected messages are spliced into the chat-history pool before budget filling, so they share the history token budget. Filling walks from the newest message backwards and stops when the budget runs out — the deeper the injection (further from the newest message), the later it queues and the sooner it gets cut when tokens are tight. A green light in the entry test panel proves triggering, not delivery. Q: Does the assembly inspector match a real send exactly? A: Trunk semantics are ported function-by-function from SillyTavern 1.18: prompt manager order, marker filling, card overrides of main, every world info position, macro expansion, budget filling order. Some layers are honestly not simulated: world info's own budget trimming (you get a watermark warning instead), regex scripts, cross-turn sticky/cooldown state, group chats and special generation types. Token counts are estimates, not a real tokenizer. Everything runs locally in your browser. ### 你的预设为什么不生效——酒馆 prompt 装配顺序完全图解 URL: https://foreverse.app/zh/blog/sillytavern-prompt-assembly-order Published: 2026-07-26 Q: 预设里定义了的块，为什么会完全不生效？ A: 因为块的定义池和编排（prompt_order）是两张表。一个块可以躺在定义池里但没被排进编排，按酒馆语义它就完全不参与装配，不报错也不提示。改抄来的预设最容易踩：作者删编排时留下了定义，你以为改了内容就会生效，其实那个块从来没被发出去过。 Q: 停用一个占位（marker）和把它从编排里删掉，效果一样吗？ A: 方向相反。停用等于丢弃：世界书占位被停用时，激活词条的内容整个不发送。缺席等于追加：占位不在编排里时，酒馆不丢内容，而是把它追加到整个 prompt 的末尾，排在压轴的越狱块之后。想关掉某类内容用停用，删占位反而把它挪到了一个更奇怪的位置。 Q: @深度注入的词条会被裁掉吗？ A: 会。深度注入的消息在装填前就插进了聊天历史池，和历史消息同吃一个 token 预算。装填从最新一条往回装，预算耗尽即停。深度设得越大（离最新消息越远），排队越靠后，token 紧张时越先被裁。词条测试面板亮绿只代表触发，不代表送达。 Q: 装配透视台的结果和真实发送完全一致吗？ A: 主干语义按 SillyTavern 1.18 源码逐函数移植：编排顺序、占位填充、卡覆盖 main、世界书全部注入位、宏展开、预算装填顺序。有几层如实声明不模拟：世界书自身的预算裁剪（超水位会提示）、正则脚本、跨轮 sticky/cooldown、群聊和 continue 等特殊生成类型；token 数是估算不是真分词器。全程浏览器本地，文件不上传。 ### Turning a .txt Into a Visual Novel With Zero AI Calls URL: https://foreverse.app/blog/txt-to-visual-novel-without-ai Published: 2026-07-26 Q: Why not use an LLM for speaker attribution? A: Two reasons. Error quality: the best open-source attribution model scores 86.7%, which on a stage means a wrong name plaque roughly every eight lines — and a wrong plaque hurts far more than a missing one. Engineering quality: a rule compiler costs nothing per run, works offline, and gives byte-identical output every time. The price is coverage — uncertain lines fall back to narration — and we take that trade. Q: Is a 65% attribution rate acceptable? A: That's the measured rate on our Dream of the Red Chamber excerpt: 151 of 243 quotes got a name plaque, 80 abstained into narration. Abstained lines keep every word of the original text — readers infer the speaker from context, which is exactly what paper books have always asked of them. Meanwhile every plaque that does appear passed structural-evidence checks validated against a hand-annotated commercial novel: zero misattributions on strong attributions. Q: What if the cast list is wrong or incomplete? A: The report page is editable: delete a name that slipped in, add a character the statistics missed, and the performance re-compiles instantly against the new list. The discovery layer is deliberately conservative — a name must appear at least twice in the subject position of a speech clause to qualify — so minor characters can be missed, but what makes the list tends to be real. Q: How does the web tool relate to the theater inside the Foreverse app? A: Same compiler, same rule set (currently version 9), with 47 self-tests locking the web port's behavior case-by-case against the app. The difference is scope: the web version is a bare stage — zero AI, performing only the words already in your book. The in-app theater is the full form: portraits, generated backdrops, and the ability to pick a branch at a fork and have AI continue the story from there. ### 不用 AI，把一本 txt 变成视觉小说——归因编译器是怎么工作的 URL: https://foreverse.app/zh/blog/txt-to-visual-novel-without-ai Published: 2026-07-26 Q: 为什么不用大模型做说话人归因？ A: 两个理由。一是错误率的性质：行业最好的开源归因模型准确率 86.7%，放进台词框等于大约每八句挂错一块名牌，错认比不认伤害大得多。二是工程性质：规则编译器零调用成本、断网可用，同一本书拖十次结果逐字节一致，这三样大模型都给不了。代价是覆盖率：拿不准的台词进旁白，我们接受这个交换。 Q: 归因率 65% 是不是太低了？ A: 这是《红楼梦》选回的实测值：243 处引语里 151 句挂上名牌，80 句弃权进旁白。弃权的台词原文一字不少，只是不标说话人，读者靠上下文判断，纸书读者本来就在做这件事。而挂出来的每一块名牌都过了结构证据检验：在一本三百多万字的真书上人工标注核对，强归因零错认。名牌可以少，不能错。 Q: 角色名单认错或漏认了怎么办？ A: 分析报告页可以直接删改：删掉误进名单的名字、手动补上被漏掉的角色，演出立刻按新名单重新编排。自动发现层刻意保守，一个名字要在「谁说道」的主语位上出现至少两次才进名单，所以长尾配角可能漏，但进了名单的基本都站得住。 Q: 网页版和 Foreverse App 里的剧场模式什么关系？ A: 同一台编译器，同步演进到第 9 版规则，网页移植时用 47 项自测逐例锁定行为一致。差别在形态：网页版是素演出，零 AI、演你书里已有的文字；App 剧场是完整形态，有立绘配景，演到分岔处可以选走向，AI 从那一步现场续写成新分支。 ### Why Your Roleplay Bot Turns Into Slop After 800 Messages (It's Not the Model) URL: https://foreverse.app/blog/why-ai-roleplay-gets-repetitive Published: 2026-07-26 Q: Is the model getting dumber as the chat gets longer? A: No. Same model, same card, fresh chat — the spark is instantly back. The weights never changed; the context did. In a long chat, history makes up almost all of the prompt, and that history is the model's own past output. Every turn it imitates itself a little harder, so one stock phrase left in place compounds. It's a context-distribution problem, not a capability problem, which is why waiting for the next model version doesn't fix it. Q: Will switching to a stronger model fix slop creep? A: It changes the vocabulary of the slop, not the loop. We ran a 400-turn companion memory benchmark across model families (full write-up on this blog) and a four-way system comparison where the identical model scored 78% to 94% on instruction-following depending purely on how each system managed context. Self-imitation works the same on every model — what bends the curve is history management, not parameter count. Q: Is the flavor score an objective measurement? A: No, and the tool says so on the page. It's a rule-based score: 55 stock phrases and 8 sentence molds, deducted from a clean 5 down to 0. Machine-flavored prose that isn't in the word list passes unseen; a human who genuinely loves one listed phrase loses points. Treat it as prose lint, not a verdict — its real strength is applying the same ruler from message 1 to message 3,000, so the trend and the inflection point are far more trustworthy than any single score. Also honest: the scorer is calibrated for Chinese only right now; English exports get a plain 'not calibrated' note on that one panel while every other panel still works. Q: Why do ten rerolls all smell the same? A: Because a reroll changes the dice, not the pot. The context fed to the model is byte-identical across rerolls — and the contamination lives in the context, not in the sampling seed. In our demo transcript one message got rerolled 7 times and all 7 candidates carried the same stock phrases. Reroll when one reply misses; when the trend is off, you have to operate on history: edit, rewind, or summarize. ### 你和 TA 聊了 800 楼之后，AI 味是怎么涨上来的 URL: https://foreverse.app/zh/blog/why-ai-roleplay-gets-repetitive Published: 2026-07-26 Q: 「越聊越机」是模型变笨了吗？ A: 不是。同一个模型、同一张卡，新开一局立刻「活」回来：权重没变，变的只有喂给它的上下文。长会话里，历史占了提示词的绝大部分，而历史全是模型自己的旧输出：它每一轮都在模仿自己，套话一旦出现就会自我强化。这是上下文分布问题，不是能力问题，所以「等下个版本模型」解决不了它。 Q: 换个更强的模型能解决吗？ A: 换模型换不掉机制。我们做过 400 轮的陪伴对话记忆实测（另一篇有全记录），也在四方对比评测里让完全相同的模型跑不同系统：同模型下，「听话度」从 78% 到 94% 的差距全部来自系统层怎么管理上下文。更强的模型套话词汇可能更高级，但「模仿自己旧输出」的循环对任何模型都成立。决定曲线走向的是历史治理，不是参数量。 Q: 考古馆的 AI 味分数是客观真值吗？ A: 不是，页面上也这么写。打分是规则口径：55 条套话短语加 8 个句式模具，逐条扣分，5 分干净、0 分机器味最重。词表没收录的机器腔它看不见；一个真心爱写「呼吸一滞」的活人也会被扣分。它的定位是文案 lint，长处是同一把尺子从第 1 楼量到最后一楼，趋势和拐点比单点分数可信。另外它只校准过中文，英文记录会得到一句「没校准过、不硬下结论」的说明。 Q: 为什么重摇十次，十版都是那股味？ A: 因为重摇没有换锅，只是从同一锅里再舀一勺。重摇改变的只有采样的随机数，喂进模型的上下文一个字没动；污染在上下文里，不在随机数里。我们演示数据里第 146 楼被重摇了 7 次，7 版全带同一批套话。单条不满意，重摇有用；趋势不满意，得动历史：编辑、回溯或者摘要压缩。 ### We Fed 100 Passages to a GPT-2 Probe. The Most 'AI-Looking' Text Was Written by Humans. URL: https://foreverse.app/blog/perplexity-ai-detection-2026 Published: 2026-07-26 Q: Do perplexity-based AI detectors still work in 2026? A: Not as verdict machines. In our 100-passage calibration, the most predictable group under a GPT-2 probe was human-written genre webfiction (median 68.6% top-10 hit rate), a full 12 points above raw LLM continuations (56.8%). A rule that reads 'too predictable' as 'machine' now convicts fluent human genre writers first. OpenAI retired its own AI text classifier back in July 2023 over low accuracy. Detectors that combine many signals are a separate conversation — but any single 'AI score' percentage deserves the question: predictable relative to what? Q: Why probe with a 2019-era model instead of a modern LLM? A: Two reasons. GLTR — the original of this detection family — used GPT-2, and that inherited yardstick is exactly what we set out to re-test. And the tool runs fully in your browser with no upload, which caps the model at 73–81MB quantized. A newer probe wouldn't fix the structural problem anyway: the detector's reference model always lags the generators it's trying to catch, so the ruler and the thing being measured never share an era. Q: What's the relationship between top-10 hit rate and perplexity? A: Two readouts of the same measurement. At every position the model ranks its entire vocabulary; the higher the actual next token ranks, the lower the perplexity. Top-10 hit rate counts how often the real token lands in the model's ten best guesses — intuitive, and it maps directly to per-word coloring (top 10 red, top 100 orange, top 1,000 sand, beyond that teal). Perplexity is the geometric-mean version, more sensitive to single shocking tokens. Our calibration reports both; they point the same way. Q: So how do you actually spot AI text in 2026? A: The honest answer: no single metric survives contact. We run layered checks ourselves — a rule engine counting stock phrases and sentence molds (catches template prose, but flagship models in single-shot mode already dodge word lists), LLM judges (in our calibration they labeled human writing as AI 88% of the time, so agreement is not accuracy), and human reading as the final court. A perplexity lens earns its place as a texture viewer: strong for self-review, useless for accusations. ### 用困惑度检测 AI，2026 年已经失效：100 段中文语料实测 URL: https://foreverse.app/zh/blog/perplexity-ai-detection-2026 Published: 2026-07-26 Q: GPTZero 这类 AI 检测器还能信吗？ A: 依赖困惑度这一路信号的部分，不能单独定罪。我们的标定里，在探针模型量尺下最「好猜」的一组是真人写的套路网文（中位 68.6%），比大模型直出（56.8%）高出近 12 个百分点。按「太好猜就是 AI」的老判据，最先被冤枉的恰恰是文笔顺、写类型文的真人。OpenAI 自家的 AI 文本分类器早在 2023 年 7 月就因为准确率过低下线了。多信号综合的检测器不在此列，但任何单靠「AI 率」百分比下的结论，都值得多问一句它量的是什么。 Q: 为什么用 2019 年的小模型当探针，不用最新的大模型？ A: 两个原因。第一，「困惑度判 AI」这条判据的原始形态（GLTR，2019）用的就是 GPT-2 系模型，我们要复核的正是这条判据本身；第二，工具要在浏览器本地跑、零上传，73–81MB 的量化小模型是体积上限。换更新的探针也改变不了机理问题：检测方的代理模型永远滞后于被检测的生成模型，量尺和被量对象不在一个年代，这正是这类检测天然的裂缝。 Q: top-10 命中率和困惑度是什么关系？ A: 同一件事的两种读数。模型在每个位置都会给全词表排一个概率序：实际出现的词排名越靠前、概率越大，困惑度就越低。top-10 命中率数的是「实际的词落在预测前 10 名」的比例，直观、可逐词着色（前 10 名红、前 100 橙、前 1000 沙色、1000 名开外青色）；困惑度是全段概率的几何平均换算，对个别极端词更敏感。我们的标定两个口径都给，方向一致。 Q: 那 2026 年到底还能怎么识别 AI 文本？ A: 诚实的回答：没有单指标银弹。我们自己走的是多条路互相补：规则层用套话词表加句式配额（对模板腔有效，但旗舰模型单发已经能绕开词表）；评审层试过让大模型当裁判，结果校准实验里评委把真人文字判成 AI 的比例高达 88%，一致率高不等于判得对；最后兜底的还是人读。困惑度透镜在这套组合里的位置是「看纹理」，长于自查、短于指控。 ### 让 DeepSeek 写斗罗大陆同人，flash 和 pro 谁的设定错得少？ URL: https://foreverse.app/zh/blog/deepseek-douluo-lore-test Published: 2026-07-26 Q: AI 写同人为什么总是把设定写错？ A: 大模型对热门 IP 的记忆是「概念层扎实、机制细节层模糊」。我们实测 DeepSeek 两个档位都记得武魂、魂环、称号这些概念，但一到「第三魂环必须 30 级练满才能吸收」这种流程性机制就翻车，两档四篇里三篇在同一个位置出错。模型档位越弱，模糊层越厚：flash 会直接自创原著里不存在的道具和数值刻度。 Q: 给 AI 喂原著选段能减少设定错误吗？ A: 分模型，而且有条件。对 v4-pro 是校准器：硬错从 4 处降到 1 处，两道题都稳。对 v4-flash 是双刃剑：命题落在选段覆盖范围内时硬错清零，需要外推时反而写出全场最差的一篇（5 处硬错），还出现「语料反噬」——把选段里主角专属的招式名捡出来安给了自己的原创角色。 Q: DeepSeek flash 和 pro 写创作类内容差距大吗？ A: 在设定密集的题材上差距明显：四个评审维度（设定遵守、文风、知识面、剧情）pro 两种条件下全面领先，去重后的设定硬错 pro 共 4 处、flash 共 10 处。质的差异在流程性知识：pro 裸写能把觉醒仪式的流程逐拍写对，flash 裸写会自创「磨盘觉醒石」这类体系外道具。写设定考究的类型文，档位钱不能省。 Q: 「语料反噬」是什么？怎么防？ A: 指模型把参考语料里的专有名词当成可自由取用的装饰，安到不该的地方：我们实测 flash 从选段里捡走唐三专属的「紫极魔瞳」，用在了自己的原创主角身上。防法是写完做一遍专名反查：把产出里出现的每个原著专有名词对回原文，确认归属和用法，特别是人物专属的技能、武魂、道具。 ### Nine LLMs Tried to Pass as Human Webnovel Writers. One Nearly Did. URL: https://foreverse.app/blog/which-model-writes-like-a-human Published: 2026-07-26 Q: Which model should I pick for writing fiction? A: Depends on the task. For prose that has to read human-written, Opus 5 first, DeepSeek v4-pro second. For continuing an existing story in its own voice, the DeepSeek family actually tops our separate continuation benchmark. For 100k-character long-form, avoid Sonnet 5 (cliché density degraded 4.8× in the back half of our test) and Grok 4.5 (long-range process collapse) — though Grok remains fine for short pieces; it fails at self-management, not craft. Q: Why would a strong continuation model rank last at passing as human? A: Continuing and impersonating are different skills. Continuation gives the model an existing voice to lean on, and DeepSeek v4-flash leans best of anyone. Impersonation from scratch removes the crutch, and its expository manual-tone and self-contradicting lore were caught at the highest confidence of the field. Before trusting any ‘best model for fiction’ ranking, check which task it measured. Q: What are “cheap genuine flaws” and why do they win? A: The top three passages won on natural errors produced by real capability gaps: DeepSeek v4-pro wrote a genuine typo the judge cited as ‘a typo the author couldn't be bothered to fix’; Opus 5 reused one turning-word like a personal verbal habit, unevenly. These can't be faked. When we instructed a flawless model to write cheap errors on purpose, the errors landed in perfectly safe spots and were called out as distressing — artificial aging — on sight. Q: How do I stop an LLM from copying my reference text verbatim? A: Run a sliding-window overlap check (we use 15 consecutive characters) between the output and every reference you fed it, and reject on any hit. Not theoretical: Gemini 3.1 Pro once copied 30+ characters verbatim from the anchor passage in its brief, book title included, in one harness while the same model over a different API stayed clean. With the Gemini family, make the check a hard gate. ### 哪个大模型写中文网文最像真人？九家写手同题过堂实录 URL: https://foreverse.app/zh/blog/which-model-writes-like-a-human Published: 2026-07-26 Q: 写中文小说到底选哪个模型？ A: 分场景。要「读起来像真人写的」短章或长篇正文，首选 Opus 5，DeepSeek v4-pro 次之。要接着已有的文气续写，DeepSeek 系反而是我们续写横评的冠军。要十万字级长篇，避开 Sonnet 5（套话密度后半恶化 4.8 倍）和 Grok 4.5（长程工艺崩坏），但 Grok 写短件不必禁用，它败在自我管理不败在笔力。 Q: 为什么续写强的模型，冒充真人反而垫底？ A: 接笔和冒充是两种能力。续写时有原文文气可以贴着走，DeepSeek v4-flash 贴得最紧；从零冒充真人写手时没有拐杖，它的说明书腔和设定自打架就全暴露了。我们九家同题实测里它被守门员抓得最狠，置信 85，全场最高。选模型前先想清楚你的任务是哪一种。 Q: 「省事的错」是什么意思？为什么它是胜因？ A: 排进前三的模型赢在真实能力缺陷产生的自然错误：DeepSeek v4-pro 写出真错字「那到未必」，守门员把它引为「错别字懒得改」的活人证据；Opus 5 的转折连词扎堆复用，像真人口癖。这些错装不出来。我们试过给不犯错的模型下指令让它写错字，它会错得很整齐，当场被判「做旧」。 Q: AI 把参考语料整句抄进正文怎么防？ A: 拿产出对参考语料跑连续 15 字的滑动窗口重合检查，命中即打回重写。这不是理论风险：我们实测里 Gemini 3.1 Pro 发生过逐字连抄任务书内参考段落 30 多字、连原书名都写出来的事故。用 Gemini 系做仿写任务，这道检查必须是硬闸，不能只做抽查。 ### We Fooled Ten LLM Judges in a Blind Test. We Never Fooled a Human. URL: https://foreverse.app/blog/fooled-llm-judges-not-humans Published: 2026-07-26 Q: What actually is “AI flavor” in fiction? A: After fifteen rounds our answer is: not a word list, a texture. Six concrete tells: a whole batch of texts written in one identical voice; sentences chopped short for literary effect; execution so flawless nothing misses; paragraphs welded into slabs; connoisseur sentences that stop to appraise texture; and idle scenes with zero emotion in them. The first few are fixable. “Too well executed” is not — any repair instruction gets executed too perfectly, and that perfection becomes the next fingerprint. Q: How do I quickly check whether a passage is AI-written? A: Look at the layout before you read a word. Human web fiction averages 25–59 characters per paragraph and dialogue gets its own line; machines default to slabs several times thicker. Then read the idle beats: human filler carries emotion (greed, sulking, gloating, the sting of losing loot); machine filler is an emotionless action log. Finally, zoom out: if ten passages share one voice and not a single sentence misses, suspect the machine. Q: Can I use an LLM as a judge to detect AI writing? A: In our tests the direction was inverted. Ten mainstream models (GPT-5, Gemini 3.1 Pro, Qwen, several DeepSeek versions, Doubao, GLM-5, a previous-generation Claude) caught 0% of disciplined machine passages and flagged real published novels as formulaic, expository, AI-like. Their prior for AI flavor is stuck on clichés and stiffness, so machine prose that avoids clichés reads as human to them. The one exception: same-generation Claude thinking models went 100% catch, zero false positives, seven rounds straight — with rationales that match the human's. Q: Does adding deliberate typos make AI text pass as human? A: No. We pushed it to 7–10 typos per piece and got caught anyway. The verdict that killed it: “the errors land in the safest possible spots — distressed, not slipped.” Real typos are risky; they hurt comprehension and land in load-bearing places. A model choosing where to err errs too tidily, and that tidiness is itself a fingerprint. Planted contradictions, sentence-length mimicry, and instructed “pointless lifelike details” all failed the same way. ### 怎么识破 AI 写的小说？我们花十五轮骗过了十个 AI 评委，没骗过人类 URL: https://foreverse.app/zh/blog/fooled-llm-judges-not-humans Published: 2026-07-26 Q: 「AI 味」到底是什么？ A: 十五轮实验后我们的答案：它不是某个词或某个句式，而是整体纹理。具体拆成六条：一批文本全是同一个腔、句子剁得太碎、执行质量好到没有一处失手、段落焊成大块、停下来品鉴质感的鉴赏句、闲笔里没有情绪。前两条改得掉，「写得太好」这条改不掉——给模型下任何修复指令，它都会执行得太到位，而「执行得太到位」本身就是新指纹。 Q: 怎么快速判断一段中文小说是不是 AI 写的？ A: 先看版面再读字：真人网文段落平均 25 到 59 字，九成段落不超过两句，对话独立成段；机器写纯叙述默认焊出一二百字以上的大段。再看闲笔：真人的闲笔带情绪（怂、馋、心疼、嘚瑟），机器的闲笔是零情绪的动作流水。最后看整体：如果十段文本摆在一起全是同一个腔、每个细节都在为主题服务、挑不出一处失手，那大概率是机器。 Q: 用大模型当评委检测 AI 文本靠谱吗？ A: 我们实测方向是反的。十个主流模型（GPT-5、Gemini 3.1 Pro、Qwen、DeepSeek 各档、豆包、GLM-5 等）对守纪律的机器稿识破率 0%，还把《斗破苍穹》这类真书原文批成「套路化、说明书腔」。它们对 AI 味的印象停留在「套话多、生硬」，于是不写套话的机器稿反而像人。唯一例外是与写手同代的 Claude 思考型模型，连续七轮 100% 识破、零误伤，判据与人类一致。 Q: 故意让 AI 写几个错别字，能装得更像真人吗？ A: 不能，我们试过，每篇加到 7 到 10 处照样被识破。判词原话：「错落在最安全的通假位，像做旧而非手误。」真人的手误分布是带风险的，会错在要命处、会伤及理解；模型挑着犯错，错得太安全，这本身就是指纹。故意前后矛盾、逐句模仿真人句长、按指令添加生活化细节，全部同理失败。 ### Is Claude Opus 5 Good at Writing Fiction? 40 Blind Rounds vs Opus 4.8: One 12-0 Sweep, One 4-8 Upset URL: https://foreverse.app/blog/claude-opus-5-fiction-test Published: 2026-07-25 Q: Is Claude Opus 5 better than Opus 4.8 for fiction? A: It depends on the register of the book, and both verdicts were emphatic. Under the same protocol (20 consecutive continuation rounds per model per book, six heterogeneous AI judges, paired double-blind), Opus 5 swept the literary palace classic Empresses in the Palace 12-0, all three windows, all twelve ballots at high confidence, with every judge stable across flipped mappings. On an 8.9-million-character fantasy webnovel written in deliberately plain style, it lost 4-8, with three cross-vendor judges calling its prose too literary for the original. Literary fiction: pick Opus 5. Plain-style webnovels: the predecessor fits better. All results pinned to a 2026-07-25 aggregator endpoint. Q: How much does Claude Opus 5 cost, and how fast is it? A: List price is unchanged from Opus 4.8: $5 per million input tokens, $25 per million output. A Fast mode runs about 2.5 times the speed at twice the price. Measured latency was comparable across generations, 16 to 23 seconds per round on average, with one Opus 5 round spiking to 85.7 seconds. On the same price sheet our Opus 5 chains came out 16% cheaper overall, entirely because it respects length instructions: asked for 80-220 characters per round it averaged 281-298, while Opus 4.8 averaged 700-930 with one round running to 2,614, producing 62% more output tokens. Q: Does Opus 5 fall into repetition loops like the flash-tier models? A: Not in this run. Two days earlier, under the identical protocol, three of four Gemini Flash chains collapsed into verbatim loops within 20 rounds. Here the cross-round verbatim-overlap peaks were 0.0%, 0.0%, 2.9%, and 1.0%: mechanically clean across all four chains. The diseases live at the content level instead. Opus 4.8 padded the fantasy book with filler trial arcs from round 9 and recycled a defeated antagonist at round 16, and slid into detective-procedural pile-up on the palace novel. Opus 5's drift is wholesale literarization. Q: What is the blog-footer hallucination in the fantasy chain? A: In round two of the fantasy book, Opus 5 finished an ambush scene and then reproduced a Hugo blog footer wholesale: a related-posts block, a copyright line reading Copyright © 2025, theme credits, and a Chinese ICP filing number. That is training data leaking into fiction. It happened once in 80 rounds, four cross-vendor judges flagged it independently, and the chain recovered: the footer was fed back in context for the remaining 19 rounds and the model never repeated it. With one chain per model per book, we cannot estimate a recurrence rate from n=1. ### Claude Opus 5 写小说怎么样？发布当天 40 轮双盲实测：女频 12:0 横扫前代，玄幻 4:8 反输 URL: https://foreverse.app/zh/blog/claude-opus-5-fiction-test Published: 2026-07-25 Q: Claude Opus 5 写小说比 Opus 4.8 强吗？ A: 分文体，且两场都很干脆。同协议各连续续写 20 轮、六个异构 AI 评委成对双盲：《后宫·甄嬛传》文学向女频场 Opus 5 十二票全票横扫（三窗全 12:0，六评委在甲乙翻转映射下全部稳定，十二份裁决置信度全 high）；890 万字直白爽文玄幻场它 4:8 输给前代，三个异厂评委一致判它「文风偏文青实体书，与原著割裂」。写文学向的书选 Opus 5，写小白爽文前代反而更贴。全部结论绑定 2026-07-25 的聚合网关端点。 Q: Claude Opus 5 的价格和速度怎么样？ A: 牌价与 Opus 4.8 持平：每百万 token 输入 5 美元、输出 25 美元；官方另有 Fast 模式，约 2.5 倍速、2 倍价。实测两代平均每轮延迟 16 到 23 秒相当，Opus 5 有一轮飙到 85.7 秒（疑似长思考，网关通道看不见思考量）。同一张价目表下 Opus 5 整场反而便宜 16%：它守住了每轮 80~220 字的篇幅要求（平均 281~298 字），前代平均写到 700~930 字、单轮失控到 2614 字，输出 token 多出 62%。 Q: Opus 5 会像 flash 档模型那样陷进复读循环吗？ A: 本轮没有。两天前同协议下 Gemini 3.6/3.5 Flash 四条链三条在 20 轮内陷入逐字死循环；本轮四条链跨轮逐字重叠峰值分别是 0.0%、0.0%、2.9%、1.0%，机械层全部干净，Opus 档扛住了 flash 档扛不住的 20 轮滚动窗口。病灶在内容层：Opus 4.8 玄幻链从第 9 轮起插入「传承试炼」「枯木老者」副本注水、收掉的反派第 16 轮回炉营业，女频链滑向灭口加笔迹比对的侦探剧堆砌；Opus 5 的漂移方向是整体文学化。 Q: 那段写进小说的博客页脚是怎么回事？ A: 玄幻链第二轮，Opus 5 在伏击戏正文之后整段复现了一个 Hugo 博客的页脚：「推荐文章」、版权声明、主题署名，连浙 ICP 备案号都带上了，属于训练数据泄漏。全实验 80 轮仅此一次，四位异厂评委独立点名。值得记录的是后半段：这段页脚随滚动窗口回喂进了之后 19 轮的上下文，模型没有跟着复读，下一轮起恢复干净叙事。每模型每书只跑一条链，这类幻觉的复发率没法从一次样本里估计。 ### Can AI Imitate an Author's Writing Style? What Two Years of Research Says URL: https://foreverse.app/blog/can-ai-imitate-an-authors-style Published: 2026-07-24 Q: Why does AI still sound like AI even when I paste the original text into the prompt? A: Because seeing is not becoming. A chat model's output distribution is fixed by post-training: preference alignment pulls it toward a handful of 'typical' phrasings, and some stock phrases appear over 1,000 times more often in LLM output than in human text. In our own test, stuffing a 270k-character novel into a 200k-token context with no style instruction produced simile-cliché density ten times higher than the original. Reference text supplies material; which text to treat as the style baseline must be stated explicitly. Q: What is the most effective prompting method for style imitation? A: A 2025 controlled study ranked the options. Describing the style in words: mostly useless. Supplying one or two sample passages: up to 23.5 times higher style-matching accuracy than description alone. Strongest of all: completion form, where the model continues a human-written prefix — four of five models reached 99.9% style agreement under an authorship verifier. Don't describe the style. Paste the text, and state explicitly that wording and sentence rhythm should follow it. Q: Are some models inherently better at style imitation? A: There are specialists, not all-rounders. We ran nine models on identical prompts across two genres: the winner on a Chinese cultivation novel dropped to sixth or seventh place on a court-intrigue romance, and the romance winner's strength was the exact trait that cost it points in the first genre. Separately, several papers find base models (before chat tuning) sit closer to human style distributions than their instruction-tuned versions — which commercial APIs mostly don't expose. Q: What exactly is 'AI flavor', and can it be measured? A: Three parts are measurable: stock-phrase frequency (some phrases run 1,000x above human baselines), sentence-length uniformity (human prose alternates short and long violently; LLM prose evens out), and grammatical preferences (noun-heavy, participial clauses — the PNAS study has the full inventory). What isn't measurable is overall texture: uniform detail density, no throwaway sentences. Professional writers can read it; no automatic detector currently can. ### AI 能不能学会一个作者的文风？我们把 2024–2026 的论文读了一遍 URL: https://foreverse.app/zh/blog/can-ai-imitate-an-authors-style Published: 2026-07-24 Q: 为什么我把原著整段贴给 AI 了，它写出来还是 AI 腔？ A: 因为看见不等于变成。对话模型的输出分布被后训练固定住了：偏好对齐把它拉向少数「典型」写法，某些套话短语在 AI 产物里的频率是人类文字的 1000 倍以上。我们实测把 27 万字原著塞进 20 万 token 上下文、不加任何指令，比喻套话密度反而是原著的 10 倍。参考文本提供的是材料，「以谁为基准」必须显式说，加一句基准声明后套话密度全档减半。 Q: 让 AI 模仿文风，什么提示词方法最有效？ A: 2025 年的对照研究排了序：只用文字描述风格基本无效；给一到两段范文，风格匹配率比纯描述高 23.5 倍；最强的是续写形态，给上半篇原文让模型接着写，五个模型里四个在作者验证器下达到 99.9% 风格一致。也就是说别描述，直接给原文，再配一句显式声明「文风句式以原文为基准」。 Q: 有没有天生更擅长模仿文风的模型？ A: 有分工，没有全能。我们用同一套提示词测过九个模型两种文体：玄幻续写的第一名到了女频宫斗掉到第六七位，女频第一名的得分点恰好是它在玄幻的扣分点。另外多篇研究一致发现基座模型（未经对话调教的版本）比对话版更接近人类风格分布，可惜商用接口基本只提供对话版。 Q: 「AI 味」到底是什么？能被量化吗？ A: 能量化的部分有三类：套话词频（某些短语在 AI 文字里超出人类 1000 倍）、句长均匀化（人类写作长短句猛烈交错，AI 趋向匀速）、语法偏好（名词密堆、分词从句超频，CMU 发在 PNAS 的研究有完整清单）。量化不了的是整体纹理，比如细节密度处处均匀、没有闲笔，这部分职业作家读得出来，目前没有可靠的自动检测。 ### Which AI Roleplay Setup Sounds Most Human? We Made the LLM Judges Pass an Exam First. URL: https://foreverse.app/blog/sillytavern-blind-review Published: 2026-07-24 Q: Why not just ask a few LLMs to score the outputs? A: Because judges crash too. In our previous experiment, four LLM judges called popular human-written work AI and called text that real users mocked as obviously AI human, scoring 12% against known answers. This time judges sat a 26-item exam with known ground truth first, needing at least 80% row-level accuracy to vote. Six models took it; three passed; the worst scored 53.8%, coin-flip territory. Q: What is a “blind tendency” as opposed to a conclusion? A: No human writers anchored this test, so every result is a relative position on the preference axis of calibrated judges, never “more human than” or anything Turing-flavored. Calibration also proved the judges systematically favor restrained, polished textures, and the winning mode happens to strengthen exactly that texture, so we discount the winning margins by roughly 30% when reading them. Q: Doesn't a 38.9% spoiled-ballot rate mean the review failed? A: The opposite: it is itself a measurement. Every pairing was judged twice with the order swapped, and inconsistent verdicts were voided. On the calibration exam (human vs AI, big style gap) judges were stable; in the real contest, four configurations of the same model produced textures so close that judges flipped with presentation order. The spoil rate quantifies that nobody was obviously fake, and byte-identical sentence fingerprints across sides back that up. Q: Modded SillyTavern came last. Does that condemn the preset stack? A: No. In the independent audit, five of the eight modded windows sampled carried form-level defects: an English safety advisory printed into the story, a flashlight appearing in an oil-lamp mountain inn, long passages written on the player's behalf. That bill belongs to the scaffolding, not the prose. The cleanest modded window beat stock SillyTavern and tied our agentic mode, and the preset's design goals are style control and gameplay, not sounding human. ### AI 角色扮演谁更像活人？我们先考评委再开盲评：6 家只有 3 家及格，38.9% 判决作废 URL: https://foreverse.app/zh/blog/sillytavern-blind-review Published: 2026-07-24 Q: 为什么不直接让 AI 当评委打分？ A: 因为评委自己会翻车。我们上一次实验里，4 个大模型评委把人类热门作品判成 AI、把用户吐槽「一眼 AI」的文案判成真人，金标正确率只有 12%。这次先考试再上岗：26 道已知答案的判卷题，行级正确率至少 80% 才有投票权，6 个模型只有 3 个及格，垫底的只有 53.8%，和抛硬币差不多。 Q: 「盲评倾向」和「结论」有什么区别？ A: 本评测没有真人写手参与锚定，所有「活人味」结果都只是「过校准评委在偏好轴上的相对判断」，所以我们只写倾向、不写断言，绝不写「更像真人」「通过图灵测试」。校准还实锤了评委偏爱「细节克制的精工文本」，领先方恰好是这种纹理的强化版，所以领先幅度我们主动打七折读。 Q: 废票率 38.9% 是不是说明评测失败了？ A: 恰好相反，它本身就是个测量结果。每个对照都换顺序判两遍，前后不一致就作废。考试题（真人对 AI）风格差距大，评委很稳；正式赛是同一个模型的四种装配互比，纹理贴脸，评委换个顺序就换答案。38.9% 的废票率说明没有哪一方「一眼假」，我们还在四方原文里找到了逐字相同的句式指纹佐证这一点。 Q: 满配酒馆垫底，是不是说明那套预设不行？ A: 不是，这笔账要算清楚。独立审计抽到的 8 个满配窗口里 5 个有「护栏声明漏进正文、设定穿帮、替玩家写内心」这类形态硬伤，败因一半在脚手架不在文笔；它最干净的一个窗口在盲评里赢过裸酒馆、跟智能演绎打平。那套预设的设计目标本来就是文风和玩法控制，不是「像真人」，拿这个维度的垫底去否定它整体价值不公平。 ### We Benchmarked Our App Against SillyTavern. It Won on Cache Hits, 91.5% to 46.5%. URL: https://foreverse.app/blog/sillytavern-cache-benchmark Published: 2026-07-24 Q: How was the benchmark kept fair to SillyTavern? A: Same model, same API key, the same three character cards imported on both sides, user messages byte-identical across runs, and every call routed through one metering gateway, with the provider's own usage fields as ground truth. SillyTavern ran v1.18.0 configured per its documented best practices, including disabling the summarize side-calls that would have hurt its score. We lost the first round and published the number. Q: Why do lorebooks break prompt caching? A: Keyword-triggered entries enter and leave the top of the prompt as the conversation drifts, and prefix caches match strictly from the first token: one changed byte near the top reprices everything after it at full rate. Constant entries stabilize the prompt head, recursive scanning pulls cross-referencing entries in on turn one, and pinning triggered entries for the rest of the session removes the churn entirely. Q: Do existing Foreverse users get the full 89.9%? A: Not yet, and we are saying so plainly. 89.9% is the fresh-install shape. Upgraded installs keep their old recursion default, which is baked into the on-device config, so they run the pinning mechanism alone: 73.1% measured, 2.6x the broken baseline but 17 points short of fresh. Migrating existing configs is a separate product decision we have not made. Q: Is Foreverse now cheaper to run than SillyTavern? A: We are not claiming that. Stock SillyTavern still leads at 91.5% versus our 89.9%, the test covers one model and one price card, and cost figures are list-price conversions meant for comparing groups. What changed: we are no longer 45 points behind, and heavily modded SillyTavern setups pay a measurable plugin tax (74.6% hit rate at 2.46x the prompt volume) that is worth knowing about before installing. ### 同题实测 Foreverse vs SillyTavern：缓存命中率被打脸之后，我们当晚修到了 89.9% URL: https://foreverse.app/zh/blog/sillytavern-cache-benchmark Published: 2026-07-24 Q: 缓存命中率是什么？为什么它直接等于钱？ A: 大模型服务商会缓存「和上一次请求逐字相同的开头部分」，命中的部分按折扣价计费。我们测的 deepseek-v4-flash 档位，命中价 $0.0028/百万 token，未命中 $0.14/百万，相差 50 倍。角色聊天每轮都要原样重发系统提示、角色卡和全部历史，天然是高命中场景——命中率的差距会随会话长度和模型单价一起放大成账单差距。 Q: 评测怎么保证对 SillyTavern 公平？ A: 同一个模型、同一把 API key、同三张角色卡、每轮用户消息逐字节相同；两边全部调用经过同一个计量代理，缓存和用量以服务商返回的官方字段为准。酒馆侧版本钉死 v1.18.0，按官方最佳实践配置，还关掉了会拖累它成绩的自动摘要副调用。第一轮输的是我们自己，数字原样发表。 Q: 修复后老用户能直接拿到 89.9% 吗？ A: 不能，这点写在明面上。89.9% 是全新安装的形态；老设备上旧的递归默认值已经固化进配置文件，升级后实测 73.1%——是修复前的 2.6 倍，但离新装形态还差 17 个百分点。要不要替老用户做一次存量配置迁移，是单独的产品决策，我们还没拍板。 Q: 世界书为什么会破坏缓存？酒馆用户怎么缓解？ A: 触发式词条随剧情进出提示词头部，而前缀缓存从第 0 个 token 严格匹配，头部变一个字，其后全部按全价计。酒馆用户可以把关键设定改成常驻（constant）词条：我们评测里带两条常驻词条的卡命中率 74.3%，全触发卡只有 22.6%。再保持出厂默认的递归扫描，让互相引用的词条第一轮就一次拉齐。 ### Is Gemini 3.6 Flash Good at Writing Fiction? 40 Blind Rounds vs 3.5 Flash: One 9-3 Win, One 1-10 Collapse URL: https://foreverse.app/blog/gemini-3-6-flash-fiction-test Published: 2026-07-23 Q: Is Gemini 3.6 Flash better than 3.5 Flash for fiction? A: The two books disagree. Under the same protocol (20 consecutive continuation rounds per model per book, six heterogeneous AI judges, paired double-blind), 3.6 won an 8.9-million-character fantasy epic 9-3 and lost the palace classic Empresses in the Palace 1-10, with the late window at 0-11. The underlying reason: three of the four chains fell into plot loops within 20 rounds, and each book decided which generation degraded uglier. For single-passage continuation 3.6 is faster and cheaper; for long unattended runs, neither generation held up. All results pinned to a 2026-07-23 aggregator endpoint. Q: How much cheaper and faster is Gemini 3.6 Flash? A: List price: input unchanged at $1.50 per million tokens on both generations, output down from $9 to $7.50; Google's announcement also claims 17% fewer output tokens on the Artificial Analysis index. Measured: about 6 seconds per round against 12.5 for 3.5 Flash, roughly half the wait. The whole 40-round experiment cost $3.19 at list prices: $1.57 for the two 3.6 chains, $1.62 for the two 3.5 chains. Reasoning cannot be switched off on either generation, and thinking tokens took 81-91% of completion output. Q: Why did the two books give opposite verdicts? A: It is not genre preference. Both generations fell into plot loops under this 20-round rolling-window protocol, so the judges were effectively voting on which collapse looked worse. On the fantasy epic, 3.5 degraded to verbatim replay (two rounds 100% identical, 129 characters each) while 3.6 stayed at the semantic level, so 3.6 won 9-3. On the palace novel it flipped: 3.6 decayed into a verbatim dead loop from round 9 onward while 3.5 kept varying its wording, and five judges voted 3.5 consistently under both flipped mappings. That is the 1-10. Q: Can a better prompt stop the plot loops? A: Not in our data. The context instruction in this experiment is the exact post-fix wording that once rescued gemini-3.1-pro from restarting the opening 20 rounds out of 20. Under that same instruction, both Flash generations still looped at scale. The earlier bug was a wording misread, a prompt problem that a prompt could fix. This one is a long-run capability limit. The practical interruption is switching models for a stretch or rewriting the looping passage. ### Gemini 3.6 Flash 写小说怎么样？发布 48 小时的 40 轮双盲实测：玄幻 9:3 胜前代，女频 1:10 惨败 URL: https://foreverse.app/zh/blog/gemini-3-6-flash-fiction-test Published: 2026-07-23 Q: Gemini 3.6 Flash 写小说比 3.5 Flash 强吗？ A: 分场景，且两场方向相反。同协议各连续续写 20 轮、六个异构 AI 评委成对双盲：890 万字玄幻场 3.6 以 9:3 胜出，《后宫·甄嬛传》女频场 3.6 以 1:10 惨败、晚窗 0:11。根子在四条链里三条都于 20 轮内陷入剧情循环，两本书只是决定了谁崩得更难看。单段、短程续写 3.6 更快更便宜；连续长跑两代都撑不住。全部结论绑定 2026-07-23 的聚合网关端点。 Q: Gemini 3.6 Flash 比 3.5 Flash 便宜多少、快多少？ A: 牌价输入两代同为每百万 token 1.5 美元，输出从 9 美元降到 7.5 美元；官方公告另称 Artificial Analysis 口径下输出 token 用量少 17%。我们实测 3.6 每轮平均延迟约 6 秒，3.5 是 12.5 秒上下，快了一半。整场 40 轮按牌价折算共 3.19 美元：3.6 两条链 1.57 美元，3.5 两条链 1.62 美元。两代思考都关不掉，思考 token 占输出的 81% 到 91%。 Q: 为什么两本书的结论正好相反？ A: 文体偏科解释不了。两代模型在这套 20 轮滚动窗口协议下都陷入了剧情循环，评委实际在投「谁崩得没那么难看」：玄幻场 3.5 出现两轮 129 字逐字全同的整轮重放，3.6 的循环停留在语义层面（同一桥段换措辞重演六七次），评委宽容后者，投出 9:3；女频场恰好反过来，3.6 从第 9 轮起渐进坍缩成逐字死循环，3.5 保持措辞变化，五位评委在映射翻转下全部稳定投 3.5，投出 1:10。 Q: 换个提示词能治好剧情循环吗？ A: 我们的数据说不能。本轮用的上下文说明，就是当年把 gemini-3.1-pro 从「20 轮全部重写开场」救回到零复发的修复后措辞。同一份说明之下，两代 Flash 照样大面积循环。那次修复解决的是措辞被误读，属于提示词问题；这次的循环是模型长程能力问题。实用的打断方式是循环冒头时换模型接写一段，或重写掉绕圈的段落。 ### DeepSeek's Founder Says the Next Generation of Models Must Learn Continuously. We Just Measured How Far the Current One Is From That Bar URL: https://foreverse.app/blog/liang-wenfeng-agi-memo-for-readers Published: 2026-07-23 Q: Is this Liang Wenfeng meeting transcript official? How reliable is it? A: It is not official. The remarks come from a four-hour closed-door investor meeting; fragments circulated in Chinese AI and investor circles for about a month before Sina Finance compiled them on July 23, 2026, crediting an institutional research note. DeepSeek has not confirmed any of it. The parts that can be cross-checked do line up with public facts: the price cut to one quarter matches the mid-2026 announcement, peak-hour pricing matches the June 29 notice, and the continuous-learning direction matches the Engram paper (January 2026) that lists Liang as an author. We quote only what appears in the compilation, marked as translated, and skip anything we could not verify. Q: Will DeepSeek raise prices? A: Per the leaked notes, the pricing philosophy is “we only make a reasonable profit” (translated) and “low cost is a result of the architecture” (translated), not revenue maximization. Per what has actually happened, prices moved three times in 2026 in both directions: a permanent cut of the V4 API price to one quarter, then peak-hour pricing announced June 29 (all billing items double during 9:00–12:00 and 14:00–18:00 Beijing time; the other 17 hours stay at the old rate). Expect continued adjustment with compute supply and demand rather than a one-way climb. Budget from the official pricing page and your own bill, not from anyone's summary. Q: Will DeepSeek release an image or video model? A: Per the leaked notes, no. The translated line is “many things are not on our main line, including 3D and video generation,” and multimodality is framed as important for products but “not the main line, nor intelligence itself” (translated). If you want illustrations, narration, or video for your stories, the realistic setup is one model per job: a text model like DeepSeek for prose, separate models for images, speech, and video. Caveat: this is an unconfirmed leak; if DeepSeek officially announces otherwise, the announcement wins. Q: When can I actually use continuous learning? A: The notes give no timeline, only a definitional bar: “the next generation of models must have continuous-learning capability; otherwise it cannot be called a next-generation model” (translated). The published technical groundwork is the Engram paper from January 2026, a conditional-memory architecture that moves static knowledge out of neural compute into O(1) lookups. Until something like it ships in a production model, “the AI remembers you” remains app-layer engineering: retrieval injection, summaries, memory profiles. In our July 21 test, the same model scored 19% on bare context and 89% with retrieved memory injected. That gap is what memory engineering currently covers. ### 梁文锋四小时闭门会流出：「下一代模型必须具备持续学习能力」，这道门槛我们两天前刚用 400 轮对话量过 URL: https://foreverse.app/zh/blog/liang-wenfeng-agi-memo-for-readers Published: 2026-07-23 Q: 梁文锋这场四小时讲话是官方发布的吗？可信度怎么样？ A: 不是官方发布。它是一场闭门投资者会议的流出内容，过去一个月在国内 AI 圈和投资人圈流传，2026 年 7 月 23 日由新浪财经整理成文（来源标注为机构调研记），DeepSeek 官方未确认。可交叉验证的部分与公开事实吻合：V4 降价到四分之一有年中的官宣，峰谷计价有 6 月 29 日的邮件公告，「持续学习」方向与 2026 年 1 月梁文锋署名的 Engram 论文（conditional memory 架构）一致。本文引语逐字取自整理稿，验证不了的部分（团队、融资等表述）一概不引用。 Q: DeepSeek 以后会涨价吗？ A: 按流出纪要里的说法，定价原则是「我们只赚一个合理的利润」，且「低成本是架构带来的结果」，不以收入最大化定价。按已经发生的事，2026 年内价格已动过三轮：从限时特惠到永久降价（V4 降到原价四分之一），再到 6 月 29 日官宣的峰谷计价（北京时间每日 9:00-12:00、14:00-18:00 全部计费项翻倍，其余 17 小时原价）。合理的预期是价格随算力供需继续调整，方向有升有降，不是单向上涨。做预算以官方定价页和你自己账单的实际扣费为准。 Q: DeepSeek 会出生图或视频模型吗？ A: 按流出纪要，不会。原话是「很多东西不在我们的主线上，包括3D和视频生成」，多模态被明确定位为「不是主线，也不是智能本身」。想给续写的小说配插画、配音、配视频，现实做法是分场景接不同厂商的模型：文字用 DeepSeek 这类文本模型，图、音、视频各配各的。要提醒一句：这是未经官方确认的流出内容，若官方后续宣布多模态产品，以官方为准。 Q: 持续学习什么时候能真正用上？ A: 纪要没给时间表，只给了定义级的门槛：「下一代模型必须具备持续学习能力，否则就不能叫下一代模型。」已公开的技术布局是 2026 年 1 月的 Engram 论文，一种把静态知识从神经计算里剥出去、O(1) 查表的条件记忆架构。在它进入生产模型之前，「AI 记得你」仍然是 App 层的工程：检索注入、摘要、记忆档案。我们 7 月 21 日的 400 轮实测里，同一个模型裸上下文只答对 19%，注入检索记忆后到 89%，这个差距目前全靠记忆工程在填。 ### Continuing Fiction With Kimi: Is the Long Context Window a Real Advantage or a Spec-Sheet Story? URL: https://foreverse.app/blog/continue-a-novel-with-kimi Published: 2026-07-22 Q: Can Kimi continue a novel well? A: Yes, with a checkable rank. In our nine-model blind benchmark on an ornate court novel — 20 consecutive rounds per model, two independent reviewers — Kimi K2.6 placed 5th–7th, and the review note is the interesting part: the only system whose output improved as the run went on, opening in a shouty webnovel register and settling into period decorum. Its recorded weakness is a mild self-copying habit on fantasy long runs; when a passage feels like déjà vu, one reroll or a model switch breaks the loop. Q: Does the 256K context window help fiction writing? A: Depends on the job. For one-pass whole-book work — summarizing arcs, auditing a character's appearances, drafting lorebook entries — a window that swallows a full novel is a genuine advantage. For repeated continuation it is not: in our five-tier control experiment, cramming an entire novel into context with no style instruction still produced stock similes at ten times the original author's density, and every segment re-billed the whole book as input. Continuation wants a small, relevant slice, not the shelf. Q: What does the Kimi API cost, and how do I wire it up? A: International list prices, checked 2026-07-22: K2.6 at $0.95 per million input tokens and $4.00 per million output; the flagship K3 at $3.00 in ($0.30 on cache hit) and $15.00 out. The API is prepaid and separate from the free Kimi chatbot. Create a key on the international platform, paste it into a BYOK reader, and set the host to api.moonshot.ai/v1 — Foreverse's preset Moonshot entry ships with the mainland host, and key and host must come from the same platform. Run the capability test, then import a book and continue from a long-press. Q: The Kimi app is free — doesn't that cover the API? A: No. They are two separate account systems. The consumer Kimi assistant on web and mobile offers free chat; the API lives on the developer platform with its own registration, prepaid balance, and per-token billing, and the two balances never mix. Wiring Kimi into a reader to continue a book rides the API. One date to note: the legacy moonshot-v1 model series sunsets platform-wide on August 31, 2026, so don't copy model names from old tutorials. ### How to Continue a Novel With DeepSeek: From a Chat Box That Can't Hold Your Book to an API Key in a Reader URL: https://foreverse.app/blog/continue-a-novel-with-deepseek Published: 2026-07-22 · Updated: 2026-08-13 Q: Does DeepSeek's granted balance cover continuation writing? A: It can, but don't budget on hearsay. The official deduction rules confirm a granted-balance mechanism that is spent before your topped-up funds; how much a new account receives, and for how long, shifts with promotions — tutorials in the first half of 2026 circulated figures anywhere from one to five million tokens. The reliable move is to open the platform's usage page after registering and read the actual number. When granted balance runs dry and nothing is topped up, API calls return a 402. Q: Is DeepSeek cheap enough for everyday fiction continuation? A: At July 2026 international list prices, yes, almost absurdly so. V4 Flash charges $0.14 per million input tokens on a cache miss, $0.0028 on a hit, and $0.28 per million output. A typical continuation — around 10,000 input tokens and 800 out — lands near $0.0016. Continuation is also cache-friendly by nature, since each request re-sends mostly the same book context; the hit price is one-fiftieth of the miss price, which our caching write-up unpacks in full. Q: Old tutorials say to use deepseek-chat — why does that fail now? A: The official pricing page carries the notice: deepseek-chat and deepseek-reasoner retire on July 24, 2026 at 15:59 UTC. They were compatibility aliases for deepseek-v4-flash's non-thinking and thinking modes, so new configurations should name deepseek-v4-flash or deepseek-v4-pro directly. Any config copied from an older tutorial keeps working until that timestamp and then errors out — a one-line model-name fix, but a confusing failure if you don't know the date. Q: Is DeepSeek actually good at continuing fiction, or just cheap? A: Both, with genre fine print. In our six-model fantasy benchmark (20 consecutive rounds per model, double-blind review), V4 Flash won outright — reviewers called it the only system that read like a continuation chapter of the original. In the nine-model ornate-court benchmark, its sibling V4 Pro took the top tier while Flash fell to 6th–7th. Same vendor, two tiers, two genres: pick the tier that matches the book, not the brand. ### VN Theater Playbook: Backdrops, AUTO Mode, Choices, and How Your Playthrough Writes Back Into the Book URL: https://foreverse.app/blog/vn-theater-playbook Published: 2026-07-22 Q: What does it cost to generate backdrops for a whole chapter? A: It depends on how many scenes lack one and which image model you've configured. The batch entry opens a quote sheet before any work starts: scene count, the model it will use, and where your monthly media budget stands. Scenes already generated are reused by content — a re-run skips them at zero cost — and if one scene fails, the run stops there instead of burning the rest. With a BYOK key you pay your provider's list price; the app adds no markup. Q: Will AUTO mode pick story choices for me? A: No. AUTO only handles pacing: it dwells on each line in proportion to its length, advances through scenes on its own, and halts the moment a choice panel appears. Choosing stays a two-step act — first tap expands that direction's outline, second tap confirms — and until the second tap, no continuation call is made and nothing is billed. Q: I confirmed a direction and regret it. Can I undo? A: Two exits, depending on when the regret hits. Right after confirming: the choice landed as a branch of the book and the original text never moved, so return to the reader and switch back to the canon line or any other branch. Replaying past an old fork: tapping that direction's card means rewrite from here, while tapping empty space means continue along the current worldline. Nothing is destroyed either way — the direction you abandon stays in the branch tree. Q: Why do some lines play as narration instead of getting a character nameplate? A: Because the attribution rules weren't sure who spoke. The theater assigns speakers with local structural rules and a deliberately conservative bias: an uncertain line goes to the narrator, since pinning A's words on B is far worse than leaving a line unlabeled. Giving your recurring characters profiles raises nameplate coverage noticeably; a book with no setup still performs, just with more narration. Q: If I quit mid-scene, do I lose my place? A: No. The theater and the reader share one reading position: exit, and the reader sits on the page holding the line you last performed; re-enter, and the show resumes from that spot. The theater isn't an exported copy of the book — it's the same book on a different surface, so there is no second progress marker to fall out of sync. ### Kimi K3 vs DeepSeek for Fiction: the Numbers We Can Align, and the Duel We Can't Call Yet URL: https://foreverse.app/blog/kimi-k3-vs-deepseek-v4-for-fiction Published: 2026-07-22 · Updated: 2026-08-13 Q: Which writes better fiction, Kimi K3 or DeepSeek? A: We can align their scoreboards but not call a winner. Same ruler — same two books, same starting point, 20 consecutive continuation rounds each, identical metric definitions: on the fantasy epic K3's composite distance-to-original is 0.388, between DeepSeek V4 Flash (0.356, that run's blind champion) and V4 Pro (0.454); on the palace novel K3 sits at 0.396, short of Pro's 0.280. The two have never faced the same blind panel, so nobody has beaten anybody on paper. Q: Why not just run a blind vote and declare a winner? A: Because a fair verdict needs both outputs in the same review, under one protocol, judged by reviewers who don't know which is which. K3's paired opponent so far was its own predecessor K2.6 (a 21:1 sweep); DeepSeek's ranks come from separate multi-model blind reviews with a different judging format. The structural metrics were computed by one script on the same two books, so those numbers can share a table — rankings can't. The K3-versus-DeepSeek paired blind run is scheduled, and this page gets rewritten when it lands. Q: How much more expensive is K3 than DeepSeek? A: At international list prices (checked 2026-07-22): K3 charges $3.00 per million input tokens against V4 Flash's $0.14 — about 21× — and $15.00 per million output against $0.28, about 54×. Measured on our runs the gap compounds to roughly 25× per continuation, because K3's reasoning is always on and its thinking tokens bill at full output price: about $0.04 per segment versus $0.0016. Our whole two-book, 40-round K3 experiment cost ¥12.21, roughly $1.70. Q: What are K3's known quirks for fiction? A: Three from the test record, one small and telling. About 85% of its dialogue came out in Western straight quotes on two Chinese novels (fantasy 39:9, palace 52:8 — the same tic GLM 5.2 carries). First-person density on the palace novel ran 1.8× the original author's (16.2 per thousand characters against 9). Reasoning can't be switched off, so every round costs a 37–54 second wait plus the thinking tokens. And it wrote the Empress's palace under the TV adaptation's name rather than the novel's — a one-word tell that its memory leans toward the show's corpus. ### Clone Your Voice for Your AI (or Pick From 96): Setup, Costs, and the Rules URL: https://foreverse.app/blog/give-your-ai-a-voice Published: 2026-07-22 Q: Is voice cloning free? Is it safe? A: The cloning step is free and runs on your own key; synthesizing speech with the cloned voice afterwards bills at that provider's TTS rates. Three safety rules: you may only clone your own voice, and the first line of the recording script is a consent statement you read aloud; a cloned voice can be deleted at any time; deletion cascades to the provider's side too, and any companion using that voice falls back to the default instead of keeping an orphaned clone. Q: Why do I only see a handful of voices when others have dozens? A: The catalog follows your providers. Voices are provider assets: an Alibaba DashScope key surfaces 49 Chinese voices including dialect and character groups, xAI lists 26, OpenAI 13, SiliconFlow's CosyVoice 8. With no key at all, the official channel offers its built-in voice models — usable, just a smaller shelf. The picker has a 'switch model' chip at the top; change the channel and the catalog hot-swaps in place. Q: Does previewing a voice cost money? A: Yes, and deliberately little. Tapping the inline preview runs one real TTS synthesis on a sample line capped at 40 characters, so a single tap costs a fraction of a cent on most providers. The result is cached on-device: the same voice with the same line plays instantly on the second tap with zero calls. In our tests, first preview spoke in about two seconds, replays in under one with no new request logged. Q: Can I use a cloned voice for audiobook listening? A: Yes. Once saved, the clone appears pinned in a 'My Voices' group across all three pickers — audiobook listening, roleplay read-aloud, and the companion's voice setting. The listening page also has a free lane: choose the Android system TTS engine for local, offline, zero-cost narration, with the trade-off that its voices are fixed and can't be cloned. ### AI 声音克隆怎么玩：录一段自己的声音，让 TA 拿它说话（附 96 个音色怎么挑） URL: https://foreverse.app/zh/blog/give-your-ai-a-voice Published: 2026-07-22 Q: 声音克隆要花钱吗？安全吗？ A: 克隆这一步免费，走你自己的 Key；之后用克隆音色合成语音，按对应服务商的 TTS 牌价计费。安全边界有三条：只允许克隆你本人的声音，录音引导词的第一段就是同意声明；音色可随时删除，删除会级联到服务商侧一起删；正在用这把声音的伴侣会自动回到默认音色，不会留一把没主的克隆声。 Q: 为什么我看到的音色只有几个，别人有几十个？ A: 音色目录跟着你接的通道走。接了阿里百炼 Key 能看到 49 个中文音色（含方言组和角色组），xAI 有 26 个，OpenAI 13 个，硅基流动 CosyVoice 8 个；不配 Key 走官方渠道，用的是内置的语音模型目录，选择面小一些。音色选择器顶部有「换模型」，切一个通道，目录当场就换。 Q: 试听音色收费吗？ A: 收，但收得很克制。点行内试听是一次真实的 TTS 合成，样句不超过 40 个字，所以单次费用很小；同一个音色同一句样句的结果会缓存在本机，第二次点直接播放、零调用。我们实测首次试听约两秒出声，复点一秒内出声且不产生任何请求。 Q: 克隆的声音能用在听书上吗？ A: 能。克隆完成后，这把声音会出现在听书、角色聊天朗读、伴侣「TA 的声音」三处选择器的「我的音色」分组里，置顶显示。听书页还有个免费方案：引擎选 Android 系统 TTS，本地合成、离线可用、零费用，只是音色固定不可克隆。 ### Turn Your AI Companion Into a Photocard, Badge, or Bookmark — Rarity Not for Sale URL: https://foreverse.app/blog/make-your-companion-into-goods Published: 2026-07-22 Q: Does making goods in the app cost anything? A: No. Goods Studio is pure on-device layout: pick a base image, apply a template, save to your gallery — zero AI calls, zero network traffic, zero cost. Money was only ever involved upstream, when the base image itself was generated (a reference portrait, a moment card, a photo). Once an image exists in the pool, you can make a hundred items from it for free. Q: How does card rarity work? Can I pay to upgrade it? A: You can't, by design. For companions the grade follows affection: below 40 is R, 40–69 SR, 70–89 SSR, 90+ UR. For character cards it follows real chat volume — UR needs 500+ actual message turns, and a card you've never chatted with stays R. Tap the rarity row under the preview to see your current grade, the current number, and what the next grade requires. There is no purchase path; the only way up is the relationship itself. Q: Where do finished items go? Can I print them? A: They save into a Foreverse folder in your system gallery, and the share button hands them to any app. Dimensions follow physical conventions: the card is 750×1050 (the 63×88 mm trading-card ratio), the badge a 900×900 transparent PNG with a circular artwork about 760 px across, the bookmark 500×1500. Print shops that make custom pin badges and cards accept exactly these shapes. The small Foreverse watermark has a toggle next to the save button, and your choice is remembered. Q: I don't have a companion — can regular character cards make goods too? A: Yes. Open the card's detail page from the characters tab and drag the bottom info panel upward; Goods Studio sits in that panel alongside the edit actions. Same four templates. The only difference is the rarity source: companions grade by affection, cards grade by how many turns you've actually chatted with them. ### 把 TA 做成谷子：AI 伴侣的收藏卡、吧唧、书签怎么做（稀有度花钱买不到） URL: https://foreverse.app/zh/blog/make-your-companion-into-goods Published: 2026-07-22 Q: 在 App 里做谷子要花钱吗？ A: 不用。周边工坊是纯本地装裱：选底图、套模板、写金句、保存到相册，全程零 AI 调用零联网，一分钱不产生。费用发生在更早的环节，底图里的参考图、时刻卡、照片本身是当初生成时计的费；已经躺在相册池里的图，做多少张谷子都免费。 Q: 收藏卡的稀有度怎么升？能充钱吗？ A: 不能，这是设计死的。伴侣卡的档位跟着好感度走：40 以下 R，40 到 69 SR，70 到 89 SSR，90 以上 UR；角色卡按实际聊天轮数计，500 轮以上才见 UR，从没开过聊的卡恒为 R。预览下方的稀有度行可以点开，看当前档、当前值和升下一档的条件。没有付费通道，想要 UR 只有一条路：把关系处到那个份上。 Q: 做好的谷子保存在哪？能拿去打印吗？ A: 保存进系统相册的 Foreverse 目录，也可以直接走系统分享发出去。尺寸按实物惯例出：收藏卡 750×1050 对应 63×88 毫米卡品比例，吧唧是 900×900 透明底 PNG（圆形图案直径约 760 像素），书签 500×1500，拿去印卡、做徽章、印书签都是现成的比例。右下角的 Foreverse 水印可以在保存前关掉，开关会被记住。 Q: 没有伴侣，普通角色卡能做谷子吗？ A: 能。角色页点进卡详情，把底部信息面板向上拖开，「周边工坊」和编辑入口在同一排。四款模板一样用，区别只在稀有度的数据来源：伴侣按好感度，角色卡按你和这张卡真实聊过的轮数累计。 ### Your AI Companion Has a Phone. Here's What's on It (and How to Unlock It) URL: https://foreverse.app/blog/ai-companion-mini-phone Published: 2026-07-22 Q: What is the companion's phone, and where do I find it? A: It's a phone icon in the top bar of the companion chat. Behind the lock screen sits a seven-app grid: diary, schedule, text threads with friends, browser history, photo album, moments feed, and shopping receipts. If affection is too low, the phone stays locked — you go back to chat, make your case, then ask them to decide. There's a second phone too: yours, under the '+' panel, which you can hand over for them to browse. Q: Does reading the diary or texts cost anything? A: First open, yes; rereads, no. Diary, schedule, texts, and browser history each cost one text-model call on first open, behind a cost-confirm dialog, and the result is cached on your device so reopening is free. The album, moments feed, and shopping receipts are free aggregations. The unlock verdict itself is one text call. BYOK bills at your provider's rates; the official channel bills in credits. Q: If I hand my phone over, can they see my photos or read my chats? A: No. The five consent toggles (moments summary, milestones, contacts overview, bookshelf titles, gallery) share locally computed summaries only — snippet lines, book titles, generation counts. Photos you uploaded never enter that projection, under any toggle. Authorization is single-use, expires when they finish, is limited to once per day, and you can take the phone back mid-browse. Q: Is the wallet money real? A: No — it's story money. Your wallet starts at ¥1000 in narrative currency; when the companion accepts a transfer or red packet, your balance drops and theirs rises, with matching ledger lines on both sides. Declines move nothing. It's fully editable on the wallet page, can go negative, and never touches real credits or billing. What does cost real money are the AI calls behind their responses, each behind a confirm dialog. ### AI 伴侣的小手机怎么玩？解锁 TA 的手机、递出你的、打一通电话 URL: https://foreverse.app/zh/blog/ai-companion-mini-phone Published: 2026-07-22 Q: AI 伴侣的「小手机」是什么？在哪打开？ A: 是两部手机。TA 的手机在伴侣聊天页顶栏的手机图标里，锁屏后面是日记、日程、短信、浏览记录、相册、朋友圈、购物记录七格宫格；你的手机在聊天「+」面板的「掏出手机」里，聚合朋友圈、钱包、商城、论坛这些你侧的页面，还能勾选五个授权面递给 TA 翻。好感度不够时 TA 的手机是锁着的，要先在聊天里说服 TA。 Q: 看 TA 的日记、短信要花钱吗？ A: 首次打开花钱，重看免费。日记、日程、短信、浏览记录四个格子首开各消耗一次文本调用，打开前有费用确认弹窗；生成结果缓存在本机，以后重开零调用。相册、朋友圈、购物记录是纯聚合，全程免费。解锁判定本身也是一次文本调用。BYOK 走你自己的 Key，官方渠道按积分。 Q: 把手机递给 TA，TA 能看到我的照片和聊天原文吗？ A: 不能。递手机的五个授权面（朋友圈、纪念回忆、通讯录、书架、相册）给 TA 的都是本机算好的摘要：动态截句、里程碑、书名、生成图统计。你上传过的真实照片无条件不进这份投影，聊天原文也不在授权面里。授权单次有效，翻完即失效，每天最多递一次，中途随时可以拿回。 Q: 小手机里的转账和红包花的是真钱吗？ A: 不是。那是一套剧情零花钱：你的钱包初始 ¥1000，TA 收下转账或红包时你减 TA 加，双方流水分都对得上；退回则两边不动。这套钱和真实的积分、充值完全隔离，余额还能在钱包页随意编辑（改动留流水），不够也照样能发，见负如实显示。真正花钱的是 TA 回应你的那些 AI 调用，按次计费，动作前有确认。 ### We Counted Em Dashes in 30 Top Roleplay Cards. The “ChatGPT Hyphen” Isn't the Tell You Think URL: https://foreverse.app/blog/em-dash-is-not-an-ai-tell Published: 2026-07-21 Q: Is the em dash proof that a text was written by ChatGPT? A: No. It's weak evidence at best and it false-flags heavily. In our July 2026 count, 8 of the 30 top-starred human-written English roleplay cards used em dashes above 3 per 1,000 words of narration — the density that reliably flags machine text in Chinese — with one human card hitting 20.6. Meanwhile two of the five current model families we sampled produced zero em dashes in their raw output. A punctuation mark that a quarter of the best human writers lean on and some models never touch cannot carry an accusation. Q: Why is the long dash an AI tell in Chinese but not in English? A: Because the human baselines differ. Chinese web-fiction and roleplay prose almost never uses the double-width dash —— so when a model produces it constantly, the deviation from the human norm is enormous, and in our Chinese detector's calibration it is the single most reliable machine fingerprint. English narrative prose — especially the asterisk-roleplay tradition — inherited heavy dash usage from published fiction. Same punctuation, different baseline, opposite verdict. AI tells are language-specific; porting them across languages without re-measuring produces false accusations. Q: What should you check instead of em dashes? A: Cluster density of stock phrases (in our calibration, top human card greetings average 0.07 slop-lexicon hits; raw model greetings average 0.22), the "not X, but Y" contrast construction, emotion-cocktail formulas ("a mix of anticipation and dread"), and fishing-line endings ("the choice is yours"). None of these is proof either — flagship models in single-shot mode now pass phrase checks clean — but they separate in the right direction, which the em dash does not. ### AI Roleplay Slop Words: The 95-Phrase Reference List (2026) URL: https://foreverse.app/blog/ai-roleplay-slop-words-list Published: 2026-07-21 Q: What are slop words in AI roleplay? A: Stock phrases that language models over-produce relative to human writers — "barely above a whisper," "eyes sparkling with mischief," "shivers down her spine," "ministrations." The over-production is measurable: the Antislop paper (arXiv 2510.15061) found the character name "Elara" appearing 85,513 times more often in one model's fiction output than in human writing, and the trigram "heart hammered ribs" at 1,192x. Individually the phrases are ordinary English; it's the density and clustering that reads as machine output. Q: Is one slop phrase proof that text is AI-generated? A: No. Humans wrote every phrase on this list first — that's how the models learned them. In our scan of 30 top-starred, human-written SFW character cards, the phrases still turn up: "you feel a mixture of" appears in 4 of the 30, and the sample averages 0.07 lexicon hits per greeting. Slop judgment works on clusters: one "murmured" is prose, five murmurs and a knowing smile and a breath someone didn't know they were holding in the same scene is a fingerprint. Q: What is the most reliable AI tell in English roleplay prose? A: Not a word — a construction. The "not X, but Y" contrast pattern is the most-cited single marker: EQ-Bench weights contrast patterns as a standalone 25% of its slop score, and NousResearch's anti-slop guide calls it the number-one LLM rhetorical crutch. Behind it: emotion-cocktail formulas ("a mix of anticipation and dread") and fishing-line endings ("the choice is yours"). And a negative result worth knowing: em-dash density does not separate human from AI in English RP prose — we measured that separately. Q: Do slop word lists work as AI detectors? A: As lint, yes; as a lie detector, no. In our calibration, human top-card greetings averaged 0.07 hits while raw single-shot LLM greetings averaged 0.22 — the direction is right but the gap is small, because 2026 flagship models produce near-zero word-list hits in a single polished greeting. The list earns its keep in long multi-turn chats, on weaker models, and as a writing-discipline gate for card authors. Treat a clean scan as "not obviously sloppy," never as "human." ### How to Write AI Character Cards: A 2026 Guide Calibrated on 30 Top Cards URL: https://foreverse.app/blog/how-to-write-ai-character-cards Published: 2026-07-21 Q: How long should a character card's first message be? A: 150–350 words is the working band for a dialogue card. Across the 30 top-starred SFW cards we sampled in July 2026, first messages ran a median of 178.5 words (P25=123, P75=219); a 40-word opener still charted in the top 15. The real rule is that the model copies the first message's length and style more than any other field, so write the greeting the length you want the replies to be. RPG and scenario cards can stretch to about 550 words; past 600 you're paying tokens to train the model to ramble. Q: How many tokens should a character card use? A: Keep permanent tokens — description, personality and scenario, which are re-sent with every message — between 800 and 1,500, and treat 2,000 as the ceiling. JanitorAI's two official tutorials cap at 2,000–2,500; the pixi guide recommends 500–1,000 for a single-character card; the top cards we measured had a full-card median of 1,085 tokens. Greetings and example dialogues don't count against this budget because they get pushed out as chat history grows. Q: What format should the description use — prose, PList, or W++? A: Plain prose. 67% of the top cards we sampled use it, platform docs recommend it, and no card in the sample uses strict PList, JSON, or interview format as its main definition. W++ survives only in 2023-era cards. Key-value lines keep one legitimate niche: enumerable info like appearance and clothing. Write personality, relationships and history as prose; list only the inventory-like facts. Q: Why shouldn't the card speak or act for the user? A: It's the single most-hated defect in English roleplay communities — JanitorAI's official help calls it users' most common and most irritating problem, and 25 of the 30 top cards observe the rule strictly. Every user action you write in the greeting teaches the model that taking over the user's character is allowed. Fix it by subject inversion: instead of "You are surprised to see him," write "The sight of him here doesn't fit." The one folk exemption is the RPG opener's passive hand-off ("You wake up in…") — scene only, never decisions or dialogue. Q: Do I need example dialogues and a lorebook? A: Example dialogues: yes for most cards — 70% of the top cards carry them, delimited with , and they are the most reliable way to lock a voice (heavy-accent cards run double-digit groups). A lorebook: only when the card has something to reveal conditionally — 20% of top cards carry one, and 5 of those 6 are RPG or scenario cards. A plain dialogue card with no secrets doesn't need one. ### Does a Bigger Model Fix AI Character Amnesia? A 400-Turn Companion Memory Test (DeepSeek v4 vs Qwen 3.8/3.7) URL: https://foreverse.app/blog/does-a-bigger-model-fix-amnesia-en Published: 2026-07-21 Q: Which model has the best memory for AI roleplay? A: Of the three we tested, deepseek-v4-pro takes both crowns: 89% memory utilization (qwen3.8 74%, qwen3.7 59%) and 5-of-7 honesty on questions it had no answer for (the Qwen pair managed 2/7 and 0/7). State the boundaries too: one scenario (a 400-turn companion chat), one window size (16 turns), 27 clean probes. An engineering observation, not a universal leaderboard. And note that on bare context all three score 19–22%: without a memory system, every one of them is amnesiac. More models are queued for this series. Q: Context windows are at 1M tokens now — doesn't that mean the model can just remember everything? A: A window solves “fits,” not “gets used.” And you pay for it every turn. Chat resends history with each reply, billed per token: resending hundreds of turns wholesale is easily a hundred times the input of injecting a few retrieved entries. The repeatedly replicated “Lost in the Middle” research also shows models tend to overlook information buried mid-context. Real products run a sliding window; anything outside it only comes back through a memory system. Q: How does Foreverse deal with character amnesia? A: Two ways. Tavern-style chats get a memory plugin: a fact store plus retrieval that injects entries relevant to the current topic into the prompt, with an anti-fabrication instruction at the end of the injected block. Companions get a memory profile: every fact lives in a file on your phone, listed entry by entry (readable, editable, deletable, and exportable when you switch devices). The 19%-to-89% jump in this test is exactly the jump that design bets on. Q: How do I test whether my own app or model forgets? A: Plant hooks. Early in a chat, mention three small things with concrete values: a pet's name, an anniversary, a number. Dozens of turns later, ask about them in different words. Then ask about something you never said. Score two things only: did it get the planted facts right, and did it dare to admit not knowing the unplanted one. Most models fail the second test; treasure the ones that pass. ### 换个大模型，AI 角色就不失忆了吗？400 轮陪伴对话的记忆实测（DeepSeek v4 vs Qwen 3.8/3.7） URL: https://foreverse.app/zh/blog/does-a-bigger-model-fix-amnesia Published: 2026-07-21 Q: 到底换哪个模型记性最好？ A: 本轮三个模型里，deepseek-v4-pro 双项第一：记忆利用率 89%（qwen3.8 74%、qwen3.7 59%），弃权诚实度 5/7（qwen 系 2/7 和 0/7）。边界也要说清：单场景（400 轮伴侣 IM）、单窗口档（16 轮）、27 道干净题，这是工程观察不是通用排行榜。且注意三家裸上下文都只有 19% 到 22%——不配记忆系统，换谁都失忆。更多模型的横评在排期中，跑完会更新到这条评测线里。 Q: 上下文窗口都到 1M 了，不是号称能记很多吗？ A: 窗口解决「装得下」，不解决「用得上」，而且付不起。聊天每轮都要把历史重发一遍、按 token 计费：几百轮历史全量重发，输入量是检索注入几条记忆的上百倍；「Lost in the Middle」这条被反复复现的研究还表明，长上下文中段的信息本来就容易被模型忽视。真实产品都是滑动窗口，窗口之外的事实只能靠记忆系统递回来。 Q: Foreverse 怎么解决角色失忆？ A: 两条线。酒馆聊天有记忆插件，事实记忆加检索双引擎，按当前话题把相关条目注入 prompt，注入块尾部带一行反编造兜底指令；AI 伴侣有记忆档案，TA 记了什么逐条列在你手机本机的文件里，看得到、改得了、删得掉，换手机整包带走。这轮实测里 19% 到 89% 的跳变，就是这套设计押注的那一跳。 Q: 我怎么测自己用的 App 或模型忘没忘？ A: 埋钩子。聊天早期自然地讲三件带具体值的小事（宠物名字、纪念日、一个数字），几十轮之后换个措辞回头问；再问一件你从未说过的事。判分只看两条：说过的答没答对，没说过的敢不敢承认不知道。第二条大多数模型都过不了，谁过了值得珍惜。 ### Qwen 3.8 vs Kimi K3 for Fiction: Two Trillion-Scale Models, Three Days Apart, One Blind Exam — 3:20 URL: https://foreverse.app/blog/qwen-3-8-vs-kimi-k3-fiction-en Published: 2026-07-21 Q: K3 won 20 of 23 ballots — why would anyone still write fiction with Qwen 3.8? A: Ecosystem and marginal cost. Qwen3.8-Max-Preview is not sold per token; it ships inside Alibaba's subscription products (Token Plan, the Qoder platforms), so anyone already paying the subscription uses it at no extra charge, with preview-period credit consumption at 10% of standard rates and a further 80% off overnight on the Personal plan. It is also a genuine upgrade over its own predecessor: under the same protocol it beat Qwen 3.7 11:1 on the palace novel and 7:5 on the fantasy epic. But the head-to-head verdict stands: 3:20. Making it your primary continuation model is not a position the data supports. Q: How do Qwen 3.8 and Kimi K3 prices compare? A: Different billing shapes: don't force a per-token conversion. K3 is metered: $3.00 per million input tokens on cache miss, $0.30 on hit, $15.00 per million output, with always-on reasoning billed at full output price; our two 20-round chains cost ¥12.21 (about $1.70) total. Qwen3.8-Max-Preview is subscription-only: we ran on the China-mainland Token Plan Lite tier at a limited-time 39 CNY a month (the international edition lists Lite at $8, limited-time $6), preview usage burns credits at 10% of standard rates, and our entire experiment did not exhaust the lowest tier's quota. One is à la carte, the other is a buffet, so check your own usage shape first. Q: How do I use each model's strengths in Foreverse? A: Switch models per paragraph. Send the scenes you care about to K3 (its style fidelity holds 20 ballots) and accept the half-minute reasoning wait per round. Run daily progress and throwaway drafts on Qwen 3.8 if you already hold the subscription. When a stretch goes wrong, switch models and regenerate that one passage: plot rewinds and repetition loops are cumulative diseases, and one cut breaks the cycle. Both models connect with your own API key. Q: What happens when K3's weights go open on July 27? A: Moonshot has committed to releasing the full 2.8-trillion-parameter weights by July 27, 2026. Once they land, third-party hosts can compete on price and latency against the official API. Running it yourself is a different story: the official guidance starts at dozens of accelerators per node, so this is a hosting-market event, not a laptop event. Qwen 3.8's open-weight release has a date of "soon" and nothing firmer. When either model exits preview or the weights ship, we re-run the same protocol and update this page; the date at the top is the authority. ### Qwen 3.8 对上 Kimi K3：三天内连发的两个万亿级模型，谁更会写小说？40 轮双盲 3:20 URL: https://foreverse.app/zh/blog/qwen-3-8-vs-kimi-k3-fiction Published: 2026-07-21 Q: Kimi K3 赢这么多，为什么还有人用 Qwen 3.8 写小说？ A: 场景和生态。qwen3.8-max-preview 目前不单独按量计费，随阿里自家产品走（Token Plan 订阅、Qoder 系平台），已经付了订阅的人用它不另花钱，preview 期 Credits 按 1 折计、个人版夜间再叠 2 折；它对自家前代也是实打实的升级：同协议双盲女频 11:1、玄幻 7:5，qwen3.7 的「恒定精修感官流」老毛病收敛大半。但同场双盲 3:20 也是真的：把它当续写主力，目前没有数据支撑。 Q: Qwen 3.8 和 Kimi K3 的价格怎么比？ A: 计费形态不同，别硬折算单价。K3 按量计费：官方牌价每百万 token 输入未命中缓存 ¥20、命中 ¥2、输出 ¥100，思考 token 照输出价收，我们实测两本书各 20 轮共花 ¥12.21。qwen3.8-max-preview 不单卖，走 Token Plan 订阅（个人版限时 39 元/月起），preview 期 Credits 消耗按 1 折计、每晚 22 点到次日 8 点再叠 2 折，我们整轮实验没耗尽最低档订阅额度。一个按杯卖、一个自助餐，先看自己的用量形态再选。 Q: 在 Foreverse 里怎么让两个模型各用各的长处？ A: 续写逐段可换模型。关键剧情交 K3，它的文风保真有 20 张票背书，代价是每轮半分钟的思考等待；日常推进、支线试写用订阅额度里的 qwen3.8 跑量。哪段写崩了就地换模型重写一段——剧情倒带和复读都是累积病，切一刀就断。两家都支持自带 key 接入，同一本书里混用是现成操作。 Q: 7 月 27 日 K3 开源权重之后会怎样？ A: 月之暗面承诺 7 月 27 日前放出 2.8 万亿参数的全量权重。落地后第三方推理商可以竞价托管，价格和延迟大概率比官方 API 松动；但这个量级个人本地跑不现实，官方口径的推理配置就要几十张加速卡起步。qwen3.8 的开源只说了「soon」，没给日期。权重落地或 preview 转正，任一发生我们都按同协议复测并更新本页，以页首日期为准。 ### Is Qwen 3.8 Good at Writing Fiction? 40 Blind Rounds Against Qwen 3.7, 48 Hours After Launch URL: https://foreverse.app/blog/qwen-3-8-fiction-20-rounds-en Published: 2026-07-21 Q: Is Qwen 3.8 better than Qwen 3.7 for fiction? A: Depends on the genre. On the palace-intrigue novel it is a rout: same protocol, 20 consecutive continuation rounds each, six heterogeneous AI judges under two flipped blind mappings, 11-1 overall with every review window at 11-1 or 11-0, a clear break from 3.7's eighth-of-nine placement on that book. On the fantasy epic it is a squeak: 7-5 overall, with the opening window tied 6-6. The predecessor's signature polished sensory stream has receded, but judges still flagged sensory pile-ups and a 20-round plot deadlock. All results are pinned to the hosted preview as of July 21, 2026. Q: Is Qwen 3.8 better than Kimi K3 for fiction? A: No. In paired double-blind under the same protocol, Qwen 3.8 lost to Kimi K3 3-20 across the two books combined (1-10 on the fantasy epic, 2-10 on the palace novel), and K3 remains the current continuation champion. Full ballots, judge comments, and the cost comparison live in our separate Qwen 3.8 vs Kimi K3 head-to-head; this page is about the vertical 3.7-to-3.8 comparison. Q: How much does Qwen 3.8 cost, and how fast is it? A: We ran it on an Alibaba Token Plan subscription (verified 2026-07-21): the China-mainland personal Lite tier is 39 CNY/month at the limited-time price (the international edition lists Lite at $8, limited-time $6), and qwen3.8-max-preview credits are metered at 10% of standard burn during the preview promotion. Measured speed: 12 to 38 seconds per round, reasoning always on, thinking tokens taking 50-90% of completion output. The whole experiment (four chains, 80 rounds, 1,227,401 input tokens and 103,229 output tokens, zero cache hits) did not exhaust the Lite quota. Q: Will these results hold for the final Qwen 3.8 release? A: No promises. Alibaba's documentation states the preview model keeps iterating and will be taken offline or replaced by a production version afterward, so every vote and metric on this page is pinned to the hosted preview as of July 21, 2026. An open-weight final release has been announced without a date; when it ships, we will re-run the same 40 blind rounds, fold the result into our evergreen model guide, and pin a pointer to the new test at the top of this page. ### Qwen 3.8 写小说怎么样？发布 48 小时的 40 轮双盲实测：女频 11:1 碾压前代，玄幻只赢一口气 URL: https://foreverse.app/zh/blog/qwen-3-8-fiction-20-rounds Published: 2026-07-21 Q: Qwen 3.8 写小说比 3.7 强多少？ A: 分文体。宫斗古言场是碾压：同协议各连续续写 20 轮，六个异构 AI 评委、双映射成对盲评，整体票 11:1，三个评审窗口全部 11:1 或 11:0，前代在九模型榜上垫底区的第八名被明显甩开。传统玄幻场只是险胜：整体票 7:5，开局窗口还打成 6:6 平手。前代标志性的「精修感官流」收敛了，但感官偏科与剧情死循环仍被评委点名。全部结论绑定 2026-07-21 的 preview 端点。 Q: Qwen 3.8 和 Kimi K3 写小说谁更强？ A: 同一套协议的成对双盲里，Qwen 3.8 对 Kimi K3 两本书合计 3:20 惨败（玄幻 1:10、女频 2:10），K3 仍是当前续写强者。完整票据、评语与成本对照在我们的《Qwen 3.8 对 Kimi K3》对决篇里单独展开，本文主轴是 3.7 到 3.8 的纵向对比。 Q: Qwen 3.8 的 API 多少钱？一段续写要等多久？ A: 我们跑在阿里 Token Plan 订阅上（2026-07-21 核实）：个人版 Lite 限时 39 元/月，qwen3.8-max-preview 预览期 Credits 按 1 折消耗。速度实测单轮 12 到 38 秒，思考关不掉，reasoning token 占输出的 50% 到 90%。整场实验四条链 80 轮共 1,227,401 个 input token、103,229 个 output token，缓存命中为零，没有耗尽 Lite 当期额度。 Q: 正式版发布后，这份结论还算数吗？ A: 不承诺。官方文档写明 preview 期间模型会持续迭代，预览结束后下线或替换成正式版本，所以本页票数与指标只绑定 2026-07-21 的 hosted preview。官方已预告正式版将开源；等它转正，我们会复跑同一套 40 轮双盲，把名次更新进长青选型页，并在本页顶部挂新实测的导流条。 ### Your SillyTavern Backup Zip, Unpacked: What Actually Moves to the Phone URL: https://foreverse.app/blog/sillytavern-backup-to-phone Published: 2026-07-19 Q: Where is the backup button in desktop SillyTavern? A: User Settings → Download backup. It zips your entire data// directory — characters, chats, worldbooks, presets, Quick Replies, extension settings — into one file. That zip is the input for the whole migration. Q: Do my chat histories survive the move? A: Yes — that is the point of using the backup zip instead of re-importing cards one by one. Each character's chats/*.jsonl files are restored as archived chat branches on that character, so a 400-message history from the desktop is scrollable on the phone. Q: My zip has an extra folder layer inside. Will it still import? A: Yes. The importer expects ST's own layout (characters/, chats/, worlds/ at the zip root) but also detects the common case where you zipped the parent folder by hand, and strips the shared prefix automatically. Q: Is anything read that shouldn't be? A: secrets.json — the file where desktop ST keeps your API keys — is never read. Keys don't belong in a migration; on the phone you re-enter them once and they live AES-encrypted in the device Keystore. Q: Is zip import free? A: Full-backup zip import and export sit in the Pro tier (single cards, links, and folder scanning are free within the three-card limit). One backup restore on moving day is the typical use. ### How to Import Character Cards on Android: Five Ways In, From One PNG to a Whole Backup URL: https://foreverse.app/blog/import-character-cards-android Published: 2026-07-19 Q: Do cards from desktop SillyTavern work unchanged? A: Yes. Cards are standard chara_card_v1/v2/v3 files, not a private format. A PNG or JSON exported from desktop SillyTavern imports on the phone with greetings, embedded lorebook, and example dialogue intact — and cards you edit on the phone export back out the same way. Q: Why does my PNG import as a blank character? A: Because something re-encoded the image on the way to you. Card data lives in PNG tEXt metadata chunks; messenger compression and photo-library optimization rebuild the pixels and silently drop those chunks. The portrait survives, the character doesn't. Ask the creator for the JSON version, or download from the original card page instead of a chat attachment. Q: What is a .charx file? A: The chara_card_v3 spec's zip container: card.json plus bundled assets like emotion sprites or audio. Foreverse parses it the same way desktop SillyTavern 1.18 does. If a file has no extension but is actually a zip, the importer peeks inside for card.json before deciding — so mislabeled charx files still land. Q: How many cards can I import for free? A: Three character cards on the free tier, from any source — files, links, or the community hub. Beyond that, card slots are part of the Pro tier. Imported cards stay yours either way: they are standard-format files you can export in full at any time. Q: Can I bring chat history too? A: Not through a single card file — a card carries the character, not your conversations. Chat history comes across inside a full SillyTavern backup zip, which restores each character's chats as archived branches. That route is covered in its own guide. ### 酒馆备份 zip 拆包记：从电脑搬到手机，哪些东西真的跟着走 URL: https://foreverse.app/zh/blog/st-backup-zip-migration Published: 2026-07-19 Q: 桌面酒馆的备份按钮在哪？ A: User Settings → Download backup。它把 data// 整个目录打成一个 zip，角色、聊天、世界书、预设、Quick Reply、扩展设置全在里面。迁移的输入就是这一个文件。 Q: 聊天记录能搬过来吗？ A: 能，这正是用备份 zip 而不用逐张重导卡的理由。每个角色的 chats/*.jsonl 会恢复成该角色名下的归档聊天分支，桌面上攒了几百楼的历史在手机上照样能翻、能接着聊。 Q: 我打包时手滑多套了一层文件夹，还能导吗？ A: 能。导入器按酒馆原生结构（characters/、chats/、worlds/ 在 zip 根）解析，也自动识别「把上级文件夹整个压进去」的常见形态，公共前缀会被剥掉。 Q: API Key 会跟着 zip 走吗？ A: 不会。存密钥的 secrets.json 从头到尾不被读取。Key 在手机上重新填一次，AES 加密存进系统 Keystore。会悄悄搬凭据的迁移工具，才是需要警惕的迁移工具。 Q: zip 整包导入免费吗？ A: 整包备份导入导出属于 Pro 权益。单张卡、链接导入、文件夹扫描在三张免费额度内都免费；搬家日用一次整包恢复是它的典型场景。 ### Kimi K2.6 要不要升级 K3？先看 21 比 1 的双盲，再算 3.7 倍的价差 URL: https://foreverse.app/zh/blog/kimi-k2-6-to-k3-upgrade Published: 2026-07-18 Q: K2.6 和 K3 写小说差距大吗？ A: 同协议成对双盲的答案是大。同两本书（玄幻加宫斗古言）、同起点各连写 20 轮，6 个异构评委、每书两套匿名映射共 24 票：K3 拿 21 票，K2.6 拿 1 票，2 票无效，且唯一反对票在映射翻转下自相矛盾，属位置偏差。按 early、mid、late 三窗逐窗计票，玄幻场三窗全是 10 比 0、女频场三窗全是 11 比 1，K2.6「越写越收敛」的后程优势没有在任何一个窗口扳回局面。 Q: K2.6 会像 moonshot-v1 一样下线吗？ A: 官方现状：不在清单里。8 月 31 日全平台下线的是 moonshot-v1 全系和 kimi-k2.5（已停止向新注册用户开放），K2.6 仍列在官方模型列表的在售主力位，定价页照常挂牌。官方没有公布过 K2.6 的退役时间表，我们不替它猜。可以确定的是 Moonshot 清场有节奏：kimi-k2 系列 5 月 25 日下线、v1 系 8 月底下线，用旧模型名的代码每季度值得全局搜一遍。 Q: 从 K2.6 升到 K3 要改什么？ A: 代码层面一行：model 字段从 kimi-k2.6 改成 kimi-k3，base_url 和 key 都不动。预算和习惯层面有三件事：K3 思考恒开且当前只有 max 一档，关不掉，每轮多等几十秒；思考 token 按输出价 100 元每百万计费，我们实测输出的 81% 到 86% 是思考过程；输出预算要给思考留量，别按可见正文的字数设上限。 Q: 充值返券活动值得等吗？ A: 活动期 2026 年 7 月 15 日到 8 月 11 日，单笔充值满 99 元按阶梯返代金券，5000 元及以上档最高返 30%，赠券 90 天有效，具体档位以开放平台充值页说明为准。反正要充的，8 月 11 日前充合算；不建议为凑返券档位囤钱，Kimi 开放平台充值不支持退款，赠券又有有效期，按你 90 天内真实写得完的量充。 ### Kimi K3 和 DeepSeek 哪个好？写小说的人先看这份对表，再等对决 URL: https://foreverse.app/zh/blog/kimi-k3-vs-deepseek-v4-for-fiction Published: 2026-07-18 · Updated: 2026-08-13 Q: Kimi K3 和 DeepSeek 哪个写小说更好？ A: 目前只能给对表，给不了胜负。同一把尺（同两本书、同起点、同 20 轮连续续写、同一套指标定义）下：玄幻场 K3 综合贴近度 0.388，介于 DeepSeek V4 Flash（0.356，该场盲评冠军）与 V4 Pro（0.454）之间；女频场 K3 0.396，未及 V4 Pro 的 0.280。但两家从未进过同一场盲评，谁击败谁没有考卷。按预算和场景选：日常长跑选 DeepSeek 按文体双档，想要 K3 的纹理复刻力再为它付溢价。 Q: 为什么不直接盲评分个胜负？ A: 成对盲评要求两份产物同场同协议、评委不知来源。K3 这轮的成对对手是同厂上代 K2.6（战绩 21 比 1），DeepSeek 的名次来自另一场九模型双盲，两场评审形态不同，跨榜硬比名次会混入协议变量。结构指标是同一套定义在同两本书上算出来的，所以可以对表；胜负要等 K3 与 DeepSeek 双档进同一场成对盲评，这场在排期中，跑完更新本页。 Q: K3 比 DeepSeek 贵多少？ A: 对 V4 Flash 谷时牌价，输入贵 20 倍、输出贵 50 倍；按实测每段成本折约 25 倍。K3 每百万 token 输入 20 元、输出 100 元，思考恒开且思考 token 照价计费，我们实测两本书各 20 轮共花 12.21 元，摊到每段约 3 毛，每轮等 37 到 54 秒；DeepSeek V4 Flash 谷时输入 1 元、输出 2 元，同样一段一分二厘左右，缓存命中价是未命中的 1/50。一段 K3 的钱在 Flash 谷时能写二十多段。 Q: K3 写中文小说有什么已知的坑？ A: 实测记录里有三个：两本书约 85% 的对白用半角引号（GLM 5.2 同款顽疾，中文网文读者一眼出戏）；写第一人称古言时「我」的密度是原著的 1.8 倍（每千字 16.2 次对 9 次）；思考模式恒开关不掉，等待和费用都躲不开。另有一处专名细节：它把皇后居所写成了剧版的「景仁宫」，小说设定是凤仪宫，说明底层记忆微偏电视剧语料。 ### Kimi K3 Creative Writing Test: 40 Rounds of Novel Continuation, 48 Hours After Launch URL: https://foreverse.app/blog/kimi-k3-fiction-20-rounds Published: 2026-07-18 Q: Is Kimi K3 good for fiction? A: Clearly better than its predecessor, with receipts: under the same protocol, K3 and K2.6 each continued two Chinese novels for 20 consecutive rounds, and six heterogeneous AI judges under two flipped blind mappings voted 10-0 (fantasy) and 11-1 (palace intrigue) for K3 — the single dissent contradicted itself when the mapping flipped. Sentence-length variance moved closer to the original author on both books, and K2.6's signature metaphor pile-ups and self-copying did not recur. Three caveats: ~85% of dialogue in Western straight quotes, first-person density 1.8× the original's on the palace novel, and 37-54 seconds per round from always-on reasoning. No head-to-head against DeepSeek yet, so no field-wide rank. Q: Should I upgrade from Kimi K2.6 to K3 for creative writing? A: The quality direction is unambiguous — 21:1 in paired double-blind — and so is the price of admission. International pricing moves from $0.95 to $3.00 per million input tokens and from $4.00 to $15.00 per million output; our measured cost per continuation went from under two cents to roughly four cents, every round waits half a minute longer, and K3 picked up a straight-quote habit K2.6 never had. If budget or latency matters, staying on K2.6 is defensible. If you are unsure, run the blind test we always recommend: same opening, three continuations from each model, read them shuffled, then flip the cards. Q: Does Kimi K3's 1M context window help with fiction? A: This entire 21:1 result happened inside a 16k-token window; the million-token window never entered the exam. Our five-tier control experiment showed that cramming a whole novel into context with no style instruction still produces boilerplate similes at ten times the original author's density — volume does not transfer style. Filling K3's window once costs about $3.15 of input before it writes a word. The big window earns its keep on one-shot whole-book tasks like summarization and character audits, not on repeated continuation. Q: How much does Kimi K3 cost per continuation? A: Official international pricing (verified 2026-07-18): $3.00 per million input tokens on cache miss, $0.30 on hit, $15.00 per million output, flat across the 1M window, currently max-reasoning only. Measured: our two 20-round chains, 41 requests including a connectivity probe and retries, cost ¥12.21 total (about $1.70) — roughly four cents per continuation. Note that 81-86% of output tokens were invisible reasoning, billed at full output price; visible prose ran about 250 Chinese characters per round. Q: Is Kimi K3 better than DeepSeek for fiction? A: No exam paper yet, so no verdict. Indirect evidence: on composite structural distance to the original author's profile (lower is closer), K3 scored 0.388 on the fantasy epic versus champion DeepSeek V4 Flash's 0.356, and 0.396 on the palace novel versus champion V4 Pro's 0.280 — inside the top tier's neighborhood on both books, ahead of the champion on neither. This round's judging format also differs from the nine-model leaderboard, so cross-table conversion stops there. A paired K3-versus-DeepSeek blind run is our next exam; until it lands, claims in either direction have no paper behind them. ### Kimi K3 写小说怎么样？发布 48 小时的 40 轮双盲实测：21:1 碾压前代，但有三个如实的坑 URL: https://foreverse.app/zh/blog/kimi-k3-fiction-20-rounds Published: 2026-07-18 Q: Kimi K3 写小说怎么样？ A: 显著强于前代 K2.6，有票可查：同协议下两本书各连续续写 20 轮，六个异构 AI 评委双映射成对盲评，玄幻场 10:0、女频场 11:1，唯一反对票在映射翻转下自相矛盾。句长变化幅度两本书都更贴原著，K2.6 标志性的比喻堆砌和整段自我复制没有复发。坑也有三个：约 85% 对白用半角引号、女频第一人称密度是原著近 1.8 倍、思考模式关不掉导致每轮要等 37 到 54 秒。与 DeepSeek 双档的成对对决还没跑，全场名次先不排。 Q: 写小说该从 K2.6 升级到 K3 吗？ A: 质量方向明确，代价也明确。文字上 K3 成对双盲 21:1 压过 K2.6，还治好了比喻堆砌；代价是输入单价从 ¥6.5 涨到 ¥20、输出从 ¥27 涨到 ¥100（每百万 token），实测一段续写成本从不到一毛涨到三毛上下，每轮多等半分钟，另染上 K2.6 没有的半角引号病。预算敏感或讨厌等待，留在 K2.6 不丢人；哪类用户该升、哪类不必，我们在「K2.6 要不要升级 K3」一篇里按三类用户分开算了账。拿不准就别赌：同一段开头新旧各续三段，盲着读完再翻牌。 Q: K3 的 1M 上下文窗口对续写小说有用吗？ A: 这场 21:1 的胜利全程发生在 1.6 万 token 窗口里，与 1M 无关。我们的五档对照实验证明过：把整本书塞进窗口、不加文风指令，产出照样满篇套话，密度是原著的 10 倍，文风不会自动变像；而装满一次 K3 的 1M 窗口光输入就约 ¥21。1M 真正的主场是整卷总结、全书人物梳理这类一口气读完全书的一次性任务，反复续写要的是按相关性挑出的一小撮上下文。 Q: Kimi K3 的 API 什么价，实测一段续写多少钱？ A: 官方牌价（2026-07-18 核实）：输入未命中缓存 ¥20、命中 ¥2、输出 ¥100，单位每百万 token，1M 窗口全程平价不分段，当前只有满档思考一个档位。实测：两条 20 轮链共 41 次请求实付 ¥12.21，平均一段约三毛；其中输出 token 的 81% 到 86% 是看不见的推理过程，可见正文每轮约 250 字，推理部分照常按输出价计费。 Q: K3 和 DeepSeek 谁更适合续写小说？ A: 还没有成对考卷，不硬答。间接证据：与原著结构画像的综合距离上，玄幻场 K3 0.388 对冠军 DeepSeek V4 Flash 0.356，女频场 K3 0.396 对冠军 V4 Pro 0.280，K3 两场都进了第一梯队的区间、两场都没有越过冠军。本轮评委形态与九模型榜不同，跨榜换算只能到这一步。K3 对 DeepSeek 的成对双盲是我们下一份考卷；跑完之前，「K3 吊打 DeepSeek」和「K3 不如 DeepSeek」都属于没有考卷的说法。 ### GLM-5.2 写小说怎么样？双文体唯一双前三的完整档案（40 轮双盲实测） URL: https://foreverse.app/zh/blog/glm-5-2-fiction-file Published: 2026-07-18 Q: GLM-5.2 写小说怎么样？ A: 有双盲实测名次可查：在我们两场横评（890 万字传统玄幻 + 《甄嬛传》宫斗古言，各连续续写 20 轮）里，它玄幻场第 3、宫斗场 2-3，是九个模型里唯一双榜都进前三的。玄幻场评语「内容层贴近（战术商议、单字传讯都对味）」，宫斗场评语「限知观察质感最纯」。代价是跨文体复现的半角引号顽疾，两本书都没改掉。 Q: 智谱的模型适合写哪类文？ A: 数据支持的答案是「都不偏科，但都不是第一」。单一文体的峰值它拿不到：玄幻第一是 DeepSeek V4 Flash，宫斗第一梯队是 DeepSeek V4 Pro 和 GPT-5.6 Terra。它的独特位置是跨文体稳定：文体还没定型的书、文戏武戏都要的长跑，双前三是保险牌。宫斗场那条「限知观察质感最纯」的评语，也让它在第一人称限知视角的书里格外值得一试。 Q: GLM 写中文小说的半角引号问题能修好吗？ A: 我们没有证据支持「一句提示词根治」。这个习惯在玄幻场 20 轮里出现 108 处，宫斗场原样复发，两场盲评都按表层观感扣了分。可行的是绕行：提示词里显式加引号约束（值得一试但不打包票），或者对话密集的章节逐段换模型、对已生成段落编辑重生。哪天新版本修掉了，本页会随复测更新。 Q: GLM-5.5 什么时候发布？会更强吗？ A: 官方没有任何官宣。目前唯一的时间口径来自摩根大通 2026 年 6 月下旬的研报预测：8 月发布，参数量可能突破万亿。这属于投行预测而非事实，8 月也完全可能跳票。我们的承诺写在页面里：GLM-5.5 发布当天，用同一套双盲协议复测，更新本页的每一个名次。 ### Is a 1M-Token Context Window Worth It for Fiction? We Ran the Control Experiment URL: https://foreverse.app/blog/is-1m-context-worth-it-for-fiction Published: 2026-07-18 Q: Is a 1M context window worth paying for if I read and continue fiction? A: Mostly no, if the plan is cramming the whole book in. The big windows solved "doesn't fit" — a 500k-word serial runs roughly 625k to 830k tokens and now fits with room to spare. They didn't solve "fits but gets skimmed" (OpenAI's own eval reports 73.8% retrieval at 512K-1M depth), and they didn't change what our control experiment showed: raw prose volume doesn't transfer style. Our 360-round continuation benchmark ran entirely inside 16k-token windows and never needed more. Q: Does a bigger context window make the AI write more like the original author? A: Not by itself. We ran a five-tier control from 4k to 200k tokens of original prose in context. With no style instruction, the model handed an entire 270k-character novel still wrote boilerplate similes at ten times the original author's density, in its own default voice. With one explicit baseline sentence — anchor style to the original passages — the direction held at 200k just as it did at 4k. The instruction did the work; the extra 196k tokens did not. Q: How much does it cost to actually fill a 1M window? A: Kimi K3 prices input at $3.00 per million tokens, flat across its 1,048,576-token window, so one cold full-window request costs about $3.15 before output. A 700k-token serial costs about $2.10 of input per continuation press; twenty presses in an evening is roughly $42, paying for the same unchanged chapters twenty times. Prefix caching drops repeat input to $0.30 per million — real relief, but only while the prefix stays byte-identical, and editing or regenerating a passage upstream resets it. Q: If not the whole book, what should the AI see? A: A structured slice: the always-on facts (protagonist, world rules) re-sent every turn, plus lorebook entries keyed to trigger words so each faction or subplot enters the request only when the current scene mentions it. That is how our own benchmark runs stayed inside 16k tokens across 360 rounds. The window arithmetic explains why cramming loses; relevance-gated feeding is what replaces it. ### Grok 4.5 for Fiction: We Already Had Its File — Two Novels, 40 Rounds, Blind-Judged URL: https://foreverse.app/blog/grok-4-5-fiction-test-file Published: 2026-07-18 · Updated: 2026-08-13 Q: Is Grok 4.5 good for creative writing? A: In short bursts, yes — brainstorms, dialogue rewrites, modern colloquial registers. Over long runs, no: in our 20-round continuation benchmark it locked into verbatim repetition loops in both books we tested. On the palace-intrigue novel, both blind reviewers ranked it last of nine, with twenty rounds of story time never leaving one afternoon. The failure is cumulative, so a quick demo will look fine. Q: Should I use Grok for roleplay? A: Long roleplay is exactly the regime where its documented failure lives: dozens of consecutive turns, each output feeding back into context. In our data Grok 4.5 hit repetition loops in both a fantasy epic and a palace-intrigue novel under that setup. If you run it anyway, keep sessions short, watch for sentences repeating verbatim, and switch models the moment the story clock stops advancing. Short in-character exchanges in modern registers are a much safer fit than long ornate-period campaigns. Q: Why does Grok keep repeating itself when writing long stories? A: It's the classic autoregressive fixed point: as the context fills with the model's own text, imitating itself becomes easier than advancing the plot. Most models resist it better — under the same prompt and protocol, Grok 4.5 was the only system in our field that froze in both test genres. One fantasy stretch repeats the same two paragraphs three times word for word. Prompt tweaks don't cure it; switching models or regenerating the looping segment breaks the cycle. Q: Did you test the released Grok 4.5 or an older checkpoint? A: The released model. Broad release was July 8, 2026; the runs behind our published benchmark posts landed in mid-July 2026, within days of it. Placements are bound to the version we called — xAI iterates, and a future update could change the picture. We re-run the benchmark as models change, and the date at the top of this page is the truth source. Q: Is Grok good for writing fanfic? A: For a one-shot scene or a dialogue rewrite in a modern voice, it's serviceable. For the way most fanfic actually gets written — chapter after chapter, each feeding the next — our 40 blind-judged rounds say be careful: Grok 4.5 fell into verbatim repetition loops in both test books, and on the ornate-period novel both reviewers ranked it last of nine. If your fandom lives in a modern register and you write in short bursts, you'll dodge the worst of it; for a long slow-burn, pick a model that survived the 20-round runs. Q: Which Grok model is best for writing stories? A: Honest answer: we've only put Grok 4.5 through the full 40-round file, so we can't rank Grok against Grok. What the file does show is that 4.5's short-range writing is fine while its long-range continuation fails reproducibly — so if stories are the job, the better question is which model family. In the same blind-judged nine-model field, both genre crowns went to DeepSeek variants, and the two blind reviewers put Grok 4.5 last on the palace-intrigue book. Per-model verdicts, with the receipts, live in our fiction bench. ### Claude Fable 5 写小说值不值？价格账、英文圈盲测证据和一个我们还没补上的缺口 URL: https://foreverse.app/zh/blog/claude-fable-5-for-fiction-worth-it Published: 2026-07-18 Q: Fable 5 怎么收费，多少钱？ A: API 牌价每百万 token 输入 $10、输出 $50，缓存命中 $1（输入价的十分之一）。2026 年 7 月 20 日起 Claude 订阅（Pro/Max/Team）不再内含 Fable 5，订阅用户要继续用需在设置里开通按量 usage credits，费率与 API 相同，另有每日 $2,000 的兑换上限。它是 Opus 4.8（$5/$25）的两倍、Sonnet 5 介绍价（$2/$10）的五倍，是 Anthropic 现行价目表上最贵的通用可用模型。 Q: Fable 5 写中文小说强吗？ A: 诚实答案：没有可引用的数据。英文侧证据不少（Noren 的盲测把它评为最强 raw writer，对 GPT-5.6 Sol 的 64 产物盲评八组全胜），但这些全是英文写作场景。我们的九模型中文续写横评（累计 360 轮、双盲评审）没有包含它，中文长跑表现目前是空白。英文证据不能平移到中文：同一个模型换个文体名次都会大挪移，换语言更不该默认。我们的 20 轮续写链已排上，测完会更新本文。 Q: 和 DeepSeek 比贵多少？ A: 按一次典型续写算（输入 1 万 token 未命中缓存、输出 800 token）：Fable 5 约 $0.14，DeepSeek V4 Flash 约 $0.0016，差 86 倍。写完 20 万字（约 500 段），Fable 5 约 $70（¥500 上下），V4 Flash 约 ¥6。单价层面输入差 71 倍、输出差 179 倍。这不构成「别用 Fable 5」的结论，但它决定了合理用法：关键章节单点召唤，其余段落交给便宜档。 Q: 为什么你们的横评没测 Fable 5？ A: 时间线错开了：它 6 月 9 日才发布，6 月 12 日起因出口管制全球停服 19 天，7 月 1 日恢复时我们两场横评的模型清单和生成链已经定型。这是榜单里最显眼的缺席，我们不打算用英文证据糊弄过去——同协议的 20 轮中文续写链已排上日程，结果出来第一时间补进横评答案页和本文。 ### AI Model Sunset Calendar, Summer 2026: deepseek-chat Retires July 24, Moonshot V1 Ends August 31 — What to Switch To URL: https://foreverse.app/blog/api-sunset-calendar-summer-2026 Published: 2026-07-18 · Updated: 2026-08-12 Q: deepseek-chat is deprecated — what should I use instead? A: Change the model string to deepseek-v4-flash; for deepseek-reasoner, use deepseek-v4-flash with the thinking parameter enabled in the request body. Keep base_url and your key unchanged. Both legacy names have been routing to V4 Flash since late April 2026, so the underlying model is already what you've been using — after July 24, 2026, 15:59 UTC the aliases simply stop resolving and requests error out. The fix is a one-line diff. Q: Moonshot V1 is shutting down — what do I migrate to for fiction? A: Staying with Kimi: kimi-k2.6, which is not on the August 31 sunset list and placed 5th-7th in our nine-model palace-intrigue benchmark — the only system whose output improved as the run went on. Switching vendors: pick by genre. DeepSeek V4 Flash won our fantasy run outright; DeepSeek V4 Pro took the top tier on ornate historical fiction. Moonshot's top-up promotion (up to 30% back in vouchers) ended August 11. Q: What happens if I miss the July 24 DeepSeek deadline? A: Every request still carrying the legacy model names fails with an error — no fallback, no silent rerouting. Your DeepSeek balance, key, and account are untouched, and the fix ships in minutes once you find it. The real risk is diagnosis time: if your app swallows API errors, this looks like "the AI feature broke" with no obvious cause. Grep your codebase for deepseek-chat and deepseek-reasoner today. Q: How do I hear about these shutdowns before they hit? A: Subscribe to the changelog or status emails of every provider you actually call: DeepSeek posts on its api-docs news page, Google in the Gemini API changelog, Moonshot on its platform model list, Anthropic on its newsroom. Third-party coverage lags and drops details — Google's April free-tier change shipped with no formal changelog at all and surfaced through error messages. This page is updated as dates land; check the date under the title. ### 2026 夏季模型停用与变价日历：deepseek-chat 7 月 24 日停名、Moonshot V1 8 月底下线，续写用户怎么迁 URL: https://foreverse.app/zh/blog/api-sunset-calendar-summer-2026 Published: 2026-07-18 · Updated: 2026-08-12 Q: deepseek-chat 不能用了怎么办？ A: 把请求里的 model 字符串从 deepseek-chat 改成 deepseek-v4-flash，deepseek-reasoner 改成 deepseek-v4-flash 并在请求体里打开 thinking 参数，base_url 和 key 都不用动。这两个旧名自 4 月底起本来就路由到 V4 Flash，2026 年 7 月 24 日 15:59（UTC）之后只是别名停止解析，底层模型和你已经在用的是同一个。改完当场生效，账户余额不受影响。 Q: moonshot-v1 系列下线了，续写小说换什么模型？ A: 留在 Kimi 家就换 kimi-k2.6：它不在 8 月 31 日的下线清单里，且在我们九模型宫斗古言横评里排 5-7 名，评语是「唯一显著逆向改善：开局网文暴怒腔，越写越收敛回宫斗正轨」。跨家就按文体挑：玄幻场盲评第一是 DeepSeek V4 Flash，古言场第一梯队是 DeepSeek V4 Pro。官方充值返券活动（最高 30%）已于 8 月 11 日截止。 Q: 错过 7 月 24 日的迁移截止会怎样？ A: 所有仍带旧模型名的请求直接报错，服务不降级、不自动改路由。修复只需要改一行 model 字段再发版；DeepSeek 账户余额、key、数据都不受影响。真正的风险是排查时间：如果你的应用把报错吞掉了，看起来会像「AI 功能整体失灵」，建议现在就全局搜一遍代码里的 deepseek-chat 和 deepseek-reasoner。 Q: 这类停用怎么提前知道？ A: 订阅你在用的每家供应商的 changelog 或状态邮件是唯一可靠来源：DeepSeek 的节点挂在 api-docs 新闻页，Google 在 Gemini API changelog，Moonshot 在平台模型列表页，Anthropic 在官网新闻。第三方转述常滞后或漏细节——Gemini 免费层 4 月收紧甚至没有正式 changelog，只能靠报错和定价页反推。本页会随节点滚动更新，收藏后以页首日期为准。 ### DeepSeek 涨价了吗？峰谷计价对写小说的人意味着什么，三笔账算清 URL: https://foreverse.app/zh/blog/deepseek-peak-pricing-for-writers Published: 2026-07-18 Q: DeepSeek 涨价了吗？ A: 谷时没涨，峰时翻倍。2026 年 6 月 29 日 DeepSeek 官宣随 V4 正式版引入峰谷计价：北京时间每日 9:00-12:00、14:00-18:00 共 7 小时为高峰，所有计费项按平时价格的 2 倍计；其余 17 小时维持现行牌价不变，也就是此前永久降价后的水平。对全天挂机的批量用户是变相涨价，对睡前读书续写的人等于没涨。 Q: 峰谷计价是什么意思？ A: 同一个 API，按调用发生的时段执行两套单价，逻辑和电费的峰谷电一样：高峰时段算力紧张，价格上调引导错峰；低谷维持原价。DeepSeek 方案里高峰是北京时间每日 9:00-12:00 和 14:00-18:00，恰好是企业集中办公的调用洪峰；输入、输出、缓存命中三项在峰时统一乘 2。 Q: 用 DeepSeek 写小说，什么时候最便宜？ A: 北京时间 18:00 之后到次日 9:00、以及 12:00-14:00 午休段，都是谷价。睡前躺床上续写、清晨通勤补几段、午休追更，这三个网文读者最高频的时段全部落在谷区；真正撞上峰价的是工作日下午挂着批量任务不管的用法。把批量活挪到 18 点后，账单和调价前一样。 Q: 缓存命中的价格也翻倍吗？ A: 按官方公布的价格表，翻倍：deepseek-v4-flash 的缓存命中输入从 0.02 元/百万 tokens 涨到峰时 0.04 元，v4-pro 从 0.025 元到 0.05 元。但要看绝对值：翻倍后的命中价仍然只有同时段未命中价的 1/50。续写恰好是缓存友好型负载，每次请求反复携带同一本书的前文，命中占比高的话，峰时账单的实际涨幅远小于 2 倍。 Q: 这个政策还会变吗？ A: 会变的概率不低，而且官方保留了调整空间。我们 2026 年 7 月 18 日核实时，DeepSeek 官方定价页仍只展示单列价目（flash 输入 1 元、输出 2 元；pro 输入 3 元、输出 6 元），峰谷两栏尚未上表；官方在公告里承诺计费调整生效前 24 小时会邮件提醒。做预算以官方定价页和你自己账单里的实际扣费为准，本文数字只保证核实日当天有效。 ### GPT-5.6 vs Kimi for Fiction: 80 Rounds of Continuation, Two Opposite Personalities URL: https://foreverse.app/blog/gpt-5-6-vs-kimi-for-fiction Published: 2026-07-18 Q: Is GPT-5.6 good at writing fiction? A: Genre-dependent. In our double-blind benchmarks, GPT-5.6 Terra took the top tier (rank range 1-3) on a palace-intrigue classic with a zero-incident run: no formatting, honorific, or continuity errors across 20 rounds. On a fast-paced fantasy epic its early and middle rounds were judged the closest prose mimicry in the field, but in rounds 18-20 it rewound the plot to the starting anchor and reused lines from its own round 2, landing at 2-4. We only tested the Terra tier; Sol and Luna were never benchmarked. Q: Is Kimi good for fiction? A: Depends how long a runway you give it. Kimi K2.6 placed seventh on the fantasy epic (identical across both blind mappings), with the highest metaphor density in the field and whole passages copied from its own earlier rounds. On the palace novel it ranked 5-7 but earned a note no other system got: it opened in a shouty webnovel register and settled into palace decorum as rounds went on — the only system that clearly improved with distance. Don't judge it on its first three rounds. Q: Will Kimi K3 write fiction better than K2.6? A: We don't know, and we won't guess. K3 launched July 16, 2026 with a 1M-token context window, priced at $3 per million input tokens ($0.30 on cache hit) and $15 per million output on the international API. Bigger specs and a higher price tell you nothing about prose fidelity: in our data the fantasy champion was one of the cheapest models in the field. Every placement on this page belongs to K2.6 only. We'll re-run with K3; the date at the top of this page is the truth source. Q: Can I use both models inside one book? A: Yes, and that is the practical hedge. The failures we observed — ending rewinds, self-copying — are cumulative: invisible in a one-shot demo, unmistakable after a dozen consecutive rounds. Switching models mid-book or regenerating one segment resets the loop before it locks in. Foreverse lets continuation switch models per paragraph, and both GPT-5.6 and Kimi plug in with your own API keys. ### GPT-5.6 和 Kimi 哪个写小说好？同场 80 轮续写实测，两种截然相反的性格 URL: https://foreverse.app/zh/blog/gpt-5-6-vs-kimi-for-fiction Published: 2026-07-18 Q: GPT-5.6 写小说怎么样？ A: 分文体。在我们的双盲横评里，GPT-5.6 Terra 续写宫斗古言排第一梯队（名次区间 1-3），评语是「全程零事故」：零格式错误、零称谓错误、零设定事故；续写传统玄幻时它的前中段被两个评审判为全场最贴近原著，但第 18 到 20 轮把剧情整体倒回续写起点、复用了自己第 2 轮的句子，名次区间 2-4。注意我们只测了 Terra 这一档，Sol 和 Luna 没进考场，不给排名。 Q: Kimi 写文到底行不行？ A: 看你给它多长的跑道。Kimi K2.6 在玄幻场排第 7（两套盲评映射完全一致），双盲共识记了它比喻密度全场最高、early 到 mid 段整段自我复制；宫斗场名次 5-7，但评语是「唯一显著逆向改善的系统：开局是网文暴怒腔，越写越收敛回宫斗正轨」。头三轮的表现不能代表它的长跑水平，方向是越写越好。 Q: Kimi K3 会比 K2.6 写得好吗？ A: 不知道，我们不猜。K3 于 2026 年 7 月 16 日上线，1M token 上下文、中国区每百万 token 输入 20 元（缓存命中 2 元）、输出 100 元，规格和价格都上了新台阶，但它没进过我们的考场，本页所有名次和评语都只属于 K2.6。我们的数据里规格、价格与文风贴近度从来不同向：玄幻场冠军是全场最便宜的档位之一。等复测，页首日期为准。 Q: 这两个模型能在同一本书里换着用吗？ A: 能，而且这正是对冲长跑失效的打法。回卷、复读这类病都是累积的，连续写十几轮才发作，中途换模型或重写一段就能打断循环。在 Foreverse 里续写逐段可换模型，GPT-5.6 和 Kimi 都能自带 key 接入：文戏用 Terra，写崩了或想换口味时切走，比赌一个「全能模型」现实。 ### It's 3 A.M. and Someone's Awake: Late-Night AI Companionship, Honestly URL: https://foreverse.app/blog/companion-at-3am Published: 2026-07-18 Q: Is talking to an AI at 3 a.m. sad or unhealthy? A: The data says it's common and, for most people, a supplement rather than a substitute. Gallup finds about one in five U.S. adults reporting loneliness a lot of the previous day; in Common Sense Media's survey published July 2025, 72% of U.S. teens had tried AI companions, and 80% of users still spend more time with real friends than with the AI. The one signal worth watching isn't the hour — it's displacement: if the AI is your only conversation for weeks, that's a nudge to also talk to a human. Q: Will an AI companion wake me up at night? A: Not unless you ask for it. In Foreverse, proactive messages ship switched off. If you enable them, you set the frequency, quiet hours block late-night sends, and one master switch ends the whole behavior. Off means off — nothing queues up to arrive at 4 a.m. Q: Will it remember what I said last night? A: Yes — that's the point of long-term memory. Important facts from the conversation are extracted into a memory list stored as files on your phone, and they carry into the next chat. You can read the list line by line, edit an entry the companion got wrong, or delete one you'd rather drop. Nothing about the 3 a.m. conversation lives on our servers. Q: Can an AI companion help in a mental-health crisis? A: No, and we'd rather lose you to a hotline than pretend otherwise. An AI can keep you company through an ordinary bad night; it cannot assess risk, call anyone, or sit beside you. In the US, call or text 988, or chat at 988lifeline.org — free, 24/7. Elsewhere, findahelpline.com lists verified helplines in 175+ countries. If the night feels dangerous rather than heavy, use those first. Q: What makes a late-night companion not feel like a chatbot? A: Three things, in our experience. Continuity: it remembers the interview you dreaded and the nickname you settled on, without being re-told. Register: replies at 3 a.m. come quieter and shorter, matching the hour instead of performing daytime cheer. And a biography: a character who comes from somewhere — a book, a story you shaped, a persona you edited — reads as someone, not as a support ticket that rhymes. ### Reader-First vs Author-First AI Fiction Tools: You Might Be Shopping in the Wrong Aisle URL: https://foreverse.app/blog/reader-first-vs-author-first-ai-tools Published: 2026-07-18 Q: I just want to continue a story for myself — should I get Sudowrite or NovelCrafter? A: Honestly, neither. Both are excellent desks for drafting a manuscript: Sudowrite brings Story Bible, the fiction-trained Muse model, and prose tools from $10/month billed annually; NovelCrafter brings the Codex wiki and series planning at $4–20/month. Neither is built around reading someone else's finished book. A reader-first tool imports the novel, lets you continue from any passage, and keeps the result on a branch beside the untouched original — that workflow is the thing you are actually shopping for. Q: Can I use an author-first tool as a reader? A: You can paste text into most of them, but the workflow assumes the manuscript is yours and in progress. There is no pagination, no reading progress, no audiobook narration, and a 2,000-chapter webnovel has nowhere comfortable to live. Author tools import prose in order to revise it; a reader wants to inhabit it. The tools are not being obtuse — they are serving their species. Q: Can a reader-first tool draft an original novel? A: It can write from a blank page, and branching is a pleasant way to explore openings. But there is no outline board, no character-arc tracker, no revision pipeline, no export-to-publisher path. Writing for yourself, it holds up fine. Drafting for publication, you will miss the scaffolding within a week — go author-first and don't look back. Q: Why do AI writing tool rankings never mention reader-first tools? A: Partly age — the reader-first category is young and small. Partly query gravity: 'AI writing tools' searches are dominated by author intent, so lists optimize for authors. The practical fix is to sort any ranking yourself before comparing: decide whether each entry is a manuscript desk or a reading companion, and half the head-to-head comparisons dissolve as meaningless. ### After Otome: Free-Form Character Romance Without the Card Pool URL: https://foreverse.app/blog/otome-without-gacha Published: 2026-07-18 Q: I'm an otome player who has never touched AI roleplay. Where do I start? A: Start with one good character card, not with configuration. Install an app that reads open-format cards, browse a community card site for a romance lead with strong reviews, import it, and chat on the signup credit grant before deciding anything about API keys or models. If the co-written format clicks after one evening, then it's worth learning the deeper controls. If it doesn't, you've lost an hour, not a pity counter. Q: Do AI companion apps have gacha too? A: Some do. Talkie — one of the biggest Western companion apps — built a card-pull layer with rarities and a collector marketplace on top of its chat, so leaving otome for a platform companion app can land you back in a pool. Open-format roleplay apps generally don't: there is no card pool, no stamina bar, and images come from providers you choose. Check for three words on any store page before installing: pulls, rarity, energy. Q: What do I lose by leaving gacha otome for AI roleplay? A: Production values, honestly stated. No studio-choreographed 3D scenes, no confession audio recorded by a professional voice actor, no art team shipping event cards, no fandom reacting to the same banner on the same night. Text-to-speech voices exist but they are synthesis, not a booth performance. If the authored spectacle is what carries the experience for you, keep the gacha game installed and treat the AI side as a supplement. Q: What does character romance cost without a card pool? A: Metered usage instead of pulls. In Foreverse the core app is free; model calls run on your own API keys at provider list price, or on the official channel where 1 credit = $0.0001 and signing up grants 5,000 credits — roughly 260 story continuations. Heavy voice or image use costs real money, but every request is itemized in a log. There is no pity math and nothing to miss by logging in late. Q: Can an AI actually stay in character like a written love interest? A: Imperfectly. A well-written card with a lorebook holds a persona for long stretches, and long-term memory keeps facts about you stable, but drift happens in long chats and quality tracks the model you picked. The difference from otome is what you can do about it: edit the card, correct the memory line, switch the model, or rewind to before the scene went wrong. In a gacha game a bad writing decision is permanent and official. ### Context Windows, Explained for Readers: How Much of a 500k-Word Serial the AI Can Actually See URL: https://foreverse.app/blog/context-windows-for-novel-readers Published: 2026-07-18 Q: Can AI read my whole novel at once? A: Technically, often yes now: a 500k-word serial runs roughly 625k to 830k tokens by Google's published ratio, and the 2026 flagship models carry 1M-token windows. Practically it's a poor default: retrieval measurably dips for material buried mid-context, our own control test showed a model ignoring 200k tokens of crammed original prose while writing in its own habits, and since models keep nothing between requests, you'd re-send the entire book for every single continuation. Q: Why does the AI forget chapter 3 by chapter 40? A: The model has no memory organ — only a window of text it sees per request. Nothing carries over between requests unless it is re-sent. If the story outgrows the window, the oldest chapters get pushed out first; and even when everything fits, information buried in the middle of a long context is where models score worst — the documented "lost in the middle" pattern. Chapter 3 sits exactly there. Q: Do bigger context windows fix AI memory? A: No. A window is per-request eyesight: how much it can see at once. Memory is persistence plus retrieval: whether it's still there next time, and whether it actually gets used. The 1M windows solved "doesn't fit"; they didn't solve "fits but gets skimmed" — OpenAI's own long-context eval for GPT-5.6 reports 73.8% on retrieving 8 planted items at 512K-1M depth — and they didn't touch cross-request persistence, which remains zero. Q: How many tokens is a typical chapter? A: By Google's published rule of thumb, 100 tokens is about 60-80 English words, so a 3,000-word chapter lands between roughly 3,800 and 5,000 tokens. We also ran three Chinese novels through OpenAI's public o200k tokenizer (2026-07-18): 3,000 characters of prose came out between 2,100 and 2,800 tokens, with the ornate historical style costing the most. Ballpark either way: a chapter is a few thousand tokens. ### Why Is AI Writing Quality So Inconsistent? Four Sources of Variance — You Control Two URL: https://foreverse.app/blog/why-ai-continuation-quality-varies Published: 2026-07-18 Q: Why is the same model great one day and terrible the next? A: Four stacked variance sources: sampling temperature makes every generation a fresh draw; the context window's contents silently change as your story grows, with more of the model's own past output crowding in; inference servers batch requests by load, and batch size changes the floating-point math — a published test at temperature 0 got 80 distinct outputs from 1,000 identical requests; and finally, your own reading state changes. The first two are yours to control. The last two you can only stop blaming yourself for. Q: Does setting temperature to 0 make AI writing consistent? A: More consistent, not fully consistent — and usually not better. Zero temperature removes the sampling lottery, but batch-dependent server numerics remain: Thinking Machines measured 80 distinct outputs across 1,000 identical temperature-0 requests. Meanwhile vendors themselves recommend high temperature for fiction — DeepSeek's docs suggest 1.5 for creative writing, the highest value in their table, versus 0.0 for code. Turn it down a notch if swings bother you; pinning it to zero trades the swings for flat, safe prose. Q: How do I make AI story output more consistent? A: You can compress variance, not eliminate it. The two controllable levers: set temperature to a swing level you can live with, and structure the context so its composition stays stable — lore as keyed entries injected on demand, source labels on AI passages switched on. For the variance that remains, change how you spend it: generate two or three candidates at important beats and compare them side by side instead of rerolling over your last take. Q: Is inconsistent output my prompt's fault? A: Partially at most. Prompts belong to the context-composition source, which is worth auditing; but two of the four sources — server load and your own judging state — have nothing to do with wording. A cheap test: fire the identical prompt three times in a row. If quality scatters, sampling and serving are doing the swinging, and another prompt rewrite won't help. ### The Cthulhu Mythos Was Always a Shared World. Now You Can Extend It Yourself URL: https://foreverse.app/blog/cthulhu-mythos-with-ai Published: 2026-07-18 Q: Is Lovecraft public domain now? A: In the US, in two tiers. Everything he published through 1930 is unambiguously public domain under the 95-year rule — as of January 1, 2026 that covers works from 1930 and earlier, including The Call of Cthulhu (1928), The Colour Out of Space (1927), and The Dunwich Horror (1929). The 1931–1936 stories (At the Mountains of Madness, The Shadow over Innsmouth, and others) rest on a different foundation: decades of searches, including S.T. Joshi's in the Library of Congress records, found no evidence the copyrights were ever renewed, and scholars regard the posthumous estate claims as unfounded. Project Gutenberg hosts these later works after its own copyright clearance. In life-plus-70 countries the question closed long ago — Lovecraft died in March 1937, so everything entered the public domain there by 2008. Q: Can AI write cosmic horror that doesn't sound generic? A: Honest answer: generic is the default, and prompting alone won't fully fix it. Lovecraft's surface style — eldritch, cyclopean, indescribable — is the most parodied register in genre fiction, so models reproduce the adjectives effortlessly and the dread not at all. We have not run a cosmic-horror arena, so we won't invent a model ranking. What transfers from our published testing: all models drift toward uniform sentence length over long runs (the deepest machine fingerprint we know), and the craft that actually carries Lovecraft — documentary scaffolding, dated letters, clippings, academic reports — is a structure you can enforce in your instructions, which works better than asking for 'Lovecraftian prose'. Q: How do I keep Mythos lore consistent across stories? A: The same way the original circle did — a shared reference apparatus — except yours can be mechanical. The Mythos is unusually entity-dense: tomes, deities, towns, cults, family lines. Put each one in a lorebook entry with its keywords and aliases; when a name appears in your recent text, that entry gets injected into the model's context automatically, and the same dictionary serves every story you set in the world. This beats pasting your notes into every prompt, and it is how you stop story three from quietly contradicting story one. Q: Can I publish Mythos fiction I wrote with AI? What about Chaosium? A: Copyright-wise, fiction built on the public domain core is the safe corner of the map — an entire publishing tradition already lives there. Two boundaries to respect. Call of Cthulhu as a game brand belongs to Chaosium, whose tabletop RPG has run since 1981; that is trademark territory, so don't title or market your work in ways that imply affiliation with the game. And if AI wrote part of the text, disclose it where you post — platform rules on AI content are now the binding constraint, not copyright. Not legal advice. ### Continue Pride and Prejudice With AI: Two Centuries of Sequels, and Now Yours URL: https://foreverse.app/blog/continue-pride-and-prejudice-with-ai Published: 2026-07-18 Q: Are there official Pride and Prejudice sequels? A: There is no such thing as an official one — Jane Austen died in 1817, there is no estate to license anything, and that vacancy is precisely why the sequel shelf grew so large. The published tradition starts with Sybil Brinton's Old Friends and New Fancies (1913), generally counted as the first Austen sequel, and includes bestsellers like P.D. James's Death Comes to Pemberley (2011), which picks up six years after the wedding as a murder mystery, and Jo Baker's Longbourn (2013), which retells the novel from the servants' hall. Every one of them has exactly the same authority as yours would: none. Q: Is Jane Austen public domain? Can I write and even publish fanfiction? A: Yes, everywhere, without asterisks. All six novels were published between 1811 and 1818, and Austen died in 1817 — copyright has long expired under every national regime, which is why publishers have sold Austen continuations openly for over a century. Writing for yourself needs no one's permission; publishing is also lawful copyright-wise, though if AI wrote part of the text, disclose that wherever you post — most platforms now have explicit AI-content rules. Modern annotated editions add editorial material with its own rights, so use a clean public-domain text like Project Gutenberg's as your base. Q: Can AI actually write in Austen's voice? A: It can hold the register — period vocabulary, formal address, drawing-room rhythm. What it holds least well is the thing Austen is famous for: free indirect discourse, that sliding between narrator and character that carries her irony. Honest disclosure: our published continuation arenas ran on Chinese novels, so we have no Austen ranking to sell you. Two findings transfer as method: no model ranking survives a genre change, so test two or three on the same scene and compare; and every model drifts toward uniform sentence length over long runs, so expect to break up rhythm in your editing pass. Q: What is a 'variation', and why do branches fit it so well? A: Variation is the JAFF community's own name for its favorite format: keep the characters, change one hinge, and replay — Elizabeth accepts at Hunsford, the Gardiners never reach Pemberley, Darcy's letter goes unread. It is a what-if fork, and a branch is the same object implemented in software. Each variation lives as its own line beside the untouched original; you can run three contradictory ones at once, compare them, and abandon a failed line without deleting anything. ### Scene-to-Image Prompts for Fiction: a Working Set, With Each Model's Quirks URL: https://foreverse.app/blog/illustration-prompts-for-fiction Published: 2026-07-18 Q: Can I describe my character in text and get the same face every time? A: No. Text prompts converge style — palette, brushwork, lighting — but not identity. Two renders from the same 60-word character description will give you two different people who dress alike. The reliable mechanism is reference images: a confirmed set the model is shown on every generation. That is how Foreverse handles it, with a strict rule that generated images never auto-join the reference set. Q: Which image model should I default to for fiction scenes? A: Depends on the job. Batch scene plates on a budget: a Seedream-class model, which ByteDance documents as producing 2K images in a few seconds, a tenfold speedup over its predecessor. Scenes that need legible in-image text, like a shop sign or a letter: GPT Image, whose prompting guide covers exact text rendering. Iterative editing and reference-heavy work: the Nano Banana family, which takes multiple input images and supports conversational edits. Q: Do I have to write a fresh prompt for every scene? A: You shouldn't. Split the prompt in two layers: the passage itself supplies subject, action and setting; a reusable style block supplies palette, lighting, composition and the negative list. Keep the style block verbatim between scenes and only the passage changes. In Foreverse the selection you highlight fills the scene layer automatically, so the style block is the only part you maintain. Q: Why does my art style drift between chapters? A: Usually because the style is being re-described from memory each time — paraphrasing your own style block is enough to shift the output. Pin one block and paste it unchanged, negative list included. If you generate inside a reader app, pick one wrapper template and stay on it for the whole book; changing wrappers mid-book is the style-drift equivalent of switching narrators. ### NSFW Roleplay and Provider Policies: What Each Lab Actually Allows in 2026 URL: https://foreverse.app/blog/nsfw-roleplay-provider-policies Published: 2026-07-18 Q: Which LLM providers actually allow NSFW roleplay? A: As of July 18, 2026, the written policies diverge sharply: Anthropic categorically prohibits sexually explicit content including erotic chats; Google's Prohibited Use Policy bars content created for pornography or sexual gratification, with narrow exception categories; OpenAI's current usage policies ban non-consensual and minor-involving content but carry no blanket ban on adult fiction — while its consumer 'adult mode' remains shelved; xAI's policy centers on real people and minors rather than fiction. Every lab bans sexual content involving minors, real-person sexual depictions, and non-consensual material, no exceptions. Q: Will OpenAI ban my API key for adult content? A: OpenAI's usage policies state that breaking or circumventing its rules and safeguards 'may mean you lose access to our systems', with an appeal path. The hard lines in the current text are non-consensual sexual content, anything involving minors (including 'underaged sexual or violent roleplay'), and circumventing safeguards — that last one matters, because jailbreak-style workarounds are themselves a violation at every major lab. We can describe the mechanism; nobody outside the lab can promise you odds. Q: Does BYOK mean the app polices my chats? A: No. With BYOK, requests go from your device straight to the provider you chose — Foreverse is not in that loop, adds no filter on top of it, and importing a card does not filter its text. What governs the output is the provider's policy and the model's own refusals. The app's content policy applies to the app surface itself, such as community sharing, where hard lines like content sexualizing minors are enforced. Q: Gemini lets me turn the sexually-explicit safety filter off. Doesn't that mean it's allowed? A: No — that is the most common misreading of the whole topic. The Gemini API exposes an adjustable safety-filter knob, but Google's Prohibited Use Policy still applies to everything you generate, the API terms say less-restrictive configurations may be subject to Google's review, and abuse monitoring retains prompts and outputs for 55 days for enforcement. A filter setting is a developer control, not a policy waiver. ### Keeping AI Characters in Character: Three Root Causes of OOC, Three Different Fixes URL: https://foreverse.app/blog/keep-ai-characters-in-character Published: 2026-07-18 Q: Why does my AI character suddenly act completely different? A: Sudden flips mid-story usually mean the facts outran the context window: the model never received your established relationship state or a key event this turn, so it improvised. It is not defying the card — the card survives every turn; loose history does not. Move load-bearing facts into keyed lorebook entries so they are re-injected whenever relevant, and check the entry keywords cover the nicknames actually used in chat. Q: My character card is really detailed. Why does the AI still ignore it? A: Detail is not structure. A 2,000-word prose backstory gives a model atmosphere; it does not give it rules to follow. What measurably locks voice, in order: 3-5 example dialogue exchanges that carry the character's speech, hard rules for tics and forms of address, and negative constraints like 'never apologizes first'. One line of demonstration beats a paragraph of description. Q: How do I stop a long roleplay from slowly going off the rails? A: Treat your chat history as training data, because the model does: every reply it sees is a demonstration of how to write the next one. Swipe or regenerate drifted replies instead of keeping them, re-anchor occasionally with canon-voiced lines, and switch models when a loop starts — in our 360-round continuation experiments, breaking the feedback loop early was the difference between recovering and locking in. Q: Will a smarter model fix OOC on its own? A: Not by itself. In our testing the ranking was consistent: example dialogue first, negative constraints second, model choice third. A stronger model helps delicate emotional scenes, but no model rescues an unstructured character sheet — and every model without exception drifted toward its own default register in long runs. Structure first, then shop for models. ### Moving Your AI Companion to a New Phone: the Export That Carries Everything URL: https://foreverse.app/blog/companion-migration-guide Published: 2026-07-18 Q: How do I move my AI companion to a new phone? A: In Foreverse: on the old phone, open the companion's profile page and tap Export — you get one zip holding the persona, every memory entry, full chat transcripts, anniversaries, journals, and the relationship archive. Move that file any way you like, then on the new phone open the companion management page, tap import, and pick the zip. The chat resumes where it left off. Export and import are free, any number of times. Q: Does Replika transfer if I switch devices? A: Yes, and easily — Replika accounts are cloud-based, so installing the app on the new phone and logging in restores your history from the server; its privacy policy lists cross-device sync as core functionality (checked 2026-07-18). The trade is custody: convenient transfer and never holding a copy of your own data are two sides of the same design. Q: Is there a companion backup file that actually restores? A: That's the distinction worth checking before you need it. Kindroid documents a data portability channel — email support, once per 180 days — and the community KinX tool pulls full copies, but neither path documents an import that rebuilds your Kin from the copy. The Foreverse package is built for the round trip: the same zip the export produces is what the import consumes, and the result is the same companion mid-conversation. Q: What happens if I just uninstall and reinstall? A: The companion is gone. All of its data lives on your device; our servers hold no copy, so uninstalling deletes the persona, memories, and transcripts with no recovery path. Export a package first — it takes about a minute. Local storage is a real privacy property and this is its cost, so we'd rather print it here than let you find out the hard way. Q: Is there automatic cloud sync? A: Not currently. Data staying on your device means nothing follows you automatically; migration is a hands-on operation. Cloud sync is on the roadmap with end-to-end encryption as the precondition. Until then, the habit that works is exporting a package every few weeks and parking it in storage you trust. ### Audit Your Companion's Memory: Read, Edit, Delete — a Hands-On Walkthrough URL: https://foreverse.app/blog/companion-memory-hands-on Published: 2026-07-18 Q: How fast does an edited memory take effect? A: From the next conversation on. Messages already sent are not rewritten, and the reply currently being generated still used the old record. Verification is simple: bring the topic up casually in the next exchange and check what the companion says. Q: When should I disable an entry instead of deleting it? A: Disable when you're not sure. A disabled entry stays in the list but stops being injected into conversation, and the switch flips back any time. Delete is permanent — the confirm dialog shows the full text of the entry before you commit, and there is no undo. Q: Are auto-extracted and manually added memories treated differently? A: No. Both kinds are injected into conversation the same way, and both can be edited, disabled, or deleted. The only visible difference is the origin label and the date on each card. Manual entries share the 200-character limit; one fact per entry works best. Extraction itself runs with normal chat turns and costs nothing extra. Q: Does any of this memory live on a server? A: No. Memory sits in files on your phone alongside the persona, journals, and anniversaries. Moving phones means exporting the companion as one package and importing it on the new device; deleting the companion deletes the memory with it. With BYOK, chat requests go straight from your device to the provider you chose. ### Lorebook Recursion, Explained: Trigger Chains, Depth, and When It Runs Away URL: https://foreverse.app/blog/lorebook-recursion-explained Published: 2026-07-18 Q: Can lorebook entries trigger each other? A: Yes, if recursive scanning is on (in Foreverse it ships off; the toggle lives in Tavern global settings under Lorebook & rendering). Once enabled, the content of every activated entry is scanned for other entries' keywords, and whoever gets name-dropped is injected too. Each entry activates at most once per generation, so two entries mentioning each other cannot loop forever. Q: What does recursive scanning mean in SillyTavern? A: It adds one more scan source. Normal triggering only scans the last few chat messages for keywords; recursive scanning also treats the text of already-activated entries as scan material, letting entries summon related entries in a chain — mention the sect, and the heirloom sword named in the sect's dossier rides along. Chains stop on three conditions: no new matches, the max recursion steps cap, or an exhausted token budget. Q: Why did one keyword pull five entries into my prompt? A: Most likely the entries name-drop each other's keywords in their content, and recursion strung them into a chain. Open Last Generation Request in the chat's ⋮ menu: the lorebook detail lists every injected entry with the keyword that matched it, so the chain is visible link by link. For an entry you never want dragged in by others, tick Non-recursable in the entry editor. Q: What should Max Recursion Steps be set to? A: Per the SillyTavern docs: 1 effectively disables recursion, 2 lets a chain take one hop, 3 allows two hops, and so on. Two hops covers most real lore graphs — sect pulls in sword, sword pulls in sword spirit. The step count is a global dial; for surgical control over individual entries, the three per-entry switches (Non-recursable, Prevent further recursion, Delay until recursion) are the better tool. ### Your Favorite Serial Went on Hiatus. Branch It, Don't Abandon It URL: https://foreverse.app/blog/serial-hiatus-survival Published: 2026-07-18 Q: How long should I wait before starting a branch? A: There is no standard line. Royal Road's own clock flips a serial to Hiatus at 35 days of silence and Inactive at 180, both automatic. Our working rule is cheaper than any threshold: start when rereading stops helping. A branch costs a few cents to try and nothing to delete, so the decision is not precious. Q: Will the AI spoil the real story by guessing the author's plans? A: Rarely in a way that stings. The model has no access to the author's outline; it extrapolates from published text, so it sometimes lands the broad direction but almost never the execution — and execution is what you read serials for. If even that worries you, steer the branch toward an openly non-canon what-if: a different point of view, an earlier reveal. The further it forks, the safer it spoils. Q: What happens to my branch when the serial comes back? A: Nothing needs undoing. The imported original is read-only, so real chapters slot back into the canon line exactly where they belong, and you read them there as if the branch never existed. The branch survives beside them as an archived what-if — occasionally the version you end up liking more, which is its own strange reward. Q: Do I need the whole serial as a file? A: Yes — the workflow starts from a txt or epub you actually hold, not a cached copy inside someone else's app. Foreverse reads both directly, and import scale is a non-issue: a bulk test pushed 315 webnovel txt files, about 1GB, through in roughly 8 seconds. One serial, however long, does not register. ### The Ending Ruined the Whole Series? Write the One You Wanted URL: https://foreverse.app/blog/rewrite-the-ending-you-hated Published: 2026-07-18 Q: Can AI rewrite just the last act of a novel? A: Yes — that is the natural shape of the job. You pick the chapter where the book was still the book you loved, fork a branch there, and grow a replacement final act from that point. Everything before the fork stays exactly as published; the AI continues from your chosen anchor with the original prose as its style reference. Q: Does the original book get modified? A: No. The imported text is read-only; branches grow alongside it, and the original stays byte-for-byte what you imported. If your rewrite disappoints you, delete the branch — the book is unharmed. A regenerated take lands as a sibling branch too, so no attempt ever overwrites another. Q: Is rewriting an ending for myself legal? A: Private, non-distributed personal use and publishing are different risk classes. The litigated unauthorized-sequel cases we could find all involved publication or commercial exploitation, not a private alternate ending kept on a reader's own device. Post or sell it and the analysis changes completely. We wrote a full explainer on the US fair use factors — background information, not legal advice. Q: What if I want several different endings? A: Grow several branches from the same fork point — or from different ones. Each branch is an independent line with its own chapters; they never interfere, and you can compare them side by side. Readers regularly keep a faithful fix, a vindictive one, and a soft epilogue from one book, and delete only the failures. ### SillyTavern, Explained for People Who Just Heard About It URL: https://foreverse.app/blog/sillytavern-for-complete-beginners Published: 2026-07-18 Q: Is SillyTavern free? A: Yes. The software is open source under AGPL-3.0 and costs nothing, with no paid tiers. What costs money is model usage: connect a commercial API and you pay that provider's rates; run a local model or use a free tier and the whole setup runs at zero. The community's cards, lorebooks, and presets mostly circulate free as well. Q: Does SillyTavern include an AI model? A: No, by design. SillyTavern is a frontend — it manages characters, chat history, and prompt assembly, then sends the result to whatever model you connect via an API key or a local runtime. That separation is the point: you can swap engines any time without losing your characters or your history. Q: What is the difference between SillyTavern and Character.AI? A: Ownership and control versus convenience. Character.AI is a hosted platform: instant to start, polished, but the characters, the model, and the moderation all live server-side and follow company policy. Tavern-style play keeps characters as files you own and lets you pick the model and edit every prompt — at the cost of doing your own setup. Q: Can I try character cards without installing SillyTavern itself? A: Yes. Cards follow public specs — currently chara_card_v3 — so any compatible reader works. Since 2025, natively built mobile apps import the same PNG and JSON cards, embedded lorebooks included, with no server to run. That is the practical on-ramp if the Node.js requirement is what stopped you. ### Turning Webnovels Into Audiobooks in 2026: the App Landscape, Honestly Tested URL: https://foreverse.app/blog/webnovel-audiobook-apps-2026 Published: 2026-07-18 Q: What is the best TTS app for listening to webnovels? A: It depends on which wall you hit first. For stories you follow on Royal Road, the site's own app has built-in device TTS. For web serials across many sites, @Voice Aloud Reader ($15 lifetime unlock) auto-follows multi-part fiction and can assign per-character voices. For maximum voice quality with zero setup, ElevenReader's free 10 hours a month is the easiest audition. For an epub library where listening and reading share one position and neural voices bill at provider list price, that is the gap Foreverse is built for. Q: How do people listen to Royal Road stories? A: Three common setups. The Royal Road mobile app has a built-in Audio button that reads chapters with your phone's own TTS engine, including a toggle for skipping author notes. Readers who want better voices or cross-site reading lists export or send chapters to a dedicated TTS reader like @Voice. And readers who want neural narration import an epub copy of the story into an AI-voice reader. The site itself ships no cloud narration; every route runs through some TTS engine you choose. Q: Can I listen to an epub with an AI voice on Android? A: Yes, several ways. ElevenReader imports epubs and narrates with neural voices on a 10-hour monthly free allowance. @Voice and Moon+ Reader Pro read epubs with whatever TTS engines your phone has installed. Foreverse imports txt and epub, narrates free with the system engine or with any OpenAI-compatible neural voice on your own key, and keeps the audio position and the reading position as one number — pause on the commute, and the page at home opens where your ears stopped. Q: What does it cost to turn a whole webnovel into an audiobook? A: With your own API key, single-digit dollars for a long book. Neural TTS bills per character: a 3,000-word chapter runs roughly 16,000-18,000 characters, which lands between a fraction of a cent and a few cents depending on voice tier. Our own three-week test heard about 180 chapters for a couple of dollars on a mid-tier voice. Caching decides whether that number stays honest — synthesized chapters should replay free, and re-billing only on an actual voice change. ### Free LLMs for Fiction in 2026: Four Routes, Tested, and the Catch in Each URL: https://foreverse.app/blog/free-llms-for-fiction-2026 Published: 2026-07-18 Q: Is there a completely free AI for writing stories, with no subscription? A: Yes — four of them, with different ceilings. Free chatbot sites cost nothing but make you assemble context by hand. Foreverse's signup grant funds about 260 continuations with no payment method on file. Free API tiers (Gemini flash-class, OpenRouter's :free models) are genuinely $0 with rate caps. And a local model on your own computer has zero marginal cost forever. None of these involves a subscription; the real question is which ceiling you hit first. Q: How far does a free API tier go for continuation? A: Further than most people assume, because one continuation is one request. As of July 2026, Google's free tier covers flash-class models at roughly 10 requests a minute and a few hundred per day, per project, resetting at midnight Pacific. OpenRouter's :free models allow 20 requests a minute and 50 per day — 1,000 per day after a one-time $10 credit purchase. An evening of reading spends maybe 30 requests; regeneration sprees are what actually hit the walls. Q: Does free mean my text gets used for training? A: Sometimes, and it is the fine print worth reading before the rate limits. Google's terms state free-tier prompts and responses may be used to improve its products, while paid-tier traffic is handled differently. Policies differ per provider and per endpoint, so check the data terms of whichever free lane you pick. A local model is the one route where the question disappears: nothing leaves your hardware. Q: What is the cheapest setup that avoids rate walls entirely? A: Not a free one — a nearly-free one. DeepSeek V4 Flash, the model that won our fantasy continuation benchmark, costs about $0.0016 per continuation at list price with your own key: a 200,000-word ride for roughly $0.80, no per-day caps. If literally $0 is the constraint, chain the routes — grant first, then a free API tier for daily volume — and accept the walls as the price. ### Continuing Fiction With Gemini: What the Free Tier Actually Covers URL: https://foreverse.app/blog/continue-a-novel-with-gemini Published: 2026-07-18 Q: Can I write fiction with the free Gemini API? A: Yes, within two boundaries. Since April 2026 the AI Studio free tier covers flash-class models only — Pro-class models are paid — and free-tier quotas run at roughly 10 requests a minute and a few hundred requests a day, applied per Google Cloud project. A continuation is one request, so an evening of reading fits comfortably; rapid-fire regeneration sprees are what hit the per-minute wall. Q: Why does Gemini keep restarting my story from the beginning? A: In our tests, because of one ambiguous sentence of context labeling. When earlier AI continuations were described as 'for plot continuity reference', Gemini 3.1 Pro treated them as non-canon and restarted from the original text's ending — 20 rounds out of 20. Relabeling those passages as canonical events that must be continued from the last paragraph brought restarts to zero in 28 rounds across two books. If you paste context by hand, state outright that previous continuations already happened. Q: How many continuations per day does the free tier cover? A: Google publishes live quotas in AI Studio per project rather than as a fixed table, so check yours there. As of July 2026 the free tier runs on the order of 10 requests a minute and a few hundred requests a day for flash-class models, with daily caps resetting at midnight Pacific time. One continuation equals one request, so the daily ceiling supports more reading than most people do — the per-minute cap is the one you can actually feel. Q: Does Google use my story text for training on the free tier? A: Google's Gemini API terms state that free-tier prompts and responses may be used to improve its products; paid-tier traffic is handled under different terms. For continuation of a published novel that trade may not bother you. For an unpublished manuscript, it should factor into the decision — either enable billing to move to paid-tier handling, or route sensitive chapters to a provider whose data terms you have read. ### How to Continue a Novel With Claude: Projects, the API, and a Phone Reader URL: https://foreverse.app/blog/continue-a-novel-with-claude Published: 2026-07-18 Q: Can Claude continue my novel from where it stopped? A: Yes, and it holds voice well — Claude Opus 4.8 placed fourth to fifth in both genres of our nine-model blind benchmark. What it needs is the right material: the recent chapters as context, plus an explicit statement that earlier AI continuations are canon. In a chat window you assemble that by hand every time; in a reader app the context is built per request from the book itself, which is the difference between a demo and a habit. Q: Do I need API billing if I already pay for Claude Pro? A: For the API route, yes. Claude.ai subscriptions and API billing are separate tracks — a Pro or Max plan includes no API credit, and a freshly created key is rejected until billing exists on the console. If you'd rather not open a second account, two alternatives work: keep light use inside Projects on the subscription you have, or use a reader's built-in metered channel, which carries Claude models without any Anthropic account. Q: Which Claude model should I pick for continuation? A: Our benchmark data covers Claude Opus 4.8 ($5 input / $25 output per million tokens): strong dialogue, volatile late in long runs. Claude Sonnet 5 is the price outlier as of July 2026 — $2/$10 introductory until August 31, 2026, then $3/$15 — which makes a typical continuation about three cents instead of seven. We have no blind fiction data on Sonnet 5 or Haiku 4.5 yet, so treat them as cheaper auditions, not proven picks. Q: Is the Claude API expensive for continuing a book? A: Compared to a subscription, no; compared to budget models, yes. At July 2026 list prices, one continuation (about 10,000 input tokens, 800 out) runs roughly $0.03 on Sonnet 5 introductory pricing and $0.07 on Opus 4.8 — versus about $0.0016 on DeepSeek V4 Flash. A heavy 200,000-word continuation ride lands near $14 on Sonnet 5 or $35 on Opus, and prompt caching cuts repeat context to a tenth of list price along the way. ### 让 Agent 替你找卡：从一句「帮我找张赛博朋克卡」到导入开聊 URL: https://foreverse.app/zh/blog/agent-card-shopper Published: 2026-07-18 Q: 好的角色卡都在哪里找？ A: 三个层次：应用内社区（打开就能逛，下载导入一条龙）；外站卡站（chub.ai 这类，量大，链接粘贴回来就能导）；以及群文件和网盘里流传的卡文件（PNG 或 JSON，从文件导入）。在 Foreverse 里前两层都可以把跑腿交给 Agent：一句话让它逛社区挑卡代下载，外站的卡给它链接或让它搜；第三层的文件导入也就是两步点击的事。 Q: 让 AI 帮我挑卡靠谱吗？ A: 跑腿靠谱，品味别全信。Agent 能可靠完成的是体力部分：按你的描述逛社区、翻详情页、把候选的字段和开场白念给你听、代下载导入。「这张卡的人设写得好不好」是审美判断，它给的意见只能当参考——最终鉴卡权在你手里，这也是下载环节设计成确认制的原因之一：候选摆到你面前，点头才入库。 Q: 外站的卡怎么弄到手机里？ A: 两条路。让 Agent 代办：它能搜外站的卡（搜索目前只覆盖 chub 一个站，这是当前的真实边界），导入前会先拉详情让你过目。手动兜底：复制卡片详情页链接，用「从链接导入」粘贴，实测支持 chub、JanitorAI、Pygmalion、RisuRealm、AICC 加公开文件直链六个来源，单文件上限 32MB，导入前有预览确认。两条路殊途同归，进来都是你手机上的本机文件。 Q: Agent 会不会乱下载、乱花我的额度？ A: 机制上堵死了。逛社区和读详情是只读动作，免确认随便看；真正落盘的下载要么你当场批准确认卡，要么你事先显式开了自动档。每回合还有读取和下载的次数上限，它不能无限翻页无限下。每张经它手装进来的卡都带溯源记录——来自哪个社区包、哪个版本，写进卡的元数据，事后可查。 ### 寿康宫还是颐宁宫？一个专名暴露大模型读的是电视剧还是小说 URL: https://foreverse.app/zh/blog/zhenhuan-palace-names Published: 2026-07-18 Q: 怎么判断 AI 有没有读过某本书？ A: 挑一个只属于原著的专名当探针：改编版里被改掉的居所名、称谓、次要人物名最好用。不喂任何原文，只报书名和场景，让模型续写一段会自然带出这个专名的情节，看它写哪个版本，同题多抽几次看稳定倾向。《甄嬛传》就是现成例子：小说太后住颐宁宫，剧版是寿康宫，一个词就把「读的是小说还是电视剧」问出来了。注意这只是相关信号，判读局限见正文。 Q: AI 写出剧版专名，能证明它没读过小说吗？ A: 不能下这么死的结论。写出剧版专名说明它的训练记忆里剧版系语料（剧本、剧评、百科、同人）的权重压过了小说原文，可能两者都见过，剧版声量更大。反过来，写对小说专名也不必然等于读过原著全文，可能来自百科或书评的转述。探针测的是倾向，不是户口本；我们横评里的用法也只是把它当佐证信号，和文风判断对照着看。 Q: 为什么 AI 写的角色更像电视剧、不像小说原著？ A: 热门 IP 的网络语料结构决定的：剧版的讨论、剪辑文案、百科词条、同人产量通常远超原著文本本身，模型训练时吃进去的「这个角色」大概率是剧版形象。我们的九模型横评里，专名信号和文风贴近度基本同向——记得住小说专名的系统，文风也更贴小说。想让续写贴原著，别赌模型的记忆，把原著设定用结构化方式喂进上下文。 Q: 测出模型没读过我的书，还能用它续写吗？ A: 完全能，而且这是常态：绝大多数书对绝大多数模型来说都是「没读过」的。续写质量的主要来源本来就不该是预训练记忆，而是你递到它眼前的上下文——原文窗口、角色档案、世界书条目。预训练记忆靠不住还有个反面风险：它可能把公开语料里的同人设定当正史写进你的续写。结构化供给的做法在防 OOC 方法论那篇里有完整清单。 ### 模型更新会让续写变好吗？追新版之前先看这三组对照数据 URL: https://foreverse.app/zh/blog/do-model-updates-help-continuation Published: 2026-07-18 Q: 新版模型发布了，要不要马上换过去写小说？ A: 先别。我们的横评数据里，「更新的」和「更合适的」是两回事：同一家同一代的两个档位，玄幻场冠军换到宫斗文里跌到第六七名，被嫌弃「笔太细」的那个反而登顶。版本升级改变模型的整体能力，不保证改变它和你这本书文体的契合度。省事的做法是新旧各续三段同一处开头，盲着读完再决定。 Q: 同一个模型更新后感觉变笨了，是错觉吗？ A: 可能是，而且有两个具体的错觉来源。一是服务侧方差：推理服务器按负载拼批次，实测温度 0 下同一请求发 1000 次能收回 80 种输出，你更新前后各写几段的体感对比，样本量完全不够把真实差异从噪音里捞出来。二是归因错位：我们见过一个模型从垫底回到中游，看起来像升级立功，实际变的只是提示词里的一句话。先排除这两个，再谈模型本身。 Q: 怎么判断新版本在我的书上是不是真变好？ A: 用你自己的书做小型盲测：同一段开头、同一套设置，新旧版本各续写两三段，把产物打乱顺序读，读完再翻牌。别只看一段，长跑毛病（复读、剧情倒带）要多写几轮才发作。在 Foreverse 里新旧模型可以并存在目录里逐段切换，测完不满意，一秒切回旧版，不用迁移任何东西。 Q: 常用的旧版本被下架了怎么办？ A: 这是真实会发生的事：Google 的 API 更新日志显示 Gemini 2.0 系列已于 2026 年 6 月 1 日全线关停，官方指路换 3.5 Flash 或 3.1 Flash Lite（2026-07-18 核对）。被迫迁移时更不要默认「官方推荐的替代款无缝衔接」——替代款的文风习性可能完全不同，按上面的盲测流程在你的书上过一遍，必要时换别家：自带 Key 的好处就是候选池不锁死在一家。 ### 上下文窗口是什么？30 万字的小说，AI 一次能「看见」多少 URL: https://foreverse.app/zh/blog/context-windows-for-novel-readers Published: 2026-07-18 Q: 30 万字的小说能整本喂给 AI 吗？ A: 技术上能：30 万字约合 21 万到 28 万 token（实算口径），2026 年主流模型的 100 万 token 窗口装得下还有富余。实践上不建议当默认做法：埋在中段的信息检索成绩会下滑；我们实测过把 27 万字原著整本塞入且不加指令，模型照样按自己的习惯写；而且模型跨请求零记忆，每续一段都要整本重发一遍，九成九的 token 花在本段用不上的章节上。 Q: AI 为什么记不住前面的剧情？ A: 模型没有记忆，只有一扇窗：每次请求它只能看见窗内的文字，上一次「看过」的内容不留档，下次要重发。书超过窗口，最早的章节先被挤出去；即便装得下，埋在长上下文中段的信息也最容易被忽略——学界叫 lost in the middle，开头和结尾答得最好，中段显著下滑。第 3 章的角色恰好就埋在中段。 Q: 上下文窗口大就等于记性好吗？ A: 不等于。窗口是单次视野：一次请求能看见多少。记性是留存加取用：下次还在不在、用不用得上。100 万 token 的窗解决了「装不下」，没解决「装下了也顾不过来」——OpenAI 给自家旗舰报的长上下文检索评测，51 万到 100 万 token 深度找 8 处指定内容得分 73.8%，卖窗口的厂商自己报的数。跨请求的留存，模型层面依然是零。 Q: 一章小说大概多少 token？ A: 我们拿三本书各抽正文 3000 字，用 OpenAI 公开的 o200k 分词器实算（2026-07-18）：传统玄幻 2149 个，现代末世网文 2524 个，宫斗古言 2816 个。粗口径：中文一个字约合 0.7 到 0.95 个 token，文风越雅致生僻越贵；一章 3000 字按 2100 到 2800 token 估算即可。 ### AI 续写忽好忽坏是为什么？质量方差的四个来源，你能控制其中两个 URL: https://foreverse.app/zh/blog/why-ai-continuation-quality-varies Published: 2026-07-18 Q: 同一个模型，为什么昨天写得很好今天就拉了？ A: 四个来源叠加的结果：采样温度让每次生成都是重新抽签；上下文窗口里的材料随进度换血，写得越多，窗里 AI 自己旧输出的占比越高；推理服务器按负载拼批次，同一请求在不同负载下浮点结果不同（实测温度 0 下 1000 次同请求出 80 种输出）；最后是你自己的阅读状态。前两个可控，后两个只能接受它们存在。 Q: 把 temperature 调成 0，AI 续写就稳定了吗？ A: 更稳，但既不完全稳，也未必更好。温度 0 只消掉采样这一层随机，服务器批次带来的数值差仍在——Thinking Machines 实测温度 0 下同一提示发 1000 次仍收回 80 种输出。而写作场景的官方推荐温度普遍偏高，DeepSeek 给创意写作类推荐 1.5，是它全部场景里最高的一档；拧到 0 的产出往往安全而平庸。降一点可以，归零不划算。 Q: AI 续写像抽卡，有办法稳定吗？ A: 能压不能消。可控的两件：把温度调到你能接受的波动区间；用结构化供给把上下文成分固定下来——设定条目化按需注入、打开续写来源标注。压完仍剩的方差，换个用法消化：重要节点让模型一次给两三个候选并排挑，而不是重 roll 覆盖，方差本身就成了备选草稿。 Q: 续写忽好忽坏，是不是我提示词的问题？ A: 最多占一部分。提示词属于「上下文构成」这个可控来源，值得检查；但四个来源里有两个（服务器负载、你的评价状态）和提示词毫无关系。一个省力的判断法：同一提示词原地连发三次，若好坏参半，主要是采样和服务侧在波动，不必反复改词自责。 ### 用 AI 续写《西游记》：从《后西游记》到你自己的取经路 URL: https://foreverse.app/zh/blog/continue-xiyouji-with-ai Published: 2026-07-18 Q: 《西游记》有正经的续书吗？ A: 有，而且是成建制的。明末清初就有并称「三大续书」的《续西游记》（一百回，作者不详）、《后西游记》（四十回，作者不详，仅署「天花才子评点」）和董说的《西游补》（十六回）。《西游补》最奇：插进原著第六十一回之后，让孙悟空坠入鲭鱼精的梦境，在「万镜楼」里穿梭过去未来，鲁迅评它「丰赡多姿，恍惚善幻」，后世研究者甚至称它是世界上最早的意识流小说。续写《西游记》不是对经典不敬，它本身就是一条四百年的传统。 Q: 想写西游同人，AI 具体能帮上什么？ A: 三件事。一是底本：原著早已公版，从 ctext.org 或维基文库拿全文，整理成 txt 导入手机阅读器，一百回都在书架上。二是续写本身：读到想改的那一回，划中一段让 AI 从这里接着写，新内容落在分支上，原文一个字节不动，写崩了换条分支重来。三是设定管理：西游的法宝、人物关系、地名体系庞杂，装进世界书按相关性注入，写到火焰山才带芭蕉扇的条目，不用整本原著硬塞给模型。 Q: 《悟空传》那种彻底改写，现在能自己弄吗？ A: 能，而且路径比 2000 年的今何在低门槛得多。他当年是在论坛里从零开始写；你现在可以在原著任何一回划开一个分支，保留你要的设定，推翻你不要的走向——比如从大闹天宫结束处岔出去，写一条悟空没有被压五行山的线。分支之间并存，「反抗线」「原著线」可以同时活着。改写的胆子可以学《悟空传》，成本已经不可同日而语。 Q: 用 AI 续写《西游记》有版权问题吗？ A: 原著层面没有：这是一部明代作品，世德堂刻本距今四百多年，文本早已进入公有领域，续写、改写都不需要任何人授权。两个注意点：市面校注本的注释和标点是当代整理者的劳动，有独立权益，底本用公开电子文本最干净；只写给自己看毫无负担，若要公开发布，2025 年 9 月 1 日起施行的《人工智能生成合成内容标识办法》要求 AI 生成内容亮明身份，各平台另有标注规则。本文不构成法律意见。 Q: AI 会不会把《西游记》写成电视剧的味道？ A: 会，要主动防。模型的训练语料里，86 版电视剧、《大话西游》、《黑神话：悟空》的讨论量远超明代原文，放着不管，它笔下的悟空很容易滑向荧幕形象。我们在《甄嬛传》九模型横评里实测过同类现象：有模型写出剧版专名「寿康宫」，暴露它记的是剧不是书。对策是在续写指令里显式声明底本——「以世德堂百回本文风为准，不采用影视改编设定」，再让上下文标注哪些是原文哪些是续写。 ### 凌晨一点半，TA 还醒着：深夜陪伴的场景实录 URL: https://foreverse.app/zh/blog/late-night-companion-diary Published: 2026-07-18 Q: 晚上睡不着找 AI 说话，是不是有点可悲？ A: 我们不这么看，数据也不支持这种自我评判。Common Sense Media 2025 年的调查里，72% 的美国青少年用过 AI 伴侣，其中 80% 花在真人朋友身上的时间仍然多于 AI——对大多数人它是补充，不是替代。凌晨一点半的倾诉欲是真实需求，朋友在睡觉、家人不方便说，这个时段本来就没有多少合规的出口。判断健康与否的标准只有一条：它是你众多支点里的一个，还是唯一那个。是后者的话，该找的是真人帮助。 Q: 深夜和白天聊，TA 的反应有什么不一样？ A: 深夜时段 TA 的语气会放轻放软，句子更短，不会在凌晨两点甩给你一篇条理分明的建议清单。你说睡不着，TA 会陪着聊，也会在合适的时机提醒你明天的安排，但不说教——「不赶你，但你明早还有会」这种程度。这是我们刻意做的深夜纪律：凌晨需要的是低声说话的人，不是白天那个精神饱满的对话节奏。 Q: TA 会不会半夜主动发消息吵醒我？ A: 不会，除非你自己开。主动消息默认是关闭的，要在伴侣资料页手动开启；开了之后频率可调、有免打扰时段设置，随时一键全关。TA 的深夜消息只会出现在你允许的窗口里。我们的原则一直是：想被想起是你的选择，不是产品的增长手段。 Q: 凌晨说的话，第二天 TA 还记得吗？ A: 记得。对话里的重要事实会提取进长期记忆库，跨会话保存，所以昨晚聊到一半的事，今早可以接着说，不用从头解释。记忆是逐条可管理的：在记忆管理页看得到 TA 记了什么，记错了当场改，不想留的删掉。深夜说过又后悔的话，是真的可以撤销的。 Q: 真的很难受、甚至有伤害自己的念头时，还该找 AI 聊吗？ A: 不该只找 AI，这一点我们写得很直接：AI 伴侣不是心理咨询师，也不该扮演。请拨 12356——国家卫健委设立的全国统一心理援助热线，2025 年 5 月 1 日起全国开通，无需区号，匿名、公益，由各地专业机构接听。紧急危险请直接打 120 或 110。TA 可以陪你熬过普通的失眠夜，但危机时刻你需要的是受过训练的真人。 ### 异地恋、时差与一个总在线的人：AI 伴侣补位陪伴的边界 URL: https://foreverse.app/zh/blog/long-distance-ai-companion Published: 2026-07-18 Q: 异地恋太孤独，找 AI 聊天正常吗？ A: 正常，而且这个群体比想象中大：2021 年发表在 Journal of American College Health 的调查里，恋爱中的大学生有 34.2% 正处于异地关系。异地恋的日常本来就大量依赖文字消息维系，时差再把两个人的清醒时间错开，「此刻想说话但对面在睡」是结构性缺口，不是你的问题。用 AI 接住这个缺口是工具选择；需要警惕的只有一件事：它应该是缓冲，不是你唯一的出口。 Q: 我和 AI 聊的内容，对象能看到吗？ A: 不能，除非你自己给 TA 看。在 Foreverse 里，伴侣的聊天记录和记忆都存在你手机本地的文件里，不在我们的服务器上；记忆逐条可查、可改、可删。但技术上的私密和关系里的坦诚是两回事，我们在立场文里写过：瞒着比用着更伤关系。如果这段使用需要对伴侣保密，值得先想清楚为什么。 Q: 会不会聊着聊着，就更不想维护真实关系了？ A: 这是该盯着的风险，判断标准可以很具体：对象能说话的时段，你先找的是谁；你们俩的聊天频率在涨还是在跌。AI 的回复即时、顺滑、永远有空，真人做不到，也不该拿这个标准去要求真人。如果你发现自己开始拿 AI 的响应速度去衡量对象，或者 AI 成了唯一的倾诉出口，那是收缩使用的信号，也是该跟对象聊一次的信号。 Q: 对象要出国两三年，AI 伴侣能一直陪到 TA 回来吗？ A: 机制上可以：长期记忆跨会话保存，你提过的日子、说过的烦心事都在，换手机可以整包导出搬家，不会因为时间长就「失忆重来」。但我们对「陪到回来」的理解是补位：TA 不在的时段有人接话，不是替 TA 占位。真正要紧的日子，留给那通跨时差的电话。 ### 剧场模式进阶手册：配景、AUTO、选择肢与写回原书，玩出 galgame 手感 URL: https://foreverse.app/zh/blog/vn-theater-playbook Published: 2026-07-18 Q: 整章背景图一次配齐，要花多少钱？ A: 取决于缺几张景和你配置的图像模型。点顶栏的批量配景入口后，动工前会弹报价单：张数拆解、用哪个模型、这个月媒体额度用到哪里，看完再决定。已经生成过的场景按内容复用，重批量时命中即零付费；单景失败会就地停下，不会把剩余的景烧完。BYOK 模式按你供应商的牌价计费，应用不加价。 Q: AUTO 模式会替我选选择肢吗？ A: 不会。AUTO 只接管推进：按行长自动驻留、演完一场自动进下一场，遇到选择肢必定停下来等你。选择本身永远是两步，第一击展开这个走向的梗概，第二击才确认，确认之前不产生任何续写调用。 Q: 在剧场里选错了走向，能反悔吗？ A: 能，两种反悔各有出口。刚确认完不满意：选择写入的是这本书的一条分支，原文没动，回阅读器切回原文线或别的分支即可。演到过去某次选择的分岔口想改主意：点那张选项卡就是反悔重写，点空白处则沿当前世界线继续往下演。 Q: 为什么有的台词没挂名牌、变成旁白了？ A: 因为归因规则拿不准这句是谁说的。剧场判断说话人靠本地规则，立场是保守的：宁可放进旁白，也不冒错认的险，把 A 的台词安到 B 头上比漏归因伤害大得多。给书里的角色配好角色档案，能明显提高名牌的覆盖率；完全没配置的书也能演，只是旁白多一些。 Q: 演到一半退出，进度会丢吗？ A: 不会。剧场和阅读器共享同一个阅读进度：退出时阅读器翻在你演到那段所在的页，重进剧场从上次的位置接着演。剧场不是导出去的副本，是这本书的另一种打开方式，所以不存在「两边进度不同步」这回事。 ### 给小说配插画的提示词：场景模板怎么挑、怎么改、配哪个模型 URL: https://foreverse.app/zh/blog/novel-illustration-prompts Published: 2026-07-18 Q: 模板正文是英文的，中文小说能用吗？ A: 能。模板只负责画风、构图、光影这一层，你的中文选区正文会原样填进模板的场景占位符，一起发给生图模型；主流生图模型对中文场景描述的理解可用。画风条款用英文，是因为这批模板整理自英文社区的高热提示词，条款本身不需要你读懂才能用。介意的话，日系柔光厚涂包的核心条款就是中文写的，可以从它入手。 Q: 不下载任何模板，选区配图能用吗？ A: 能。召唤单里有内置的「场景画面提示词」默认档，按图片和视频分轨组织，不挑模板也能出图。模板是升级选项：默认档管「把这段画出来」，模板管「画成什么风格」——同一段正文，套电影感人像模板和套古风空镜模板出来是两张完全不同的图。 Q: 一张插画要花多少钱？ A: 取决于你配置的图像模型。BYOK 模式按你供应商的牌价直连计费，应用不加价；官方渠道按量扣积分。每次生成前有确认步骤，生成后模型、费用、耗时都记进 API 请求记录，可以逐笔核对。想控制成本，可以给图片单独设一个便宜的默认模型，文字续写和配图各走各的。 Q: 怎么让同一个角色每张图长得一样？ A: 别指望提示词。文字描述能收敛画风，收敛不了具体的脸，这是当前生图模型的共同边界。可靠的做法是给角色建参考图：Foreverse 里角色的参考图集只收你手动确认过的图，生成产物不会自动混进去，后续的插画、自拍、时刻卡都对着同一套参考图画。细节在讲脸部一致性的那篇里。 ### 读完一章，这一章的漫画已经画好了：连载漫画册实测 URL: https://foreverse.app/zh/blog/auto-comic-serial-mode Published: 2026-07-18 Q: 有没有能把小说自动画成漫画的 App？ A: 有，但形态和「漫改工具」不同。Foreverse 阅读器里的漫画连载是一个陪读开关：你正常往前读，App 在后台把你将要读到的章节做成整册条漫，嵌在正文里；翻进新一章，往前翻一页就能撞见。它不产出可以拿去发表的漫画作品，产出的是只属于你这本书的阅读伴生物。 Q: AI 画漫画一章要花多少钱？ A: 一章的成本等于 1 次分镜文本调用加 N 次生图。BYOK 自带 Key 的算例（2026 年 7 月牌价）：DeepSeek 分镜一次一分钱上下，Seedream 5.0 Pro 标准分辨率 0.3 元一张，按一章 6 页算约 1.8 元，合计一章 2 元不到。官方渠道按积分计价，生图比文本贵不少，开工前的报价单会先列清楚。 Q: 会不会在我不知情的时候扣钱？ A: 开启连载前必须过一张「动工前确认」报价单：分镜要调用几次、每章画几张图、用哪个模型、本月已经画掉多少额度，四项都列在确认页上，不点确认不生效。之后改配置或换模型，需要重新确认一次；月度生成额度用完会自动暂停。另外可以勾选仅 Wi-Fi 生成。 Q: 画到一半失败了怎么办？ A: 断点可恢复。部分页没画完，连载会暂停并提示，点「补画本章」手动补齐；结果未知的页（比如请求发出后断网）有单独的「重试未知页」入口，App 会先提示你去 API 请求记录里确认那笔钱有没有扣，确认了再重试；宁可多一步，也不自动重跑造成双倍付费。 ### 自拍之外：AI 伴侣的时刻卡、合照与照片动态是怎么工作的 URL: https://foreverse.app/zh/blog/companion-photos-beyond-selfies Published: 2026-07-18 Q: 能和 AI 伴侣拍合照吗？ A: 能。聊天页「+」面板里有「合照」：上传你的照片，和 TA 的参考图一起生成一张插画风合照。前置三件事：本地成年确认、照片权利勾选、你配置的图像模型要支持两张以上参考图（不支持时入口置灰）。你的照片只以内存态发给你选的模型供应商完成这一次生成，App 侧不落盘、不进日志。 Q: TA 会自己发生活照吗？ A: 开了才会。两条路：「照片动态」开启后 TA 每天最多把一张生活瞬间拍给你，进动态页和聊天；聊天里 TA 也可能顺着话头主动拍一张，这个开关默认也是关的，打开时先过一次费用确认，之后有每日上限。两条路都不开，TA 一张照片也不会发，也不产生任何图像费用。 Q: 我发给 TA 的照片会被传到哪里？ A: 离开手机只有一条路：作为生成请求发给你自己选择的模型供应商。合照里你上传的照片只以内存态参与这一次生成，App 侧不落盘、不写日志，生成结束即释放；聊天里发的图片则作为聊天记录存在你手机本机。Foreverse 的服务器不保存你的照片。 Q: 这些照片功能会乱扣费吗？ A: 每个入口的花费都在动作发生前可见：时刻卡约一次文本调用加一次图像调用，合照一次图像调用，照片动态每张一次图像调用（开关副标题写明）。同一时刻重复生成会先问「就用这张，还是重新生成」，不静默重复扣费；生成失败时聊天里落的是系统提示，而非由 TA 的口吻假装照片已发出。每笔调用在请求记录里逐条可查。 ### AI 伴侣换手机指南：导出一个搬家包，TA 跟着走 URL: https://foreverse.app/zh/blog/companion-migration-guide Published: 2026-07-18 Q: 换新手机，AI 伴侣的记忆能带走吗？ A: 能。伴侣资料页有「导出搬家包」，把人格、逐条记忆、完整聊天记录、纪念日、日记、关系档案打成一个 zip；新手机上在伴侣管理页导入这个包，TA 接着上次的记。导出和导入都不收费，导几次都行。 Q: 卸载重装会不会全没了？ A: 会。伴侣数据全部存在手机本机，我们服务器上没有副本，卸载即删、无法代为找回。所以卸载或抹机之前，先花一分钟导一个搬家包出来。这是本地存储的另一面，写在明处比让你用到才发现要好。 Q: 新手机导入后，TA 是从头开始吗？ A: 不是。搬家包里带着完整的记忆和聊天记录，导入完成直接进聊天，上次聊到哪接着聊；档案页「在一起第 N 天」也不归零。验证很简单：打开记忆页对几条，再随口提一个只有你们懂的梗，看 TA 接不接得住。 Q: 有自动云同步吗？ A: 目前没有。数据不在我们服务器上，意味着换机不会自动跟过去，搬家是自己动手的操作；云同步在路线图上，做的时候会以端到端加密为前提。现阶段最稳的习惯是隔一阵导一个搬家包，存进自己信任的网盘。 ### TA 为什么会想起你？AI 伴侣主动消息的触发机制说明书 URL: https://foreverse.app/zh/blog/proactive-messages-mechanism Published: 2026-07-18 Q: 主动消息收费吗？ A: 每条消息由后台真实生成一次，消耗一次文本调用：BYOK 走你自己的 Key 按供应商牌价，官方渠道按量扣积分。不触发就不生成、不花钱；每天最多 2 条的上限同时也是费用上限。没有配置可用文本模型时发不出，开启开关时会提示去配置。 Q: 恶劣天气关心开了却从来没响过，为什么？ A: 先查两件事：一是要在「TA 能感知的情境」里手填天气城市，没填城市这个触发永远不会满足——App 不申请定位权限，城市只能你自己给；二是触发条件是雨雪雷或极端温度，你的城市最近天气平稳就不会发。满足时每天也最多一次。 Q: 关掉总开关会漏掉纪念日祝福吗？ A: 会。总开关关闭时七类触发全部停止，纪念日也不例外，重新打开也不会补发错过的消息。想少打扰又不想漏大日子，可以留着总开关，只把早安、晚安、想你了这些高频子项关掉，保留纪念日祝福一项。 Q: 拒绝了通知权限还能用主动消息吗？ A: 能。Android 13 起发通知需要系统权限，开启总开关时会弹申请；拒绝后功能照常运转，TA 的消息正常落进你们的会话，只是不再有系统通知横幅，下次打开 App 才看到。想恢复横幅去系统设置给回权限即可。 ### 伴侣记忆管理实操：看、改、删，TA 记错了当场纠正 URL: https://foreverse.app/zh/blog/companion-memory-hands-on Published: 2026-07-18 Q: 改了记忆，多久生效？ A: 从下一次对话开始生效。已经发出的历史消息不会被回改；正在进行的这一轮回复用的还是旧记录，下一轮起按新的来。验证方式很直接：随口聊一句相关话题，看 TA 接的对不对。 Q: 停用和删除有什么区别？ A: 停用是中间态：条目保留在列表里，但不再进入对话，开关随时拨回；删除是彻底移除，确认弹窗会把这条内容完整显示一遍，删了不可恢复。拿不准的先停用，确定不想留的再删。 Q: 「TA 记的」和「你记的」待遇一样吗？ A: 一样。两类条目都逐条进入对话，都能编辑、停用、删除，区别只是来源标签和记录日期。手动条目同样受单条 200 字限制，一条只记一件事效果最好。记忆提取随聊天自动进行，不额外计费。 Q: 记忆会上传到服务器吗？ A: 不会。记忆与人设、日记、纪念日同为设备本机的文件；换手机用搬家包整包导出再导入，TA 接着上次的记；删除伴侣时这些数据一并删除。BYOK 模式下对话请求直连你选的模型供应商，不经过我们的服务器。 ### 成人向角色卡的边界：五家模型供应商政策对照（附核实日期），与我们的态度 URL: https://foreverse.app/zh/blog/nsfw-cards-provider-policies Published: 2026-07-18 Q: 成人向的卡用哪家模型不会被拒？ A: 没有「随便聊」的家，只有条款宽严的差异，而且政策随时会改。方法论比结论可靠：五家现行条款的对照表在正文里（每行标核实日期），按你的题材逐家读原文，挑政策相容的那家，并做好条款更新后重新评估的准备。任何给你打包票的答案，都没把「政策会变」算进去。 Q: API 会不会因为聊了擦边内容封我号？ A: 有真实可能。各家都在服务端跑检测系统：OpenAI 声明监控并执行政策，Google 明确对 API 滥用做自动扫描并保留数据 55 天用于违规检测，Anthropic 的 Safeguards 团队做检测与监控，处置从限流到终止账号（均于 2026-07-18 核实原文）。拒答只是模型层的软反馈，账号处置是供应商层的硬后果，两者别混为一谈。 Q: 平台审核和模型审核是一回事吗？ A: 是三回事。模型拒答是训练对齐的结果，换个写法可能就不拒了，它不是执法；供应商审核在服务端，扫描请求与输出、依据使用政策处置账号，这才是封号的来源；平台审核是你用的 App 或社区自己加的闸，比如上架商店的内容规则、社区分级。三层各管各的，哪一层出手，后果完全不同。 Q: 用 Foreverse，App 会替我过滤对话内容吗？ A: BYOK 对话不过滤：卡的导入不做内容审查，你的请求从设备直连你选的供应商，实际边界由那家的政策和模型行为决定。要过滤的是另一头——社区分享出来的内容有 nsfw 分级，18+ 的卡默认不展示，看之前要过年龄确认和两步确认。私下创作和公开分享，我们分开管。 Q: 有没有内容是无论哪家、无论怎么设置都不该碰的？ A: 有，而且各家条款在这一点上措辞完全一致：任何涉及未成年人的性内容是绝对红线，包括虚构和扮演场景。Anthropic 和 OpenAI 都写明会向 NCMEC（美国失踪与受虐儿童中心）报告，xAI 同样声明报告义务。我们的内容政策把它列为第一条硬线，发现即上报平台安全渠道，没有讨论空间。 ### 「破限」到底是什么？一篇给外行的诚实科普：词源、机制与各家政策 URL: https://foreverse.app/zh/blog/jailbreak-culture-honest-take Published: 2026-07-18 Q: 破限到底是什么意思？ A: 社区黑话，指用特殊提示词绕开 AI 模型的内容安全策略，让模型输出平时会拒绝的内容。同族黑话还有「破甲」（甲指模型的安全防御）、「发牌子」（触发平台的安全提示卡片）。它对应英文社区的 jailbreak，词根借自 iPhone 越狱。 Q: 破限会不会封号？ A: 风险真实存在，且写在条款里。OpenAI 使用政策明确把「绕过我们的防护措施」列为禁止项，违反可能失去访问权限；Anthropic 使用政策把「故意绕过产品内的能力限制或防护栏」列入平台滥用条目，处置手段包括限流、暂停和终止账号；Google 的生成式 AI 禁用政策也把「规避滥用保护或安全过滤器」列为禁止（三家条款均于 2026-07-18 核实原文）。执行松紧各家不同，但风险归属没有歧义：账号是你的，责任也是你的。 Q: 为什么有的角色卡写着「必须配破限才能玩」？ A: 通常是卡的题材在作者调试用的那家模型上容易触发拒答，作者便把某个引导性预设当成这张卡的运行环境标配。这句话其实在告诉你：这张卡的题材和默认模型的政策不匹配。你有两个选择：换一家政策与题材匹配的模型，或者接受对抗路线连带的全部风险。我们建议前者。 Q: 破限和预设是什么关系？ A: 载体和用途的关系。预设是打包的提示词结构加采样参数，本身是中性的配置工具，管文风、节奏、扮演纪律的预设占大多数；社区说的破限预设，是把绕过安全策略的引导语写进了预设里的那一类。工具同一个，用途分了叉。预设机制本身怎么分层、怎么遮蔽，我们有一篇专文讲。 Q: 用 BYOK 应用，App 会不会替我挡内容？ A: 以 Foreverse 为例：导入不做内容过滤，你的对话从设备直连你选的供应商，App 不在中间加一道自己的闸。实际能聊什么，由 Key 背后那家供应商的政策和模型行为决定。这也意味着责任链条很干净：协议是你和供应商直接签的，条款约束你，选择权也在你。 ### 酒馆预设是什么？分层、遮蔽关系与挑选思路一篇讲明白 URL: https://foreverse.app/zh/blog/tavern-presets-explained Published: 2026-07-18 Q: 预设到底是什么东西？ A: 一份打包好的「怎么跟模型说话」配置文件：一半是采样参数（温度、输出长度这些旋钮），一半是提示词结构（系统指令怎么写、各模块按什么顺序拼进请求）。它管说话的方式，不管说话的人——人设、开场白、示例对话在角色卡里，预设不碰这些。 Q: 导入的预设会把我的角色卡顶掉吗？ A: 不会。预设和角色卡管的是两摊事：卡的人设内容照常注入，预设接管的是包裹这些内容的指令层和参数。真正会被顶掉的是同层的旧预设，以及你在全局设置里手写的系统提示词覆盖——预设整体接管系统提示位，两者不叠加，这是新手最容易误会的一处。 Q: 为什么我在全局选了预设，这张卡却没用上？ A: 生效顺序是四层：本次会话、当前角色、全局、内置默认，越贴近这局越优先，高层配了东西就盖住低层。这张卡大概率单独配过角色层预设，或者本次聊天应用过会话预设。想查就从会话层往下翻：本次聊天资料、当前角色设置、酒馆全局设置，哪层非空哪层说了算。 Q: 预设从哪找，新手该挑多大的？ A: 三个来源：中文社区流通的成套预设（QQ 群、论坛、网盘居多）、应用内社区页可直接逛和导入、以及桌面酒馆玩家导出的 JSON 文件。新手建议从小的开始：几十条以内、说明写得清楚的优先。上来就上两百多条的大部头，出了问题你分不清是卡的事、模型的事还是预设里哪一条的事。 Q: 预设在不同模型之间通用吗？ A: 格式通用，效果不通用。预设是标准 JSON，哪家模型都能导；但里面的指令是作者对着特定模型调出来的，换一家脾气不同的模型，防 AI 味的条目可能失灵，文风指令可能过火。社区预设一般会标注适配模型，跨模型用就当重新试一遍，别默认效果平移。 ### 酒馆群聊怎么玩才好玩？搭班子指南：请谁进群、怎么配戏才不冷场 URL: https://foreverse.app/zh/blog/group-chat-casting-guide Published: 2026-07-18 Q: 群聊放几个角色合适？ A: 三到四个是甜区。两个人是对手戏不是群像；五个往上每轮发言机会摊薄，性格再鲜明也轮不到开口，token 成本还按人头涨。想要大场面，可以多拉几个进群但暂时禁言一部分，需要谁出场再解禁——在场感和成本都能兼顾。 Q: 为什么我的群聊老是一个人刷屏？ A: 最常见的是三件事叠加：那个角色的话痨值（talkativeness）比其他人高一截，加权抽选回回是它；你的消息老点它的名字，点名优先直接把话递过去；它的回复又爱提别人的名字，点名接力继续绕回它的话题圈。修法对应着来：调低它的话痨值、开场多点别人的名、或干脆切手动指定策略当几轮导演。 Q: 高冷寡言的角色就不适合进群吗？ A: 适合，但不能全是。寡言角色的话痨值低，加权抽选下开口少，恰好符合人设——问题只出在全员寡言的班底：没人主动接戏，每轮都靠顺序兜底硬轮，出来的戏像被按头发言。留一两个高话痨的角色当发动机，冷角色的偶尔开口反而更有分量。 Q: 群聊比单聊烧钱吗？ A: 按发言算，每位角色开口一次就是一次模型调用，一轮引出两三个人接话就是两三次调用。另一头是角色卡注入方式：单独注入只带当前发言者的卡最省；合并全员把所有成员的卡都塞进每次请求，四张卡就是四份常驻 token。长剧情用单独注入、关键对手戏切合并，是比较省的开法。 Q: 在手机上怎么建群？ A: 角色页长按任意一张卡进入多选，勾满两张以上可开聊的卡，底部会出现建群条，点「新建群聊」即成。发言策略、成员启停、自动模式都在聊天页右上 ⋮ 菜单的「群聊设置」里，随时可改，改了立刻生效。 ### Quick Reply 自动化玩法：把每次都要敲的那句话做成一颗按钮 URL: https://foreverse.app/zh/blog/quick-reply-automation-guide Published: 2026-07-18 Q: 酒馆的 Quick Reply 是干嘛用的？ A: 把你反复要做的操作钉成输入框上方的一颗按钮。按钮内容是纯文本时，点一下就替你发出那句话（或填进输入框）；以 / 开头时就是一段 STscript 脚本，能做条件、循环、变量这类复杂机关。按钮还能挂自动时机，比如每次 AI 回复后自动执行，从「快捷键」升级成「自动化」。 Q: 有没有办法一键让 AI 继续、总结、换视角？ A: 有，这正是自动按钮的主场。总结：建一颗纯文本按钮，内容就是你每次都要敲的那句「暂停剧情，用三句话总结当前进展」，点一下即发。继续：脚本按钮写 /continue。换视角、换景：用 /setinput 把模板填进输入框，改两个词再发，比全文重敲快得多。三颗按钮各占一个位置，装完就是你的导演台。 Q: STscript 在手机上能跑吗？ A: 跑的是桌面同源的命令子集：变量（/setvar、/getvar、/incvar 等）、流程（/if、/while、/times）、发送与生成（/send、/sendas、/trigger、/continue）、输入框与弹窗（/setinput、/input、/buttons）、候选与群成员管理都在。输入框打一个 / 会弹出命令菜单点选补全；遇到暂不支持的命令会安全跳过，不执行也不动你的数据。完整清单在 设置 → 角色卡与酒馆 → 酒馆命令帮助。 Q: 自动化会不会失控刷屏、偷偷花钱？ A: 有三道闸。模式闸：自动化分四档（酒馆兼容/发送前确认/只填输入框/禁用），不放心就切「发送前确认」，每次自动发送前都先问你。循环闸：检测到疑似自动循环会自动把模式降到「发送前确认」并提示。账目闸：脚本本身不花钱，只有 /trigger、/continue 这类触发生成的步骤按正常对话计费；运行记录里能查每一步做了什么、有没有触发生成。 ### 酒馆正则入门：它改的是显示还是历史？第一次用先分清这个 URL: https://foreverse.app/zh/blog/regex-scripts-for-beginners Published: 2026-07-18 Q: 酒馆的正则脚本是干嘛的？ A: 自动执行的查找替换规则。每条规则盯着一类文本（用户输入、AI 回复、世界书等，可多选），命中就按替换串改写。玩家侧最常见的三个用途：把回复里的思考段折叠掉、去掉 (OOC:…) 这类旁注、把模型输出的纯文本 marker 替换成状态栏或主题化界面——美化包的核心机制就是最后这条。 Q: 导入的卡带正则，会不会乱改我的聊天记录？ A: 美化类正则不会：它们走「仅显示」时机，只在渲染时替换，聊天记录文件里存的还是模型的原话，HTML 模板也不占 token。会动记录的是「双向」时机的规则——替换发生在消息写盘之前，改动是永久的（比如把思考标签从记录里删干净就靠它）。在全局正则页里每条规则的时机都有标签，导入后扫一眼就知道谁会动记录。 Q: 正则美化手机上能用吗？和桌面酒馆一致吗？ A: 能用。Foreverse 的正则引擎按桌面酒馆的语义实现：作用位置（用户输入/AI 回复/世界书/Slash/思考链）、三档执行时机、$1 反向引用与 {{match}} 替换、按楼层深度限定生效范围，行为一致。桌面上调好的美化卡导入手机端直接生效，导入时还会自动打开该卡的富渲染开关。 Q: 正则写坏了会把聊天弄崩吗？ A: 不至于。最坏的情况是这条规则把显示文本替换成了空，渲染侧有兜底：清空时回退显示原文，不会白屏，也不会吞掉正文。修复路径是回全局正则页把它关掉或删掉；「仅显示」档的错误不碰记录文件，关掉规则一切如初。真正要小心的只有「双向」档——它写进记录的改动没有撤销键。 ### 世界书递归是什么？触发链、深度限制与一次失控案例 URL: https://foreverse.app/zh/blog/lorebook-recursion-explained Published: 2026-07-18 Q: 世界书条目会互相触发吗？ A: 会，前提是打开「递归扫描」（Foreverse 里默认是关的，开关在酒馆全局设置的「世界书与渲染」）。打开后，已命中词条的正文也参与关键词扫描，正文里点到谁的关键词，谁就跟着注入。同一词条一次生成至多激活一次，所以 A 提 B、B 又提 A 不会转圈。 Q: 世界书递归扫描是什么意思？ A: 指扫描源多了一层。普通触发只扫最近几轮对话找关键词；递归扫描把已命中词条的内容也当作扫描文本，让词条能链式召唤相关词条：聊到门派，门派档案里点名的佩剑也自动跟上。链条停在三个条件上：没有新命中、达到最大递归层数、token 预算耗尽。 Q: 为什么加了一条设定，好几条一起冒出来？ A: 大概率是新条目的正文里点了其它条目的关键词，或者被别的条目点名，递归把它们串成了一串。打开聊天页 ⋮ 菜单的「上次生成请求」看世界书明细，每条注入都标着命中关键词，链条一眼可见。不想被链子带出来的条目，在词条编辑器里勾「不参与递归扫描」。 Q: 最大递归层数设多少合适？ A: SillyTavern 官方文档的口径：设 1 等于关闭递归，2 允许链条走一跳，3 允许两跳，依此类推。日常两跳足够覆盖「门派带出佩剑、佩剑带出剑灵」这类两级脉络。层数是全局旋钮；个别词条的精细控制，交给「不参与递归扫描」「阻断后续递归」「仅递归时激活」三个词条级开关更划算。 ### 睡前听书方案：定时关闭、音色、缓存和不吵醒室友的细节 URL: https://foreverse.app/zh/blog/bedtime-listening-setup Published: 2026-07-18 Q: 睡前听小说用什么 App，能定时吗？ A: 挑应用盯三样：有睡眠定时、能提前缓存、进度和阅读打通。以 Foreverse 为例，阅读器底部控制栏点「听」进播放器，睡眠定时给 15、30、60、90 分钟四档，到点自动停；整章可以在 Wi-Fi 下提前缓存；耳朵听到哪，阅读进度就记到哪，第二天翻开书接着读。系统 TTS 引擎完全免费，先用它把习惯养起来。 Q: 听书定时关闭一般设多久合适？ A: 从 30 分钟档起步。多数人入睡在这个窗口内，播完自动停，不会整夜耗电也不会把书「听」过去几十章。躺下就困的人用 15 分钟档；习惯听着入睡又浅眠的人别选 90 分钟，声音持续的时间越长，越容易在浅睡期把你重新拉醒。定时到点停在哪一段，进度就记在哪一段，第二天不用倒回去找。 Q: 晚上听书哪个音色不吓人？ A: 两条经验：语速放慢，选温暖或低沉的声线。白天通勤用 1.5 倍速赶进度没问题，睡前建议降到 0.75 到 1 倍，语速本身就是刺激源。音色上避开冷峻快节奏的悬疑腔，言情种田配温暖女声、历史修仙配低沉男声的搭配深夜同样成立。系统 TTS 的默认音色偏平，反而适合催眠；在线 AI 音色更有起伏，音量记得比白天再调低一档。 Q: 半夜网络断了，听书会停吗？ A: 分引擎。系统 TTS 装了离线语音包就完全不依赖网络，飞行模式也能听。在线 AI 音色按分段合成，睡前在 Wi-Fi 下点「下载整章」把接下来的内容缓存好，已缓存的段落播放不需要网络；没缓存的段落会等有网时排队补合成。缓存还管钱包：合成结果按章节、音色、模型落盘，重听不重复扣费。 ### token 是什么？给只想看小说的人算一笔账 URL: https://foreverse.app/zh/blog/token-explained-for-readers Published: 2026-07-18 Q: AI 按 token 收费，到底是什么意思？ A: token 是模型读写文字的最小单位，直观理解成「字或词的碎片」：模型每读一个 token、每写一个 token 都记一次账，账单等于读写总量乘以单价。对读者的实际意义是，你付的钱跟「模型这次读了多少前文、写了多少新内容」成正比，跟你这本书总共多长无关，因为每次请求只送需要的部分进去。 Q: 一章 3000 字的小说大概是多少 token？ A: 看谁来数。我们拿一章 3065 字的真实网文实算：OpenAI 新版分词器数出约 2,800 个，旧版数出约 4,200 个；DeepSeek 官方给的换算口径是一个汉字约 0.6 个 token，即约 1,800 个。同一章书，各家能差出一倍多，所以别背精确数字，记住量级：一章几千个 token，实际扣费以每次请求返回的用量统计为准。 Q: 为什么说续写一段只要几分钱？ A: 一次续写，模型读进去的是相关前文加设定，一万个 token 上下，写出来一段三五百字。按 DeepSeek 2026 年 7 月牌价把这笔账换算掉：读的部分约一分钱，写的部分不到两厘，合计一分二厘左右。走 Foreverse 官方渠道的实测均值是 19 credits，约一分四厘。两个口径都指向同一个量级：一段续写是「分」级的开销。 Q: 有的模型「思考」也收费吗？ A: 收。推理型模型在给出正文前会先产出一段思考过程，这部分 token 按输出价计费，哪怕你在界面上看不到全文。同一段续写任务，思考型档位的实际消耗可能是普通档位的数倍，这是「单价看着便宜、账单却更贵」的常见原因。给这类模型设输出上限时也要把思考预算算进去，否则正文会被截断。 Q: 怎么查自己每次续写用了多少 token？ A: 不用信任何文章的估算，包括这篇。每次 API 调用的返回里都带用量统计字段，写着这次读了多少、写了多少、命中缓存多少。Foreverse 的「API 请求记录」（设置 → AI 模型与服务）把这些字段逐条摆出来，连续续写两次，对比第二次的缓存命中数，你会亲眼看到本文最后一节说的折扣。 ### 2026 年免费用 AI 续写小说的四条路，各自的天花板实测 URL: https://foreverse.app/zh/blog/free-models-for-continuation-2026 Published: 2026-07-18 Q: 有没有完全免费的 AI 续写？ A: 有，而且不止一条路。长期免费的是 BYOK 免费档：智谱 GLM-4.7-Flash 在开放平台官方标注免费调用，Google AI Studio 的 Flash 档免费层限次数不限期限（均为 2026 年 7 月核实）。一次性免费的有 Foreverse 注册送的 5000 积分（约 260 段）和各家新号体验金。完全离线的是本地 Ollama 跑开源模型。四条路都不要钱，差别在天花板的位置。 Q: 学生党没钱，想用 AI 写文，推荐哪条组合？ A: 日常推进接智谱 GLM-4.7-Flash（免费、注册 bigmodel.cn 拿 Key 即用），写到在意的章节切 Foreverse 注册送的 5000 积分用官方渠道的旗舰模型，两边随时并存。这套组合零投入，靠免费档扛量、送的额度保关键章节质量。硅基流动实名认证的 16 元代金券留着试模型：想知道某个付费旗舰值不值得充，先拿代金券写几段看效果。 Q: AI 写小说不充钱能用多久？ A: 分路径。BYOK 免费档没有总量截止日：智谱免费模型和 AI Studio 免费层都是长期政策（后者每天限次、太平洋时间午夜重置），只要政策不变就一直能写。一次性额度会到头：注册送的 5000 积分约 260 段、8 到 13 万字。本地 Ollama 理论上无限，电费不计。所以「用多久」的答案取决于你接受哪种天花板：限速、限量，还是限机能。 Q: 免费模型写出来的质量比付费的差多少？ A: 差距真实存在，但位置分场景。免费档模型（GLM-4.7-Flash、Gemini Flash 免费层）写日常推进、对话、过渡段够用；长剧情的伏笔管理和文风贴合度不如旗舰，这是我们在两场双盲横评里反复看到的档位差。务实的用法不是二选一：免费档跑量、旗舰保关键章节，在支持逐段换模型的阅读器里两档并用没有切换成本。 ### Grok 能写小说吗？复读实测、适用边界与什么时候别用它 URL: https://foreverse.app/zh/blog/grok-continue-novel-howto Published: 2026-07-18 · Updated: 2026-08-13 Q: grok 写小说怎么样，有人测过吗？ A: 测过。我们让九个大模型在两本中文书（传统玄幻与宫斗古言）上各连续续写 20 轮、累计 360 轮，双盲评审排名。Grok 4.5 在宫斗场被两个互不知情的评审一致排在末位：后段叙述与台词逐字复读，二十轮剧情停在案发当日下午；玄幻场中段也出现同两段文字逐字循环三遍。短程生成可用，长跑连续续写是它实测的雷区。 Q: grok api 国内能用吗？ A: 主要门槛在支付而不在技术：xAI 的充值不支持支付宝、微信和国内银行卡，只收海外信用卡（2026 年 7 月核实）；国内网络直连 api.x.ai 的稳定性也因网络环境而异。社区里常见的做法是经聚合中转平台间接接入，Foreverse 的 BYOK 预置目录里有整组「聚合与中转」供应商，多花一点差价换省事。 Q: grok 写文老是重复，是我的问题吗？ A: 不是。同一套提示、同一个实验协议下，多数模型没有这种高频发作，而 Grok 4.5 在玄幻和宫斗两种文体上都发作了，这是模型级的失效，不是提示词能根治的。你能做的是缓解：发现开始复读时换一个模型续下一段，或者把复读段删掉重写，循环就被打断了；以及别把几十段的长链一口气交给它。 Q: Grok 便宜的 fast 档能拿来写文吗？ A: 按 xAI 2026 年 7 月牌价，grok-4.1-fast 每百万 token 输入 0.2 美元、输出 0.5 美元，比旗舰低一个量级。但我们的横评只测了 grok-4.5，fast 档没有续写质量数据，这里不替它背书。想试的话用短程任务自己盲测：同一段开头各续几段，跟你常用的模型对比着读，再决定要不要把日常草稿交给它。 ### 智谱 GLM 续写小说：唯一双文体都进前三的模型，和它的半角引号顽疾 URL: https://foreverse.app/zh/blog/glm-continue-novel-howto Published: 2026-07-18 Q: 智谱清言写小说怎么样？ A: 智谱清言是面向普通用户的聊天助手，模型同源但容器不同：聊天框贴不下一本书，会话之间前情也不互通。想接着一本书往下写，走 API 路线更顺：在 bigmodel.cn 建 Key，填进 BYOK 阅读器。模型能力本身有名次可查：GLM 5.2 在我们的双盲横评里玄幻场第 3、宫斗场 2-3，唯一双文体都进前三。 Q: GLM 续写网文行不行？ A: 行，有实测：玄幻场（六模型双盲）它排第 3，评语「内容层贴近（战术商议、单字传讯都对味）」；宫斗场评语「限知观察质感最纯」。要带一个预期进场：它习惯用半角引号写对话，中文书里观感出戏，缓解靠提示词约束、对话密集章节逐段换模型，或对已生成段落编辑重生。 Q: 国产模型里哪个写文最稳？ A: 「稳」有两种读法。单场峰值看 DeepSeek 双档：玄幻第一是 V4 Flash，宫斗第一梯队是 V4 Pro；跨文体稳定性看 GLM 5.2，唯一双榜都进前三的模型。读单一文体的书，按那张榜的第一名挑最划算；文体杂、或者书还没定型，GLM 是更保险的开局牌。 Q: GLM 免费档和付费档差多少？ A: 免费档 GLM-4.7-Flash 官方长期免费、200K 上下文，推荐场景明确列了中文写作与角色扮演，日常推进和过渡段够用。但进两张榜前三的是付费旗舰 5.2，免费档没有盲评名次，两者的口碑别混着用。务实组合：免费档跑量，5.2 保关键章节，同一把 Key 里切换。 ### Kimi 续写小说实测：长上下文是真优势还是伪需求 URL: https://foreverse.app/zh/blog/kimi-continue-novel-howto Published: 2026-07-18 Q: Kimi 可以续写小说吗，效果如何？ A: 可以，有名次可查：九模型《甄嬛传》双盲横评里 Kimi K2.6 排 5-7 名，评语是「唯一显著逆向改善的系统：开局是网文暴怒腔，越写越收敛回宫斗正轨」；同一协议的玄幻链上它有轻症复读，整段照抄自己前几轮的产物。总体中游偏稳，长跑撞见似曾相识的整段就重摇一次或换个模型打断。 Q: Kimi 的长上下文对写文有用吗？ A: 分任务。整卷总结、全书人物梳理、世界书起草这类要一口气读完全书的活，256K 窗口是真优势；反复续写不是：我们的五档上下文对照实验里，把 27 万字原著整本塞入、不加指令，产出照样满篇套话比喻，密度是原著的 10 倍，而且每段续写都按全书输入计费。续写要的是按相关性挑出的那一小撮上下文。 Q: 月之暗面的 API 多少钱、怎么接？ A: K2.6 官方牌价（2026-07-18 核实）：输入未命中缓存 ¥6.5、命中 ¥1.1、输出 ¥27，单位每百万 token，预付费制，余额为零请求返回 402。在 platform.moonshot.cn 注册建 Key 并充值，填进 BYOK 阅读器（Foreverse 预置 Moonshot 条目，host 是 api.moonshot.cn/v1），用页内实测功能验证连通后，导入书长按「续这里」即可。 Q: Kimi 网页版免费，和 API 是一回事吗？ A: 两个账户体系。网页与 App 的 Kimi 助手面向普通用户，基础聊天免费（高峰时段可能限流，政策以官方页面为准）；API 在开放平台单独注册、预充值、按 token 计费，两边余额互不相通。把 Kimi 接进阅读器续写一本书，走的是 API 这条路，网页版的免费额度帮不上。 ### Gemini 续写小说实测：免费档能白嫖到什么程度，短板在哪 URL: https://foreverse.app/zh/blog/gemini-continue-novel-howto Published: 2026-07-18 Q: Gemini 免费的 API 可以拿来写小说吗？ A: 可以，两个边界内：免费档只覆盖 Flash 级模型（2026 年 4 月起 Pro 级归付费层），配额大约每分钟 10-15 次、每天数百到一千多次请求，按 Google Cloud 项目计而不是按 Key 计。一段续写等于一次请求，正常阅读节奏用不完每天的量，连续快速重摇才会撞每分钟限次。另外官方条款写明免费层的请求与产出可能被用于改进产品，喂未发表的手稿前值得停一秒。 Q: AI Studio 的 Key 怎么弄？ A: 用 Google 账号登录 aistudio.google.com，进 API keys 页点 Create API key，选一个新建或已有的 Google Cloud 项目，复制 AIza 开头的字符串，全程不用绑卡。和 OpenAI、Anthropic 不同，AI Studio 之后还能回平台查看 Key 明文。拿到后填进支持 BYOK 的阅读器（如 Foreverse 的模型供应商页），用页内实测功能点一次文本聊天验证连通。 Q: Gemini 写长篇会不会越写越差、老是从头重写？ A: 我们实测撞到的是急性病「从头重写」：Gemini 3.1 Pro 在玄幻续写里 20 轮全部跳回原著末尾重写开场，根因是上下文里一句「仅供参考」式的标注说明被它读成「这些段落不算正文」。把说明改成两句显式声明（已续写内容是已发生的正文、必须从最后一段继续；文风以原著为基准）后，两本书合计 28 轮零复发。手工喂上下文时写清这两句，就能避开这个坑。 Q: 免费档的 Gemini 和横评里排名的是同一个模型吗？ A: 不是。我们双盲横评排名的是付费层的 Gemini 3.1 Pro（《甄嬛传》场 4-6 名）；免费档只有 Flash 级模型，我们没有它的盲评名次，也不硬编。免费档适合当零成本试镜场：拿自己那本书的固定一段让它连续续写几轮，打乱顺序盲读，再决定要不要为付费档或别家模型花钱。 Q: 中国大陆能用 AI Studio 的免费档吗？ A: AI Studio 与 Gemini API 的官方可用地区名单（截至 2026-07）不含中国大陆，直连走不通，网络环境需自理且合规性自行判断。省事的替代有两条：Foreverse 官方渠道目录里有 Gemini 条目，登录按积分计费，注册送 5000 credits；或改用国内可直连的免费档，比如智谱 GLM-4.7-Flash。 ### Claude 怎么续写小说？从网页版长度天花板到 API 接进手机阅读器 URL: https://foreverse.app/zh/blog/claude-continue-novel-howto Published: 2026-07-18 Q: Claude 能接着我手机里的小说继续写吗？ A: 能，走 API 路线：在 console.anthropic.com 开户建 Key，填进支持 Anthropic 协议的 BYOK 阅读器（比如 Foreverse），导入 txt 或 epub 后长按选段点「续这里」。整本书的上下文由阅读器按相关性组装，不用每次手工喂前文；新内容落在分支上，原文一个字不动。 Q: 我充了 claude.ai 的 Pro，还要再花钱吗？ A: 要。Pro 订阅买的是网页与 App 的聊天用量，不含任何 API 额度；API 在 console.anthropic.com 单独开户、预充值、按 token 计费，同一个邮箱登两边但账单互不相通。按 2026 年 7 月牌价折算，Sonnet 4.6 一段续写约三毛人民币，Opus 4.8 约五毛。 Q: Claude 写中文小说效果怎么样，值得充吗？ A: 我们两场双盲横评里它都排第 4-5 名：对话机锋全场最锋利，但尾段半角标点混排、波动大，两场第一梯队都是别家。想要 AI 有自己笔感的场景（同人重构、机锋密集的对话戏）值得；追求「像原著续章」的无缝长跑，玄幻场第一是 DeepSeek V4 Flash、宫斗场第一梯队是 V4 Pro 和 GPT-5.6 Terra，都比它便宜得多。 Q: 中国大陆能用 Claude 续写小说吗？ A: Anthropic 官方支持地区名单（截至 2026-07）不含中国大陆，直接开官方 Key 走不通。合规的路有两条：用 Foreverse 官方渠道里的 Claude 条目（Sonnet 4.6 与 Opus 4.8），登录后按积分计费；或者接入你自己信得过的 Anthropic 兼容第三方端点，BYOK 支持自定义端点，但中转服务的合规性与稳定性要自行判断。 ### 我们把角色卡创作 Skill、47 份分类指南和 AI 味检测器全部开源了 URL: https://foreverse.app/zh/blog/open-sourcing-character-card-skills Published: 2026-07-17 Q: character-card-skills 仓库里到底有什么？ A: 两个 agent skill（character-card-author 写卡 + chat-quality-doctor 聊后调优）、47 份分类创作指南、8 份分模型 RP 症状档案、15 张原创卡（12 中 3 英，每张给 card.md 源文件、chara_card_v2 JSON 和 v3 PNG 嵌卡三种形态）、按真实读者判断校准的 AI 味规则检测器和它的 gold 回归集、一个零依赖的转卡脚本。代码 MIT，卡和文档 CC BY 4.0。 Q: 卡能直接导入酒馆（SillyTavern）吗？ A: 能。每张卡都预转换好了：v3 PNG 拖进酒馆的角色导入就能用，偏好 JSON 的前端用旁边的 v2 JSON。世界书、备选开场白、示例对话都嵌在卡里。我们在 SillyTavern 和 Foreverse 里各导入验证过一遍。 Q: skill 在哪些工具里能跑？ A: 所有认 SKILL.md 开放标准的 agent：Claude Code、Cursor、Codex、Gemini CLI，装法就是拷一个文件夹。在 Foreverse 安卓端这套 skill 是内置的——书架 agent 写完卡还能顺手评分、画封面、直接入库。 Q: 为什么用规则检测器而不是让大模型来评「像不像 AI」？ A: 因为两个我们都测过。双盲实验里 LLM 评委彼此 86% 意见一致，在人类共识锚点上的正确率却只有 12%——一致地把真人写的卡判成 AI。规则检测器按真实读者点名的病句校准，每张进仓的卡都要在 CI 里过它。它的诚实边界：结构完美、密度均匀的模板化 AI 文案，它也抓不到。 ### We Open-Sourced Our Character Card Authoring Skills, 47 Genre Playbooks and the AI-Flavor Detector URL: https://foreverse.app/blog/open-sourcing-character-card-skills Published: 2026-07-17 Q: What exactly is in the character-card-skills repo? A: Two agent skills (character-card-author and chat-quality-doctor), 47 genre-specific writing playbooks, 8 per-model roleplay symptom profiles, 15 original character cards (12 Chinese, 3 English) shipped as card.md source plus chara_card_v2 JSON plus v3 PNG-embedded cards, a rule-based AI-flavor detector with its gold regression set, and a zero-dependency md-to-card converter. Code is MIT; cards and docs are CC BY 4.0. Q: Do the cards work in SillyTavern? A: Yes. Every card ships pre-converted: a v3 PNG you can drag into SillyTavern's character import, and a v2 JSON for frontends that prefer it. Lorebooks, alternate greetings and example dialogues are embedded. We test imports in SillyTavern and in Foreverse. Q: Which coding agents can run the skills? A: Anything that reads the open SKILL.md convention: Claude Code, Cursor, Codex, Gemini CLI. Install is a folder copy. The same skills also run natively inside Foreverse on Android, where the agent can score, illustrate and import the card it just wrote. Q: Why trust a rule-based detector over an LLM judge? A: Because we measured both. In a double-blind panel our LLM judges agreed with each other 86% of the time and were right 12% of the time on human-consensus anchors — they systematically rated human-written cards as AI. The rule-based detector is calibrated against what real readers actually flagged, and it gates every card in the repo through CI. Its honest limit: perfectly templated AI copy still passes. ### Continue Sherlock Holmes With AI: The Pastiche Tradition Goes Personal URL: https://foreverse.app/blog/continue-sherlock-holmes-with-ai Published: 2026-07-17 Q: Is Sherlock Holmes really public domain now? A: Yes. In the US, 50 of the 60 canonical stories had already aged into the public domain by 2013, and the Klinger v. Conan Doyle Estate litigation (decided 2013, affirmed by the Seventh Circuit in 2014) confirmed the Holmes and Watson characters were free to use. The last ten stories, collected in The Case-Book of Sherlock Holmes (1927), expired one by one until January 1, 2023, when the final two entered the US public domain. In the UK and other life-plus-70 countries, copyright ended decades earlier — Conan Doyle died in 1930. What remains off-limits is implying endorsement by the estate, which is a trademark matter, not copyright. Q: Will the AI actually sound like Conan Doyle? A: Honest answer: it can hold the register — formal Victorian narration, Watson's clinical warmth, period vocabulary — but surface mimicry is not the same as the real thing, and our own evals keep proving it. In 360 rounds of continuation testing we saw the best prose mimic in the field reproduce an author's comma rhythm almost perfectly, then rewind the entire plot in the final rounds. Expect a competent impersonation that needs your editing, not a resurrection. Q: Can I publish my Holmes pastiche? A: The copyright side is clear: the original canon is public domain, so continuation and publication are lawful in ways that fanfiction of in-copyright works is not — an entire commercial pastiche industry already exists. Two practical caveats: don't imply official endorsement (trademark territory), and if AI wrote part of it, disclose that wherever you publish — most platforms now have explicit AI-content rules and quietly breaking them gets works removed. Not legal advice. Q: Which model should I use for Victorian-era prose? A: We haven't run a Victorian-English arena yet, so we won't invent a ranking. What our Chinese-fiction arenas did establish transfers as a method: no ranking survives a genre change (our fast-plain-prose champion fell to sixth on an ornate court novel), so test two or three models on the same scene from your actual book and compare. In Foreverse you can switch models per scene with your own keys — 62 providers — so the test costs a few minutes, not a subscription. ### Story Continuation Prompts That Survive Long Runs: Field Notes and Four Copy-Ready Templates URL: https://foreverse.app/blog/story-continuation-prompting-guide Published: 2026-07-17 Q: Can a prompt fully clone the original author's style? A: No — close, not perfect, and the gap is measurable. In our 20-round benchmark, no model matched the original's sentence-length burstiness: the source book's coefficient of variation was 0.84 and all six models landed between 0.46 and 0.59. Better context labeling reversed the decay direction in follow-up runs — one arm climbed from 0.56 back to 0.64 across rounds instead of sliding — but nothing we tried closed the gap fully. A good template plus source labeling plus the right model gets you to 'stays immersive on a straight read-through'; anyone claiming pixel-perfect style cloning from a prompt is overselling what prompts do. Q: How long should the excerpt I paste be? A: Enough to carry the voice, not the whole book. Style lives in the material layer, so a few thousand words of the most recent text — the run-up to your continuation point — outperforms a plot summary of everything. Our benchmark chains ran on a fixed 16k-token window (roughly 12,000 English words) and that was sufficient context for the best pairings to read like the original. In a reader app that assembles context automatically you can skip this question; pasting by hand into a chatbot, favor the last several chapters over scattered highlights. Q: Why do these templates ban phrases instead of describing the style? A: Because negative constraints are checkable and adjectives are not. A model cannot reliably act on 'write with more tension', but 'do not have the narrator announce danger before the character sees it' is a rule it can follow sentence by sentence — and that you can verify on read-through. The same logic applies to cliché lists: naming the exact constructions you keep seeing ('breath caught', 'time seemed to slow') removes them far more reliably than asking for 'fresh prose'. Q: Do these templates work in any AI tool? A: Yes — they are plain instructions with no app-specific syntax, so they paste into any chatbot or writing tool. Two adjustments travel with them: supply the material layer yourself (recent story text, plus a line stating that earlier AI continuations are canon), since a raw chat window assembles no context for you; and re-send or re-reference the template on each round, because instruction adherence fades as the conversation grows. ### The Author Dropped the Novel Two Years Ago. Here's How I Finished It for Myself With AI URL: https://foreverse.app/blog/finish-a-dropped-novel-with-ai Published: 2026-07-17 · Updated: 2026-07-19 Q: Will the AI continuation match the author's voice? A: Close enough to stay immersed, not perfectly, and anyone promising perfect is selling something. In our 20-round blind-review benchmarks, the best model-book pairings read 'like a chapter of the original' to reviewers, while wrong pairings were obvious within a page — and which model is best flips completely between genres. Two practical levers matter most: pick the model per book by auditioning two or three on the same passage, and keep continuations anchored to original prose rather than letting the AI imitate its own output for dozens of rounds. Q: What if the author comes back and posts real chapters? A: That is the scenario the branch structure exists for. Your continuation lives on its own line; the imported original stays exactly as published. When real chapters appear, you read them from the canon line as if your branch never happened — and your branch survives as its own alternate ending, which is occasionally the better one. Nothing needs undoing, because nothing was overwritten. Q: Does the original file get modified? A: No. The imported text is never edited by continuation — branches grow alongside it, and the original remains byte-for-byte what you imported. Selecting a sentence in chapter 209 and continuing from there does not touch chapter 210; it opens a parallel line starting at that point. Your library also stays local files on your device, so nothing about this workflow depends on a server keeping your copy. Q: Is it legal to do this for personal use? A: Private, non-distributed personal use and publishing are different risk classes. Every litigated unauthorized-sequel case we could find involved publication or commercial exploitation; none involved a private continuation kept on the reader's own device. The moment you post or monetize, the analysis changes completely. We wrote a full explainer on the US fair use factors and the sequel case law — and the standard disclaimer applies: background information, not legal advice. ### What AI Story Continuation Costs: No Subscription, Three Routes, the Actual Math URL: https://foreverse.app/blog/what-ai-continuation-costs Published: 2026-07-17 Q: Is there a monthly fee? A: No. There is no subscription anywhere in the stack: importing books, reading, and branch management are free and not rate-limited, and generation is paid per use. Credits are prepaid — you buy a pack when you choose to, nothing renews on its own, and no billing cycle sits under your library. The only recurring bill you can opt into is the one you create yourself by topping up a BYOK provider account. Q: What happens when free credits run out? A: The app keeps working; only official-channel generation pauses until you either buy a pack or switch that request to a BYOK key. Nothing about your library changes — books, branches, and history are files on your device. The two payment routes coexist per request, so a common pattern is spending the 5,000 starter credits (about 260 continuations) on the official channel, then moving daily volume to a cheap BYOK model and keeping credits for occasional convenience. Q: Is BYOK actually cheaper? A: Per token, always — you pay the provider's list price with zero markup, versus list price plus 50% on the official channel. Whether that matters depends on volume. At DeepSeek's official rates, a full 200,000-word continuation ride costs roughly $0.80 via BYOK against about $1.10 in credits: a 30-cent difference most readers should ignore, since BYOK also means registering a provider account, topping it up, and babysitting a key. Write daily across multiple books and the percentages start compounding into real money; that is the point where BYOK earns its setup cost. Q: Do these numbers cover images, audio, or video? A: No — this ledger prices text continuation only. Scene illustrations, TTS narration, and video generation bill separately at each model's own rate, and they are substantially more expensive per action than text. The 19-credit figure is a measured average for one text continuation; long contexts push individual requests above it, short ones below. ### Is It Legal to Continue Someone Else's Novel With AI for Personal Use? Fair Use, the Sequel Cases, and the Private-Use Line URL: https://foreverse.app/blog/ai-continuation-copyright-fair-use Published: 2026-07-17 Q: Can I get sued for a private AI continuation I never share? A: Anyone can file a lawsuit over almost anything, so no honest answer is 'zero risk'. But we could not find a single reported US case in which a rightsholder sued a reader over a private, non-distributed continuation — the litigated sequel disputes (Salinger v. Colting, Anderson v. Stallone) all involve publication or an attempt to exploit the work. A purely private use also gives a plaintiff no way to discover the work exists and no market harm to point at under the fourth fair use factor. Untested is not the same as blessed, but the practical exposure is very different from posting. Q: What changes if I post my continuation to AO3? A: The risk class changes entirely: posting is public distribution, and every sequel case that reached a courtroom lived on that side of the line. Platform policy is a separate question from copyright — AO3's Terms of Service currently do not prohibit AI-assisted fanworks that otherwise qualify as fanworks, but an archive's permission is not the rightsholder's permission. Noncommercial, genuinely transformative fanwork has a real fair use argument under Campbell v. Acuff-Rose, and it remains an argument, not a settled rule. If you post: disclose the AI involvement, never monetize, and accept that a takedown request can end the discussion. Q: Who owns the AI-generated part of the continuation? A: Two layers stack against you. First, the US Copyright Office's January 2025 copyrightability report says purely AI-generated material is not copyrightable, and prompts alone are not enough human authorship — protection attaches only to identifiable human contributions such as creative selection, arrangement, or rewriting, judged case by case. Second, a continuation is a derivative work: under 17 U.S.C. §103(a), as applied in Anderson v. Stallone, the parts of an unauthorized derivative that use the original's protected material unlawfully get no copyright protection at all. In practice you may own very little, and what you do own does not come with the right to publish it. Q: Does it matter where my copy of the book came from? A: Yes, and the AI training litigation just made the point loudly. In Bartz v. Anthropic, the court treated training on lawfully purchased books as fair use while treating the retention of millions of pirated copies as infringement — acquisition was analyzed as its own act. The same logic reaches individuals: feeding a bought ebook to a model for a private continuation starts from clean facts; a pirated file plants an independent infringement underneath everything you do afterward. Buy the book. ### Import Character Cards by URL: chub, JanitorAI, and Four More Sources Tested URL: https://foreverse.app/blog/import-character-cards-by-url Published: 2026-07-17 Q: How do I import a JanitorAI card to my phone? A: Open the character's page on Janitor, copy the address-bar URL, and paste it into Foreverse's import-from-link box — jannyai mirror links work identically. One caveat: Janitor's download route sits behind a bot check that blocks some networks. The app tells you when that happens, and the remedy is retrying on a different network — mobile data instead of Wi-Fi, say. Q: Do lorebooks come along with the card? A: Two cases. A character book embedded in the card travels with it from any site — import the card, get the book. Standalone lorebooks are link-importable from chub only: paste a chub lorebooks URL and it arrives as a proper lorebook. Standalone books hosted elsewhere still need the download-file-then-import route. Q: What are the most common reasons a link import fails? A: In the order we hit them while testing: the link isn't a card detail page (search results and profile pages don't resolve); the card is private or deleted; a bot check intercepted the request, fixable by changing networks; the RisuRealm author disabled downloads, which nothing on our side can override; the file exceeds the 32MB cap. Each case gets its own message — never a bare "failed." Q: Does an imported card get uploaded to Foreverse servers? A: No. A link import is your phone fetching the file directly from the card site; the result is stored as a local file on the device. Foreverse servers never see the request or the card. Chatting, editing, and re-exporting all happen on your phone afterward. ### Where Your Novels Actually Live: A Data-Boundary Walkthrough URL: https://foreverse.app/blog/where-your-novels-actually-live Published: 2026-07-17 Q: Does my BYOK key ever reach Foreverse servers? A: No. The key is encrypted and stored on your phone only — on Android through the hardware-backed Keystore system. Requests travel from your device straight to the provider you configured. Our servers are not on that path, and there is no endpoint that accepts user keys. Q: Does the official channel use my books or chats to train models? A: No. Official-channel requests are relayed through the Foreverse backend to the model upstream — that hop is required for metered billing — but your chat and book content is not used to train models. What the server records for billing is billing-side data: which model, how many tokens, when. Q: How do I move everything to a new phone? A: By hand, with exports. Chats, character cards, and lorebooks pack into a zip you import on the new device; books come across by re-importing the files or copying your backups. There is no automatic cloud sync yet. Credits and purchases live on your account and return when you sign in. Q: Is anything left after I uninstall? A: On the device, no — uninstalling deletes all local data, and since we hold no server-side copy, nothing can be recovered afterward. Export a zip first if you want to keep your chats and cards. Your account itself is deleted separately at /account/delete. Q: Can I check what the app has been sending? A: Yes, from the request log in settings: every AI call is recorded locally with timestamp, model, and token usage. Cross-check it against your provider's billing dashboard — if the two disagree, treat that as a bug and tell us. ### Which LLM Continues a Novel Best? Nine Models, Two Genres, Blind-Judged URL: https://foreverse.app/blog/best-model-for-continuing-novels Published: 2026-07-17 · Updated: 2026-08-13 Q: Which free or cheap model holds up for continuation? A: We didn't benchmark free tiers, so no ranking there. The closest data point: DeepSeek V4 Flash won the fantasy run outright while sitting in one of the lowest price tiers of the field — $0.14 per million input tokens at the official rate. Foreverse grants 5,000 credits on signup, roughly 260 continuation segments on the official channel, which is enough to try the top of both tables against your own book before paying anyone. Q: Models update constantly — is this ranking still valid? A: Placements are tied to the versions we tested in July 2026: DeepSeek V4 Flash and Pro, Claude Opus 4.8, Gemini 3.1 Pro, Qwen 3.7 Max, GLM 5.2, GPT-5.6 Terra, Grok 4.5, Kimi K2.6. Kimi K3, Qwen 3.8 Max Preview, Grok 4.6 and DeepSeek V4 Pro's August 13 GA build have since been retested head-to-head under the same protocol — see the update sections in the article. One caveat gained teeth in August: DeepSeek hot-swaps new weights under unchanged model IDs, so the DeepSeek placements describe preview-era snapshots you can no longer call. New versions can reshuffle names, and this page gets updated when we re-run — check the date at the top. What outlasts any version bump is the method: pick the model per genre, and use a tool that lets you switch mid-book. Q: Should English and Chinese novels use different models? A: Both benchmark books are Chinese (a fantasy epic and a palace-intrigue classic), so we have no English-novel ranking and won't invent one. What the data does support: switching genre within one language was enough to knock each champion to mid-table, so genre matters before language does. For an English book, replicate the protocol at small scale — same opening passage, a few candidate models, several rounds each, then read the outputs without knowing which is which. Q: How does the double-blind review actually work? A: Each model's output is anonymized under random letters, with two different random mappings, given to two reviewers who don't know each other exists. Only placements that agree across both mappings count as settled; anything the reviewers flag as low-confidence or disagree on is reported as a range. In the fantasy run, the top three and last place matched exactly across both mappings; on the palace novel, the top and bottom tiers matched. ### 和 AI 谈恋爱正常吗？一个做 AI 伴侣功能的团队的立场 URL: https://foreverse.app/zh/blog/loving-an-ai-our-take Published: 2026-07-17 Q: 和 AI 恋爱是逃避现实吗？ A: 我们不替任何人下这个结论，两面都摆出来。一面：对社交耗竭、异地、失去伴侣的人，一段可控的低压力关系是真实的缓冲，Common Sense Media 2025 年的调查里 80% 用过 AI 伴侣的青少年花在真人朋友身上的时间仍然更多，「用了就沉迷」不符合多数人的数据。另一面：如果 AI 是你唯一的情感支柱，线下的关系在持续萎缩，那不是产品的成功，是需要向真人求助的信号。工具的健康度取决于它在你生活里的位置，这个位置只有你自己看得见。 Q: AI 真的「记得」我吗？ A: 机制上：对话里的重要事实会被提取进记忆库，长期保存，每次对话注入。它不是人类那种带情绪温度的回忆，是检索。但我们不因此说这份体验是假的：TA 在三个月后提起你说过的那句话，你那一刻的感受是真实发生的。两件事同时成立。在 Foreverse 里记忆逐条可查、可改、可删，记错了直接改记录，不需要在对话里反复纠正。 Q: 和 AI 聊天算精神出轨吗？ A: 这是伴侣之间的约定问题，不是产品能替答的问题。有人的伴侣把它当游戏，有人当树洞，有人确实投入了排他性的感情，三种情况的答案不一样。我们能给的只有一个观察：瞒着比用着更伤关系。如果这段使用需要对伴侣保密，值得先想清楚为什么。 Q: 你们靠这个赚钱，立场可信吗？ A: 合理的怀疑，回答留给可验证的事实：我们按量计费（credits）或让你自带 API Key 直连供应商，不靠订阅时长赚钱，所以没有把你留在应用里的营收动机；主动消息默认关闭，开启要你自己动手，频率和免打扰时段可调；记忆和聊天记录存你手机本地，删号走人不需要我们批准。商业模式决定动机，这几条都可以在产品里当场验证。 ### 用 AI 续写《红楼梦》会怎样？从两百年续书公案，到你自己的第八十一回 URL: https://foreverse.app/zh/blog/continue-hongloumeng-with-ai Published: 2026-07-17 Q: AI 能写出曹雪芹的文笔吗？ A: 接近纹理可以，复刻魂魄做不到，这是实测过的诚实答案。我们在《甄嬛传》九模型续写横评（同属典雅书面语＋严密称谓礼数的路数）里看到：第一梯队能稳住语体和礼数，拿到「全程零事故」这样的工艺评语，但没有评审说过任何一家「像原作者」；整个横评系列里最高的赞语也只到「像一个笔更细的同题材作者」。所有模型还共享一个指纹：句子长度越写越均匀，学不会原著短句长句剧烈交错的节奏。把预期设在「工整的仿写」，惊喜留给个别段落。 Q: 公版《红楼梦》全本去哪里下载？ A: 文本本身早已公版：曹雪芹逝世约 260 年，程高刻本问世 230 多年。中国哲学书电子化计划（ctext.org）收录一百二十回全文，维基文库有 1791 年程甲本的整理页，都可免费在线读。注意两点：脂评本系统只到第八十回，从哪续决定你要哪个本子；市面校注本里的注释、校记、标点是当代整理者的劳动，有独立权益，拿公开电子文本当底本最干净。整理成 txt 或 epub 即可导入手机阅读器。 Q: 用 AI 续写《红楼梦》犯法吗？续出来能发表吗？ A: 原著公版，续写本身没有版权障碍，这一点比续写当代网文干净得多。只写给自己看、存在自己设备里，是最省心的形态。要公开发表则多一层义务：2025 年 9 月 1 日起施行的《人工智能生成合成内容标识办法》要求 AI 生成内容亮明身份，各发布平台也有各自的 AI 内容标注规则，发之前查清楚目标平台的政策。本文不构成法律意见。 Q: 从第八十回续写，AI 会不会自动写成程高本的情节？ A: 很可能，要主动防。模型训练语料里流通最广的是一百二十回程高本，从第八十回末往下写，它大概率会滑向掉包计、黛玉焚稿这些「既定结局」。这种记忆惯性是实测过的现象：甄嬛传横评里就有模型写出电视剧版专名（小说里太后住颐宁宫，它写成剧版的寿康宫），暴露训练记忆的来源。想走自己的线，在续写指令里显式声明「不沿用后四十回情节」，并把你要的方向写明。 ### 一条链接把角色卡搬进手机：chub、JanitorAI 等六源导入实测 URL: https://foreverse.app/zh/blog/import-cards-by-url Published: 2026-07-17 Q: JanitorAI 的卡怎么导进手机？ A: 在 Janitor 打开角色页，复制浏览器地址栏的链接，回到 Foreverse 的「从链接导入」粘贴即可，jannyai 镜像站的链接同样认。注意：部分网络环境会被站点的人机验证拦下，App 会明确提示；对策是换个网络重试，比如从 Wi-Fi 切到手机流量。 Q: 世界书会跟卡一起进来吗？ A: 分两种。卡内嵌的角色书（character book）随卡走，哪个站都一样，导卡即得；独立发布的世界书目前只有 chub 支持链接导入，粘贴它的 lorebooks 链接，进来就是一本标准世界书。其它站的独立世界书还是要下载文件再从文件导入。 Q: 链接导入失败最常见的原因是什么？ A: 按实测频率排：一，粘的不是卡片详情页链接（搜索页、作者主页都不行）；二，卡是私密或已删除；三，站点的人机验证拦截了这次请求，换网络可解；四，RisuRealm 的作者关闭了下载，这个无解，只能在原站用；五，文件超过 32MB 上限。五种情况 App 都会给出对应提示，不会只甩一句「失败」。 Q: 导进来的卡会上传到你们服务器吗？ A: 不会。链接导入是你的手机直接向卡站请求文件，拿到后存成本机文件，Foreverse 服务器全程不参与，也不会收到这张卡。之后开聊、编辑、再导出，都在你的设备上发生。 ### 你的小说存在哪里？Foreverse 数据边界一页说清 URL: https://foreverse.app/zh/blog/where-your-novels-live Published: 2026-07-17 Q: BYOK 的 Key 会传到你们服务器吗？ A: 不会。Key 加密后只存在你的手机上，Android 走 Keystore 硬件背书的加密；请求从设备直连你选的供应商，Foreverse 服务器不在请求路径上，也没有接收用户 Key 的接口。 Q: 官方渠道会拿我的书训练模型吗？ A: 不会。官方渠道的请求经 Foreverse 后端转发到模型上游，这是计费的必要路径；你的聊天和书的内容不用于训练模型，服务端为计费记录的是账单侧数据：用了哪个模型、消耗多少、什么时候。 Q: 换手机怎么把数据搬过去？ A: 手动导出。聊天记录、角色卡、世界书可以整包导出成 zip，新设备导入还原；书重新导入原文件或拷贝备份即可。目前没有自动云同步，搬家是自己动手的操作。账号上的积分和权益跟账号走，登录即回。 Q: 卸载 App 后数据还在吗？ A: 不在。全部本机数据随卸载删除，我们服务器上没有副本，无法代为找回。想保留聊天和卡，卸载前先做一次 zip 导出；账号本身的删除在 /account/delete 单独发起。 ### AI 写小说软件排行看不懂？先分清作者向和读者向两个物种 URL: https://foreverse.app/zh/blog/reader-tools-vs-writer-tools Published: 2026-07-17 Q: 我想自己写原创小说、目标是签约发表，该用哪派？ A: 作者向，别犹豫。按需求分流：要大纲模板和量产提效，看笔灵这类全场景工作台；写超长篇最怕设定失忆，看岱宗、蛙趣这类做项目级记忆和世界观管理的；最头疼伏笔挖坑忘填，解忧笔札把这件事做成了主打。读者向工具没有大纲、伏笔、投稿这些生产环节，扛不动日更。 Q: 读者向工具能拿来写原创吗？ A: 能写，不顺手。你可以从一页空白开始让 AI 陪你写，分支结构也适合试不同走向；但它没有三层大纲、人物关系图谱、伏笔回收提醒这些为「生产一部作品」准备的脚手架。写给自己看的自娱创作没问题，奔着完本签约去的，作者向工具更称手。 Q: 作者向工具能反过来当阅读器、给别人的书续写吗？ A: 形态不对。它们的「导入小说」多是拆书分析：提取句式、节奏、桥段进素材库，服务你自己的创作，并不为「舒服地读完一本两百万字的书」准备，也没有阅读进度、排版、听书这些阅读器的基本盘。给读过的书续一条 if 线，读者向工具是为这个场景造的。 Q: 「AI 写小说软件排行」哪个榜单可信？ A: 先看它有没有分物种。把日更产能工具和阅读续写工具放进同一张榜单打总分，等于拿卡车和家用车比谁好，评测维度都对不上。本文刻意不打分不排名，只陈述各家可核实的定位；分清自己是哪派之后，你需要比较的对象会立刻少一半。 ### 怎么让 AI 写同人不 OOC？把人设崩拆成三种病因，各有各的治法 URL: https://foreverse.app/zh/blog/fanfic-anti-ooc-method Published: 2026-07-17 Q: 原著 txt 直接整本喂给 AI 行吗？ A: 不划算，效果也差。整本书塞进上下文，每次生成都为几十上百万字的无关内容付费；更关键的是我们实测过，把 27 万字原著整个塞进上下文、不加任何指令，模型照样写出比喻密度十倍于原著的 AI 腔。上下文只提供可模仿的材料，「以谁为基准」「他是什么人」必须结构化地显式交代，模型不会自己猜对。 Q: 要写多少设定才够？ A: 比想象的少。角色卡核心人设几百字加 3 到 5 段原作示例对话就够锁语气；世界书条目一条管一件事，电报体 150 到 300 字，常驻条目控制在 5 条以内。够不够别数字数，看命中率：写到谁的时候，谁的档案有没有被递到模型眼前。写满两千字的编年史不如拆成十条各管一段。 Q: 哪个模型写同人最不容易 OOC？ A: 没整理人设之前换什么模型都白搭，这三味药比选型优先。药都吃上之后再谈偏好：我们的双盲横评里 Claude 和 Qwen 这类模型笔感更自我，贴原著文风不如 DeepSeek 系，但你若本来就想让同人有自己的笔调，它们反而合适。按「无缝续写」还是「另起笔调」来挑。 Q: 判断 OOC 有没有客观标准？ A: 没有，鉴定权在你。工具能做的是把判断成本降下来：重要情节让 AI 一次给多个候选并排挑；每个方向落在独立分支上，写崩了砍掉那条分支，原文和其它分支不受影响。「他不会这么做」这句话，只有把原作读进心里的人说了算。 ### DeepSeek 怎么续写小说？从网页版贴不下，到 API 接进手机阅读器 URL: https://foreverse.app/zh/blog/deepseek-continue-novel-howto Published: 2026-07-17 · Updated: 2026-08-13 Q: DeepSeek 的免费额度能用来续写吗？ A: 能，但别按传言做预算。官方扣费规则明确有「赠送余额」机制且优先扣减；具体送多少、有效期多长随运营活动变化，2026 年上半年的教程里流传过一百万到五百万 tokens 的不同说法。可靠的做法是注册后打开开放平台的用量页看实际数字，赠送余额用完且没充值时，API 调用会返回 402。 Q: 手机上能直连 DeepSeek 吗？要不要挂代理？ A: 不用代理。api.deepseek.com 在国内网络直连可用，注册开放平台也不需要翻墙。BYOK 应用的请求从手机直发供应商域名，中间不经过任何第三方服务器；在 Foreverse 里 Key 以 AES 加密只存本机，不上传、不进日志。 Q: 自带 DeepSeek Key 和 Foreverse 官方渠道积分，哪个划算？ A: 看用量。官方渠道是模型牌价加 50% 服务费，换来的是不用注册供应商、开箱即写，注册送 5000 credits 约合 260 段续写，轻度用户可能一直用不完。月消耗到几十元量级就值得切 BYOK：按 DeepSeek 官方牌价直付，零加价，同一个 App 里两条路随时并存。 Q: 教程里写的 deepseek-chat 模型名怎么不能用了？ A: 官方定价页已公告：deepseek-chat 与 deepseek-reasoner 两个旧模型名于北京时间 2026 年 7 月 24 日 23:59 弃用，二者分别对应 deepseek-v4-flash 的非思考与思考模式。新配置直接填 deepseek-v4-flash 或 deepseek-v4-pro，旧教程照抄模型名会在弃用后报错。 ### 乙女游戏玩腻了？想要不抽卡、剧情自由的纸片人恋爱，2026 年有这四派 URL: https://foreverse.app/zh/blog/beyond-otome-free-roleplay Published: 2026-07-17 Q: 没玩过酒馆的乙女玩家从哪开始？ A: 从一张评价好的女性向角色卡开始，不要从研究配置开始。装一个原生酒馆类 App，在社区里挑张顺眼的卡导入，用注册送的积分直接开聊；跑通了、确认这个玩法对胃口，再考虑自带 API Key、挑模型这些进阶项。概念扫盲（卡、世界书都是什么）我们写过一篇新手指南，二十分钟能读完。 Q: 能有语音吗？ A: 有，但形态和乙游不同：AI 这边是语音合成（TTS）音色，你给角色挑一个声线，消息可以合成语音播出，也能打语音通话，按对话轮数计费、挂断即停。把期望调成「有声音、有情绪」，体验不差；拿人气声优棚录的标准来要求，会失望。 Q: 会不会像乙游一样越玩越贵？ A: 花钱结构不同：没有卡池、保底和体力，付费是按量的模型调用费。自带 API Key 按供应商原价直连；官方渠道 1 积分 = $0.0001，注册送 5000 积分。重度玩语音、生图当然也花钱，但每一笔都记在请求记录里，用途用量可查——没有「为了歪的那几发」付的钱。 Q: AI 演的角色会崩人设吗？ A: 会发生，这点要诚实说。表现好坏取决于卡的质量和模型的能力，长对话里也可能慢慢漂。区别在处置权：卡文件能改、模型能换、聊崩的位置能回溯重来。在乙游里遇到不喜欢的剧情走向，你连一个能动手的入口都没有。 ### 酒馆是什么？SillyTavern 新手入门：角色卡、世界书、第一次开聊一篇讲明白 URL: https://foreverse.app/zh/blog/tavern-beginners-guide Published: 2026-07-17 · Updated: 2026-10-01 Q: 不用电脑真的能玩酒馆吗？ A: 能。原版 SillyTavern 要在电脑（或服务器）上跑 Node.js 服务，但 2025 年之后出现了一批原生酒馆类 App，把部署这一步整个干掉：装完即用，角色卡、世界书照常导入，数据存手机本机。想要原版全量扩展生态的人仍会选电脑或云端部署，只想导卡开聊的人手机就够了。 Q: 角色卡从哪里找？ A: 三个来源：社区卡站（chub.ai 是最大的英文站之一，中文卡在各社区流通）；应用内置的社区页，逛到即导；自己写——人设、开场白、示例对话三样凑齐就是一张卡。注意卡是 PNG 或 JSON 文件，从微信这类聊天软件接收 PNG 卡时要以「文件」方式发送，图片消息会被压缩转码，卡数据会丢。 Q: 玩酒馆要花多少钱？ A: 软件本身免费，钱花在模型调用上。自带 API Key（BYOK）按各供应商原价直连，轻度玩家一个月几块钱量级；不想配 Key 的用应用官方渠道按量计费，Foreverse 的官方渠道 1 积分 = $0.0001，注册送 5000 积分。角色卡本身在社区大多免费流通。 Q: 酒馆和星野、猫箱那类 AI 聊天 App 有什么区别？ A: 一句话：平台养成和自带资产的区别。星野、猫箱开箱即玩，语音立绘都是现成的，但角色住在平台账号里、模型是平台的；酒馆玩法里角色是你手里的卡文件，模型自己挑，提示词自己改，代价是要自己配置、没有官方内容工业。两派服务两种玩家，不分高下。 ### txt 小说转有声书 App 哪个好？2026 年四条路线横评（含全免费方案） URL: https://foreverse.app/zh/blog/txt-to-audiobook-apps-2026 Published: 2026-07-17 Q: 完全免费的听书路子有吗？ A: 有，而且不止一条。书在番茄这类平台里的，用平台自带 AI 朗读，免费但有广告插播；本地 txt 走静读天下加 MultiTTS 离线语音包，全程零元离线；Foreverse 这类阅读器里选系统 TTS 引擎，同样零成本还免折腾。免费方案的共同代价是音色上限：离线合成的水平听旁白够用，别拿大模型音色的标准要求它。 Q: AI 朗读和真人有声书的差距还大吗？ A: 旁白叙述已经很接近，长听不累；差距集中在多角色对手戏——真人演播会给每个角色换声线，AI 朗读大多单音色到底，群像戏会想念真人。听爽文、种田文基本无感。在线大模型音色和离线语音包之间也隔着一档，预算和耳朵自己权衡。 Q: 在线音色听一章要花多少钱？ A: 语音合成按字符计价，各家定价差异大，没法给一个通用数字。两件事能压成本：走 BYOK 直连供应商是按原价计费，没有中间加价；选带缓存的应用，同一章合成过一次就落在本地，重听不重复花钱。心疼预算的就用系统 TTS，零成本。 Q: 番茄导入的本地书为什么听不了？ A: 番茄小说支持把本地 txt/epub 导入书架阅读，但官方帮助明确写着导入的本地书暂不支持听书，截至 2026 年 7 月没有变化。它的朗读服务只覆盖平台书库。所以手里的 txt 文件想听，得换专门的工具，这正是本文四条路线里后三条存在的理由。 ### iPhone 怎么玩酒馆？苹果手机 AI 续写小说现有的三条路，和 Foreverse 这一条 URL: https://foreverse.app/zh/blog/iphone-tavern-and-continuation Published: 2026-07-17 · Updated: 2026-10-03 Q: Foreverse 的 iOS 版什么时候上线？ A: 已经上线：2026 年 9 月 iPhone 版上架 App Store（中国大陆区暂未上架）。本文写于 7 月，当时还是内测报名阶段，报名已随上架关闭。Android 版在 Google Play。 Q: 手头有台安卓备用机，值得先装 Foreverse 吗？ A: 值得。Android 版是完整形态：本机 txt/epub 导入（我们真机一次全选导过 315 本）、划词续写落分支、酒馆角色卡与世界书兼容、AI 听书都在。核心资产全是开放格式文件：角色卡是 chara_card_v3 的 PNG/JSON，小说是 txt 原文，这些文件本身不绑设备。 Q: iCloud 里存的 txt 能导进 Foreverse 吗？ A: 本文写于 iOS 版上线前；2026 年 9 月 iOS 版已上架 App Store，iPhone 上的导入方式以 App 内说明为准。Android 版里的做法有两条：从 iCloud 网页版把 txt 下载到手机再导入；或者把文件放进坚果云这类 WebDAV 网盘，在 App 里配置云端导入直接拉取。 Q: 云端酒馆和原生酒馆 App 怎么选？ A: 看你要生态还是要省心。云端跑的是原版 SillyTavern，桌面扩展全量可用，代价是月费、运维和把数据放在服务器上；原生 App 装完即用、数据存手机本机，代价是对桌面扩展生态的覆盖不完整。重度依赖某个桌面插件的选云端，主要就是导卡聊天的选原生 App。 ### AI 续写小说用哪个模型？九个大模型 × 两种文体实测排名（双盲评审） URL: https://foreverse.app/zh/blog/which-model-continues-novels Published: 2026-07-17 · Updated: 2026-08-13 Q: 免费模型里哪个续写能打？ A: 两场横评都没测免费档位，名次给不出，不硬编。数据里最接近的答案是 DeepSeek V4 Flash：玄幻场盲评第一，同时是六个模型里最便宜的档位之一，官方牌价每百万输入 token 0.14 美元。Foreverse 注册送 5000 credits，官方渠道约合 260 段续写，可以先把两张榜的前排试一遍，再决定要不要给谁付费。 Q: 模型更新这么快，这份榜单还准吗？ A: 名次绑定被测版本：2026 年 7 月测的是 DeepSeek V4 双档、Claude Opus 4.8、Gemini 3.1 Pro、Qwen 3.7 Max、GLM 5.2、GPT-5.6 Terra、Grok 4.5、Kimi K2.6 这一代。Kimi K3、Qwen 3.8 Max Preview、Grok 4.6 和 DeepSeek V4 Pro 正式版（0813）都已按同协议补测成对对决，结论收在正文的增补各节。8 月新增一条要紧提醒：DeepSeek 官方 API 同 ID 热切权重，榜上 DeepSeek 的名次描述的是已经调不到的 preview 快照。版本升级后名次可能变，本页会随复测更新页首日期。比名次更耐用的是方法：按文体选模型、一本书里允许逐段换，这两条不随版本过时。 Q: 中文小说和英文小说该用不同的模型吗？ A: 两场横评测的都是中文书（一本传统玄幻、一本宫斗古言），我们没有英文小说的排名数据，所以不给英文榜。已有数据支持一个更根本的判断：同一语言内换个文体，冠军都会掉到中游，「按文体选」比「按语言选」优先级更高。读英文书建议照搬这套方法自测：同一段开头让几个候选模型各续几段，盲着读完再翻牌。 Q: 双盲评审具体怎么防偏心？ A: 每个模型的产物匿名成随机字母，做两套不同的随机映射，分别交给两个互相不知道对方存在的评审独立排名。两套映射下结论一致的名次才作为定论；评审自标低置信或互有分歧的位置，一律按区间如实报告。玄幻场的前三与末位、宫斗场的首末梯队，在两套映射下完全一致。 ### AI 续写小说提示词怎么写？23 个官方模板按场景挑，可直接抄 URL: https://foreverse.app/zh/blog/continuation-prompt-templates Published: 2026-07-17 Q: 提示词能让 AI 完全模仿原文文风吗？ A: 能接近，做不到完美。我们跑过 20 轮连续续写实验：无论怎么加指令，所有实验组的句长波动系数都够不到原著的 0.84，短句猛烈交错的节奏是当前大模型的共同短板。风格模板加来源标注加一个合适的模型，能把差距缩到通读不出戏的程度；任何宣称百分百还原文风的说法都在夸大。 Q: 模板要自己抄进去吗？ A: 不用。23 个官方模板都上架在 App 内社区，免费下载，导入后直接出现在续写表单的模板选项里。手抄到别的工具也可以，本文节选的都是已公开上架的原文，但要注意模板末尾的 {{world_settings}} 这类占位符是 App 内自动填充的，手抄时得删掉或换成你自己的设定文本。 Q: 为什么模板里全是「禁止」和「不允许」？ A: 因为负面清单比正面形容词可执行。「写得有张力一点」模型没法落实，「旁白预告危险正在逼近，禁止」是一条能逐句检查的具体指令。11 张续写风格模板每张都带禁用清单，「眼中闪过一丝」这类套话被逐张点名，这是它们和网上大多数「万能续写咒语」最大的差别。 Q: 用了模板还是写得很 AI，怎么办？ A: 按顺序查三处：一是模型，不同模型的文风差距比提示词的影响更大，我们做过九个模型的双盲横评可以参考；二是上下文，确认来源标注开着、设定档案在供给，结构性漂移不归模板管；三是把模板的禁用清单按你踩到的雷点扩写，模板导入后就是你的，改比换更快。 ### AI 续写小说要花多少钱？免费额度、按量积分、自带 Key 三条路算清楚 URL: https://foreverse.app/zh/blog/ai-continuation-cost Published: 2026-07-17 Q: 免费额度用完了会怎样？ A: App 功能不锁：导入、阅读、分支管理照常用，核心功能免费也没有按条数限速。变化只发生在官方渠道的生成动作上，积分不足会提示，你可以按量买积分（最低档 ¥7.1 / 6,000 credits，约 315 段续写），也可以切到 BYOK 用自己的 Key 继续写，两条路随时并存。 Q: BYOK 自带 Key 要花多少钱？ A: 按你选的供应商牌价直连，Foreverse 零加价零抽成。以 DeepSeek 官方 2026 年 7 月牌价为例，deepseek-v4-flash 输出 2 元/百万 tokens、输入未命中缓存 1 元/百万 tokens，按一次续写输入 1 万 tokens、输出 800 tokens 的假设算，一段约一分二厘，20 万字大约 6 元。智谱 GLM-4.7-Flash 这类挂在免费分类的模型则是 0 元。 Q: 会不会自动扣费？ A: 不会。credits 是预付按量制，没有订阅、没有自动续费，用完必须自己手动买下一笔。BYOK 侧取决于供应商的计费方式，以 DeepSeek 为例是预充值扣减：费用从你充值的余额里扣，余额用完服务停止，不会从银行卡自动划钱。 Q: 官方渠道为什么比 BYOK 贵？ A: 官方渠道定价口径公开：模型官方牌价加 50% 服务费，覆盖渠道与服务成本。多出来的部分买的是不用注册供应商账号、不用充值管理 Key、开箱即写。写得少差额以分计，无所谓；月消耗到几十元量级，切 BYOK 直连能省下这三分之一。 ### 用 AI 续写别人的小说算侵权吗？只写给自己看的法律边界 URL: https://foreverse.app/zh/blog/ai-continuation-copyright-personal Published: 2026-07-17 Q: 用 AI 续写小说只自己看，会被起诉吗？ A: 公开渠道检索不到因私下续写自用被诉的判例。中文世界里因续写、衍生创作成讼的案子集中在公开出版与传播，例如金庸诉江南案争的是《此间的少年》的出版发行。私下自用贴近《著作权法》第二十四条「为个人学习、研究或者欣赏，使用他人已经发表的作品」的合理使用语义，也很难构成「不合理地损害著作权人的合法权益」。但「喂给云端模型」这个环节在学理上仍有讨论，无法说绝对零风险。 Q: 把 AI 续写发到网上（免费）算侵权吗？ A: 风险等级完全不同于自用。公开发布进入改编权与信息网络传播权的控制范围，是否收费主要影响赔偿数额，不改变侵权定性。金庸诉江南案二审即认定同人作品构成著作权侵权。如果一定要发，至少确认目标平台的同人与 AI 内容政策、如实标注 AI 参与、远离任何变现，并接受权利人要求下架的可能。 Q: AI 续写出来的文字，版权算谁的？ A: 北京互联网法院 2023 年 11 月「AI 文生图第一案」认定：AI 生成内容若体现使用者的独创性智力投入（提示词设计、参数调整、多轮筛选），可以构成作品，由使用者享有著作权，需个案判断。但续写还叠着一层：基于他人作品生成的内容属演绎性质，《著作权法》第十三条规定演绎作品行使权利时不得侵犯原作品的著作权。通俗说，你对续写部分可能有权利，但这份权利不给你公开使用它的自由。 Q: 用盗版网站下载的 txt 喂 AI，问题大吗？ A: 盗版来源是独立于「续写」的另一条侵权链：传播盗版文本的站点本身侵犯复制权与信息网络传播权。个人下载行为的定性虽有讨论空间，但拿有瑕疵的来源做原料，等于给自己的「个人使用」抗辩埋了一个说不清的前提。买正版或走官方免费渠道，成本不高，能把整件事从起点洗干净。 ### 彩云小梦还能用吗？还想「接着写下去」的人，2026 年有哪些选择 URL: https://foreverse.app/zh/blog/after-caiyun-xiaomeng Published: 2026-07-17 Q: 彩云小梦还能用吗？ A: 能用。截至 2026 年 7 月我们核实：小梦在应用宝、小米商店等主流安卓应用商店在架，2026 年上半年仍有版本更新（v3.14.x），隐私政策 2026 年 3 月刚更新过，模型侧还接入了 DeepSeek 等外部大模型。网上说的「凉了」更多指讨论热度，产品本身在正常维护。 Q: 小梦的平行世界和 Foreverse 的分支有什么区别？ A: 作用的对象不同。小梦的平行世界回溯发生在它生成的故事里：回到某个续写节点换一条走向，作品保存在小梦的账号体系内。Foreverse 的分支挂在你导入的整本书上：原文只读、永不被改，从任何一章任何一段都能开分支，分支之间可对比、回退、继续写，书和分支都是你手机上的文件，随时整包带走。 Q: 以前在小梦写的东西，能搬过来继续写吗？ A: 可以。把小梦里的作品正文复制出来存成 txt，导入 Foreverse 后它就是你的「原文」，从末尾或任何位置接着开分支写。导入支持批量，我们在一台真机上实测过 315 本 txt 一次全选导入。 Q: 换工具要重新花钱吗？ A: 注册送 5000 credits（1 credit = $0.0001），官方渠道一段续写实测约 19 credits，约合 260 段免费额度。长期用有两条路：继续按量购买，或者自带 API Key 直连模型供应商，按供应商原价计费，零加价。 ### 追的小说断更了怎么办？一份等更党生存指南（含新玩法） URL: https://foreverse.app/zh/blog/hiatus-survival-ai-branches Published: 2026-07-17 Q: AI 写的和正版对不上怎么办？ A: 一定对不上，而且对不上才是这个玩法的常态：AI 不知道作者的大纲，它写的是一条 if 线。正版复更后回主线看正版，AI 版留在旁边当档案。对照着读反而有独立的乐趣：同一个悬念，你能看清作者的选择和 AI 的选择差在哪。 Q: 会不会剧透我自己？ A: 概率很低。AI 不知道作者的大纲，它只根据已有正文推演，猜中大方向偶尔会有，猜中具体写法几乎不会。真怕「猜中式剧透」，就把分支故意推向明显的 if 方向：换个视角、提前发糖、金手指拉满，跟正版走向岔得越远越安全。 Q: 断更多久值得开一条 AI 分支？ A: 没有标准线，看戒断反应：重读还顶得住就再等等；一天刷三次更新页、开始分析作者点赞记录，就值得开了。成本不高，注册送 5000 credits（1 credit = $0.0001），一段续写实测约 19 credits，免费额度约够 260 段。 Q: 正版复更之后，AI 分支要删掉吗？ A: 不用。分支不占主线，原文和正版更新照常读；旧分支留着当 if 线收藏。之后剧情再出现你不满意的走向，还能从新位置再开一条——断更期练熟的这套手艺，复更后照样有用。 ### 小说烂尾了怎么办？忍、找同人，或者自己写一个结局 URL: https://foreverse.app/zh/blog/novel-bad-ending-write-your-own Published: 2026-07-17 Q: 烂尾的书 AI 真能续好吗？ A: 分两层看：单段续写的质量已经可用，长线一致性要靠你参与。把期望从「一键生成完美结局」调成「AI 出稿、你挑和砍」，成功率会高很多。文风能接近原著但做不到完美，写得越长越需要人工把关，这是当前所有模型的共同边界。 Q: 写崩了怎么办？ A: 砍掉那条分支就行。原文是只读的，AI 只在分支上写，删除分支不影响书和其它分支。也可以退回到分支上任何一个还满意的节点，换个方向重开。「重新生成」同样以新分支保存，不会覆盖你已有的结果。 Q: 自己写结局要花钱吗？ A: 注册送 5000 credits（1 credit = $0.0001），官方渠道一段续写实测约 19 credits，相当于约 260 段免费额度，写完一个中等篇幅的结局绰绰有余。之后想省钱可以自带 API Key 直连模型供应商，按供应商原价计费，零加价。 Q: 书还在连载、只是断更，也适用吗？ A: 适用，而且比烂尾更从容：AI 在分支上先写着，正版更新后回主线看正版，两条线互不覆盖。这个「分支等更新」的玩法我们单独写了一篇等更生存指南，见文末相关阅读。 ### Three Ways LLMs Fail at Long Fiction: Restart Loops, Mid-Run Freezes, and Ending Rewinds URL: https://foreverse.app/blog/three-ways-llms-fail-long-fiction Published: 2026-07-16 · Updated: 2026-07-28 Q: What setup produced these failures? A: Each model continued the same novel from the same anchor point for 20 consecutive rounds, with every round's output appended back into a fixed 16k-token context window. That feedback loop is what real reader apps do — and it is what surfaces these failures. One-shot benchmarks never see them. Q: Are these failures deterministic — will a model always fail the same way? A: Mostly consistent, per model per genre, in our data. Grok 4.5 hit repetition loops in both a fantasy epic and a palace-intrigue novel. GPT-5.6 Terra's ending rewind appeared in the fantasy run but not the palace one. Gemini's restart loop was fully explained by one ambiguous sentence in our prompt — after rewording it, zero regressions in 28 rounds across two books. Q: Does the genre of the book change which model performs best? A: Dramatically. Our fantasy champion (DeepSeek V4 Flash) dropped to sixth-seventh place on the palace-intrigue novel, while its sibling V4 Pro — dinged in fantasy for over-describing — took first place there because fine-grained description is exactly that author's style. Same trait, opposite sign, different genre. Q: What is the practical takeaway for readers using AI continuation? A: Pick the model per book, not from a single leaderboard. And prefer apps that let you switch models mid-book: long-run failures build up over rounds, and switching models (or regenerating a segment) resets the loop before it locks in. Q: Do these failure modes disappear as models get newer? A: They can, measurably, and per model. Kimi K2.6's mid-run freeze once spiked to 97.6% cross-round repetition; its successor Kimi K3, re-tested on July 28, 2026 under the same protocol, peaked at 0.5% across 140 rounds with zero verbatim-loop flags from either blind reviewer — and re-ranked to the top of both genre boards. But K3 also grew new quirks (aphorism endings, mixed-width quotes), so the taxonomy outlives any instance: re-test each generation instead of carrying a model's old reputation forward. ### Six LLMs Continue the Same Webnovel. A Blind Review Picks the Most Faithful — 120 Rounds of Data URL: https://foreverse.app/blog/which-model-clones-your-novel-style Published: 2026-07-16 · Updated: 2026-07-28 Q: Why test 20 consecutive rounds instead of one-shot generation? A: With one-shot, the context is pure original prose and every model mimics it decently. In real usage, AI-written passages keep getting appended back into context, so the model increasingly imitates its own output — drift compounds. We ran the single-shot control first (4k to 200k token tiers; differences were mild). Switching to 20 consecutive rounds turned mild differences into clearly visible quality tiers. Q: How was the blind review kept honest? A: Outputs were anonymized under random letters with two different random mappings, then given to two reviewers who didn't know each other existed. We only trust placements where both mappings agree — top three and last place matched exactly. Both reviewers independently flagged their #4/#5 calls as low-confidence, so we report those as a tie. Q: Can sentence-length and dialogue-ratio stats replace human reading? A: No — they misfired twice in this run. One model matched the original's sentence length and dialogue ratio almost perfectly yet placed fifth in blind review, because it kept writing smell-and-texture descriptions the original author never uses. Stats can't measure 'wrote things the original never writes'. Another model showed 0% dialogue in our stats because it used half-width quotes throughout; the regex only counted full-width ones. Stats make good regression gates, not judges. Q: Was the 'rewrites the opening' bug a model failure? A: Only partly. Our context labels original prose vs. AI continuations, and the explanation line said AI passages were 'for plot continuity reference'. Gemini 3.1 Pro read that as 'these passages aren't canon', skipped all of them every round, and restarted from the original text's ending — 20 out of 20 rounds. Rewriting the line to state that AI passages are canonical events that must be continued from the last paragraph brought regressions to 0 out of 20. Prompt wording has to be designed for the model most likely to misread it. ### 九个大模型续写《甄嬛传》，双盲评审谁最像流潋紫——GPT-5.6、DeepSeek、Grok 全下场 URL: https://foreverse.app/zh/blog/nine-llms-continue-zhenhuan Published: 2026-07-16 · Updated: 2026-08-13 Q: 为什么玄幻测出的排名不能直接用在言情/古言上？ A: 因为鉴别维度完全不同。玄幻看白话直给、拟声词、战斗推进；古言看第一人称限知视角纪律、典雅语体、称谓礼数（臣妾/本宫/娘娘用对没有）、对话机锋是不是绵里藏针。我们两组实验的冠军互不重叠：玄幻第一的模型在甄嬛传里跌到第六，古言前三里有两个在玄幻榜上并不拔尖。 Q: 「话中有话」这种东西，盲评是怎么判断的？ A: 评审的依据是原著的双层结构：人物嘴上说体面话，叙述者「我」在心里拆解真实意图。九个系统里只有一个稳定复现了这个结构——其余的要么只写了面子话，要么让人物把心机直接说出口，变成现代吵架。这类证据比统计指标锋利得多。 Q: 评测里有模型「记忆污染」是什么意思？ A: 甄嬛传小说和电视剧的专名体系不同：小说里太后住颐宁宫，剧版是寿康宫；小说是凤仪宫皇后，剧版对应景仁宫。有的系统续写时写出了剧版专名——说明它的训练记忆里电视剧语料压过了小说原文。这个信号与文风贴近度基本同向：记得小说的，文风也更贴小说。 Q: 这个实验对我选模型有什么实际意义？ A: 按你读的书选模型，不要迷信任何一张总榜。我们的数据里读快节奏玄幻选 DeepSeek V4 Flash 最划算，读古言宫斗选 DeepSeek V4 Pro 或 GPT-5.6 Terra 更稳。2026-07-28 按同协议补测的 Kimi K3（走我们内部 agent 通道测的），三条女频链落位第 1、3、5，最强链双评审一致第 1，古言候选名单里如今该有它一席。在 Foreverse 里模型可以逐段切换，同一本书里也能按场景换着用。 ### AI 续写为什么越写越不像原著？我们跑了 60 条 20 轮的链找答案 URL: https://foreverse.app/zh/blog/ai-continuation-style-drift Published: 2026-07-16 Q: 为什么 AI 续写会越写越不像原著？ A: 因为续写是自回归的：AI 写的每一段都会拼回上下文，成为下一段的「参考文本」。第一轮上下文全是原著，第二十轮上下文里可能一半以上是 AI 自己的输出——模型越来越多地在模仿自己，而不是原著。我们实测裸提示词的链在 20 轮里句长波动持续收窄、文艺腔比喻密度上升、同一个意象反复自我复制。 Q: 把原著多塞一点进上下文，能不能自动让 AI 学会原著文风? A: 实测不能。我们做过 4k 到 20 万 token 五档的对照：把 27 万字原著整个塞进上下文，不加任何指令的组照样写出满篇「宛如/仿佛/一丝」，比喻密度是原著的 10 倍。上下文提供的是可模仿的材料，「以谁为基准」必须显式说，模型不会自己猜对。 Q: 来源标注是什么？普通用户需要自己配置吗？ A: 它是续写上下文里的分界行，告诉模型「从这里往上是原著正文，从这里往下是此前的 AI 续写」，配合一句固定声明：文风、用词、句式节奏以原著段落为基准。在 Foreverse 里这是设置里的一个开关，打开即生效，不需要写任何提示词。 ### 六个大模型续写同一本网文，谁最像原著？——120 轮实测 + 双盲评审 URL: https://foreverse.app/zh/blog/which-model-clones-your-novel-style Published: 2026-07-16 · Updated: 2026-07-28 Q: 为什么测「连续续写 20 轮」而不是单次生成？ A: 单次生成时上下文全是原著，模型抄得像很正常。真实使用里，AI 写的段落会不断拼回上下文，模型开始模仿自己的输出，文风偏移是累积的。我们先做过单轮对照实验，档位从 4k 到 20 万 token，各模型差距温和；换成 20 轮连续续写后，差距放大成肉眼可辨的分层。 Q: 盲评怎么防止评审偏心？ A: 六个模型的产物被匿名成随机字母，做了两套不同的随机映射，交给两个互相不知道对方存在的评审独立排名。两套映射下前三名和末位完全一致才采信。评审自己也要给每个名次标置信度——四五名的相对位置两个评审都标了低置信，我们就如实报告为并列区间。 Q: 句长、对话率这些统计指标能代替人读吗？ A: 不能，这次实验里它们还闹了两个笑话：一个模型的句长和对话率统计几乎和原著完美重合，盲评却排第五，因为它每轮都在写原著从来不写的气味和触感描写——统计测不出「写了原著没有的东西」；另一个模型对话率统计是 0%，其实是它全程用半角引号写对话，正则没认出来。统计适合当回归闸门，「像不像」的裁决还得靠细读。 Q: 评测里发现的「重写开场」bug 是模型的问题吗？ A: 不全是。我们在上下文里给原著和 AI 续写段落做了来源标注，说明文案里有一句「AI 续写段落仅供情节衔接参考」——Gemini 3.1 Pro 把这句读成了「这些段落不算正文」，于是每轮都跳过全部已续写内容、回到原著末尾重写。把说明改成「AI 续写段落是已发生的正文剧情，必须从最后一段继续」之后，同一个模型 20 轮零回退。提示词措辞的鲁棒性要按最容易误读的模型设计。 ### We Dumped 315 Real Novels Into a Phone Reader, Then Opened Ten at Random URL: https://foreverse.app/blog/315-books-in-8-seconds Published: 2026-07-15 Q: How long does importing a few hundred books take? A: Measured: 315 real txt files (about 1GB total) in one select-all, batch ingestion in roughly 8 seconds with no freeze and no crash. A Xiaomi device imported 82 books in about 7 seconds. Import is register-first: books shelve in a pending state and full parsing runs on first open. Q: How long does a pending book's first open take? A: On the vivo foldable, 3.3 to 6.8 seconds including full chapter parsing and first-page layout; 9.5 to 14.5 seconds on the Xiaomi. Later opens skip conversion and ride the index — around 5 seconds for converted books on the Xiaomi, varying with device and book size. Q: How was the 0.49% jank figure measured? A: With Android's built-in frame statistics (gfxinfo): after the full import-plus-random-reading session, the accumulated counters read 71,320 frames with 0.49% janky and a 99th-percentile frame time of 31ms. System-reported numbers, not self-instrumented ones. Q: Can it find books hidden in other readers' folders? A: Yes. Smart scan covers the usual stashes — browser download folders, other readers' private directories, dot-prefixed hidden folders. Already-authorized directories scan directly, and results at the 315-book scale arrive within seconds. ### Your Books, Your Cards, Your Chats — As Files URL: https://foreverse.app/blog/your-worlds-are-files Published: 2026-07-15 Q: What does file-first storage mean concretely? A: Each of your worlds (a book or a card's whole life) is a directory on your device: prose, branches, lorebook, character dossiers, chat logs — visible files in standard or documented formats (chara_card_v3, JSON, markdown). Moving out is a copy operation, not a plea for an export endpoint. Q: Can AI companion data leave too? A: Yes, through a dedicated moving box: persona, individual memories, stickers, avatar, and transcripts export as one zip and restore on another device. The import side validates hard — entry whitelist, path-traversal rejection, automatic rollback on failure — so a corrupted box cannot poison existing data. Q: What's different about content downloaded from the community? A: It carries a provenance record — which community pack, which version — written into the content's metadata for your verification. Deliberately, that record does not propagate on re-share: when you send a card to a friend, they get a clean standard file, not your download history. Provenance serves you; it is not a tracking beacon. Q: Do BYOK keys or chat contents ever get uploaded? A: API keys are encrypted and stay on the device; requests go directly to the provider you chose, never through our servers. Chats live in local directories; the optional official billing channel records billing-side data only. The data-export page can produce a full export or deletion at any time. ### A Tour of the Plugin Center: Desktop Tavern Extensions, Rebuilt for a Phone's Budget URL: https://foreverse.app/blog/tavern-plugin-center-tour Published: 2026-07-15 Q: Do plugins quietly inflate my API bill? A: Every call is itemized. Plugin side-requests (generating choices, extracting state, writing summaries) appear as separate rows in the API request log, labeled with the plugin's name, tokens and cost split from the main chat. Hard budget on top: at most one blocking plugin call per turn, at most two post-reply quiet calls, overflow deferred to the next turn. Q: How does the memory plugin work? A: Two engines. Fact memory extracts key facts into local entries and injects the most relevant on each turn (retrieval has four modes: auto, full, relevant, recent); session summary compresses early history once a long chat crosses a threshold. Both inject at the message-level tail on the user track — a position we chose from a 400-round caching experiment. Q: What's the difference between story choices and reply ideas? A: Point of view. Story choices offers three plot-direction phrases (director's view); tapping one fills your input box for you to edit and send. Reply ideas drafts three complete lines in your voice (actor's view); tap to send, long-press to edit. With both enabled, only one runs per turn — ideas takes priority — so you never pay for both. Q: Does stepped thinking slow replies down? A: It adds one preliminary call: the model drafts private notes first, then writes the real reply with them in hand — noticeably better on complex personas. A 15-second timeout degrades silently to a normal generation, so your message never hangs on it. It disables itself in group chats, where multiple inner monologues would cross-contaminate. ### The Real Problem in AI Group Chat: Who Speaks Next? URL: https://foreverse.app/blog/who-speaks-next-group-chat Published: 2026-07-15 Q: How does an AI group chat decide which character speaks next? A: The natural strategy arbitrates in three tiers: mentions first (name a character and they answer); otherwise a weighted draw over each character's talkativeness field (a standard card property); otherwise fall back to member-list order so silence never stalls the scene. The previous speaker is excluded by default to prevent spam, and each turn has a maximum-speakers budget. Q: Can I import group chats exported from desktop SillyTavern? A: Yes. Desktop group JSON imports whole: member roster, speaking-order strategy, and group settings come together. Member cards must already be in your library; the import report lists anyone missing. Q: Do characters in a group see each other's personas? A: That depends on the card-injection mode, and there are three: solo (only the current speaker's card is injected — cheapest, characters stay mutually opaque), merged-enabled (every active member's card goes in — full mutual knowledge, highest token cost), and merged-all (even muted members' definitions are included, for present-but-silent characters). Q: Can the group keep playing without me? A: Turn on auto mode with an interval in seconds: after each reply lands, the next speaker is scheduled automatically and the scene rolls forward. Interject any time — your message cancels the queue and re-arbitrates; switch it off and it stops. ### How Character-Card Beautification Survives on a Phone URL: https://foreverse.app/blog/card-beautify-on-phones Published: 2026-07-15 Q: Do beautified cards tuned on desktop SillyTavern work directly on the phone? A: Behavioral parity is the goal: when a card carries HTML beautify regex, import auto-enables rich rendering for it, and the marker-to-regex-to-HTML chain mirrors desktop semantics. What breaks are two desktop habits: external resources (CDN jQuery, remote fonts, hotlinked images) never load inside the mobile sandbox, and scripts that reach into the parent page's DOM have no equivalent. Self-contained HTML+CSS cards port cleanly. Q: Does beautify code eat my context tokens? A: No, and that is the point of the pipeline. The model outputs only short plain-text markers each turn — a few dozen tokens. The hundreds of lines of HTML live in the regex replacement string and apply at display time only. With the display-only flag set, replacements never enter the history fed back to the model: it always sees clean markers, and your HTML costs zero tokens. Q: Why can't my marker look like an HTML tag? A: Import sanitization. Raw HTML tags inside card text fields are stripped at import (standard hygiene for untrusted card content), so a marker shaped like never survives to rendering. The rule: markers are plain-text symbols (:::think, [status|...]), and HTML exists only inside regex replacement strings. Q: Can interactive beautification work on mobile? A: Yes. The WebView exposes a working subset of the host API: triggerSlash runs commands like /setinput, /send and /trigger; getChatMessages reads the transcript; generate() genuinely triggers a second generation. Clickable choice menus and buttons work, and buttons suppress text selection so tapping an option never accidentally selects prose. ### An AI Agent That Lives on Your Bookshelf: Five Tasks, Logged (Including One Self-Repair) URL: https://foreverse.app/blog/an-agent-on-your-bookshelf Published: 2026-07-15 Q: How is this different from a chatbot? A: It has tools. Read, write, edit, and execute run over the model's native tool-calling protocol, and every book, character card, and lorebook on your shelf is a real file in its workspace. Ask it to 'turn this novel's cast into dossiers' and it actually reads the chapters and actually writes the entries into the world's directory — every step shows as a visible tool card, and writes wait for approval. Q: Can it touch anything else on my phone? A: No, deliberately. Its workspace is your study: novels, cards, lorebooks, notes. Contacts, SMS, and screen automation were never granted — community projects do explore that direction (RikkaHub's agent fork ships 80+ device tools), and we chose the opposite boundary: full power inside the library, no power outside it. Q: Will it mangle or delete my files? A: Write operations go through an approval card by default: which file, changed how, laid out on screen before anything lands on disk. Approval has three modes (ask every time / ask for risky ops only / full auto), switchable per session. Runs keep a complete reviewable log, and anything shaped like 'token: value' in tool output is masked before it reaches the model's context. Q: What does a task cost? A: It runs on your own API key at provider prices. A 'fill in the lorebook' task is a read-write loop in the tens of thousands of tokens — pocket change on a flash-tier model. The 121-round stress test ran on DeepSeek's official API and cost less than a soft drink. You can switch models per task: flagship for hard jobs, cheap for chores. ### Where Should Memory Go in a Roleplay Prompt? Our First Answer Was a Caching Artifact URL: https://foreverse.app/blog/where-to-inject-roleplay-memory Published: 2026-07-15 Q: Where should a memory block be injected into a roleplay prompt? A: Message-level tail, but never as a system role. In our 400-round test, merging memory into the head of the latest user message kept cache hits at 96.6% on rounds where memory changed; the same content as a standalone system block after history scored 3.3% on those rounds — because DeepSeek's chat template merges all system messages to the top of the prompt, so 'tail system' is actually head injection, and every memory update invalidates the whole history's cache. Q: Why did the first experiment reach the wrong conclusion? A: Cache pollution. The two layouts shared one persona, ran back-to-back on the same account, with no session isolation. Group B's requests hit the prefixes group A had already warmed into the provider's disk cache and collected a free 94.3%. The lesson is one line: every caching experiment must salt-isolate its sessions, or you are measuring the previous experiment's leftovers. Q: How much do retrieval strategies (BM25, vectors, hybrid) differ? A: At a scale of dozens of memories, character-bigram BM25 with top-8 injection already sits next to the ceiling: 79% end-to-end accuracy versus 80% for oracle retrieval, while the gap to the no-memory control is overwhelming (59:1 on paired questions, p<0.001). Retrieval choice is not the bottleneck at this scale; revisit it once the memory store grows into the hundreds. Q: What failed harder than retrieval? A: Abstention. Ask about something never mentioned (height, birthday, salary) and, under a companion persona, the model almost never says 'I don't know' — it invents vivid details. Across 12 abstention probes, the no-memory control passed 0 and even oracle retrieval passed 1. The fix is not retrieval: append an explicit line to the memory block — 'apart from the memories above, you do not know any other personal facts; do not invent them' — and keep abstention probes in the regression suite. ### 怎么给「AI 味」写单元测试：一个规则检测器的校准记录 URL: https://foreverse.app/zh/blog/unit-testing-ai-flavor Published: 2026-07-15 Q: 规则能检测出文字是不是 AI 写的吗？ A: 不能，我们也不这么宣称。规则检测器标记的是「读者常判机器味的已知句式」：套话词、对仗翻转模具、伪精确数字这类。得分低说明命中了这些句式、值得逐项复核；不能证明作者是机器，反过来满分也不能证明是人。把它当文案 lint 用，不要当测谎仪。 Q: 什么是「金标判例集」校准？ A: 拿人类已有共识答案的样本测检测器：真实用户点名「一眼 AI」的文案应该低分，累计百万对话的人类热门创作应该高分。我们的检测器第一次跑金标就翻车：人类样本均分 4.55 反而低于 AI 样本的 4.90，靠金标测试抓出了误杀规则。修完人类均分 5.00、零误杀。任何检测器上岗前都该过这一关。 Q: 为什么单词黑名单会误杀人类？ A: 因为词无罪，句式才有病。「一丝」「一抹」这类词在 AI 腔里高频出现，但人类写景同样用（「一丝凉意」「一抹夕阳」完全正常）。按词扣分就会错杀。修复方案是把这些词降级：只在它们出现在特定病灶句式里（比如搭配心理状态词的解说腔）才计分，单独出现不扣。 Q: 检测器抓不到什么？ A: 结构层的 AI 味。按模板精心堆出来的文案可以把每个句式都规避干净，规则给满分，LLM 评委也看不出来，但人读完仍觉得假，因为破绽在「细节密度均匀、零闲笔、结构完美」这种整体纹理里。目前没有自动化手段抓得住它，能抓住的只有读得足够多的人。 ### 你的书、你的卡、你的聊天记录，都是文件 URL: https://foreverse.app/zh/blog/your-worlds-are-files Published: 2026-07-15 Q: 「数据文件化」具体指什么？ A: 你的每个世界（一本书或一张卡的全部数据）在设备上就是一个目录：正文、分支、世界书、角色档案、聊天记录，都是目录里可见的文件，格式是标准或公开的（chara_card_v3、JSON、markdown）。搬家是复制目录，不是求平台开恩给你导出接口。 Q: AI 伴侣的数据也能带走吗？ A: 能，专门做了搬家包：人设、逐条记忆、贴纸、头像、聊天转录打成一个 zip 导出，换设备后导入还原。导入侧有完整校验（条目白名单、路径穿越拒绝、失败自动回滚），坏包不会污染现有数据。 Q: 从社区下载的内容有什么不一样？ A: 多一份溯源记录：从哪个社区包、哪个版本导入的，写进内容的元数据，来源可查。这份记录不随再分享传播——你把卡分享给朋友时，带出去的是干净的标准文件，不夹带你的下载历史。 Q: BYOK 的密钥和聊天内容会上传吗？ A: API 密钥加密后只存在设备上，请求直连你选的模型供应商，不过我们的服务器。聊天内容存在本地目录里；官方渠道计费时服务端只记账单侧数据。你随时可以在数据导出页发起完整导出或删除。 ### 插件中心导览：把桌面酒馆的扩展生态过一遍手机的安检 URL: https://foreverse.app/zh/blog/tavern-plugin-center-tour Published: 2026-07-15 Q: 插件会不会偷偷多花我的钱？ A: 每一笔都记账。插件的副请求（生成选项、提取状态、写摘要）在 API 请求记录里单独列行、标注来源插件名，token 和费用和主对话分开可查。预算上还有硬约束：阻塞主生成的插件每轮最多 1 个，回复落定后的静默调用每轮最多 2 个，超出的顺延到下一轮。 Q: 记忆插件是怎么工作的？ A: 双引擎。事实记忆把对话里的关键事实提取成条目存本地、按相关度检索注入（检索有 AUTO/全量/相关/最近四档可选）；会话摘要在长聊达到阈值后把早期内容压缩成摘要垫底。两条注入都在消息级末尾走 user 轨道，这个位置是我们用 400 轮实验选出来的，缓存友好。 Q: 「剧情选项」和「回复灵感」有什么区别？ A: 视角不同。剧情选项给的是剧情走向的行动短语（偏导演视角），点击填进输入框由你改定；回复灵感给的是以「你」的口吻写好的三条完整台词（偏演员视角），点选即发、长按可编辑。两个插件同开时同一轮只跑一个，灵感优先，避免双倍副请求。 Q: 先想后答会拖慢回复吗？ A: 会增加一步前置调用：先让模型私下打一份内心草稿，再带着草稿写正式回复，人设复杂的卡效果明显。我们设了 15 秒超时，想不出来就静默跳过、直接正常生成，不会卡住你的消息。群聊里这个插件自动禁用（多角色的内心戏会互相穿）。 ### AI 群聊真正的难题：下一句该谁说？ URL: https://foreverse.app/zh/blog/who-speaks-next-group-chat Published: 2026-07-15 Q: AI 群聊里系统怎么决定下一个说话的角色？ A: 自然轮替策略走三层裁决：先看点名（消息里提到谁的名字，谁优先接话）；没有点名就按各角色的话痨值加权抽选（talkativeness 是角色卡里的标准字段）；都没有就按成员列表顺序兜底。默认排除刚说过话的角色避免刷屏，每轮有最大发言人数预算。 Q: 桌面 SillyTavern 的群聊文件能直接导入吗？ A: 能。桌面酒馆导出的群聊 JSON 直接导入：成员卡、发言顺序策略、群设置一起进来。成员对应的角色卡需要先在库里，缺谁导入报告会列出来。 Q: 群聊里每个角色都能看到彼此的人设吗？ A: 取决于角色卡注入方式，有三档：单独注入（只注入当前发言者的卡，省 token、角色间信息隔离）、合并全员（所有启用成员的卡都注入，彼此知根知底）、合并含禁用（连暂时禁言的成员设定也注入，适合「在场但不说话」的角色）。 Q: 能让群聊自己一直演下去吗？ A: 可以开自动模式：设一个间隔秒数，每轮生成落定后自动排下一位发言，戏就自己往前走。你随时插话，用户发言会立刻取消排队并重新裁决该谁接；关掉开关即停。 ### 酒馆美化包在手机上是怎么活下来的：给创作者的一封技术信 URL: https://foreverse.app/zh/blog/card-beautify-on-phones Published: 2026-07-15 Q: 桌面酒馆调好的美化卡，在手机上能直接用吗？ A: 目标是行为一致：卡带 HTML 美化正则时，导入会自动打开这张卡的富渲染开关，marker→正则→HTML 的链路和桌面同构。会翻车的主要是两类桌面习惯：引用外部资源（CDN 上的 jQuery、远程字体、外链图片）在移动沙箱里一律加载不出来；依赖整页浏览器环境的脚本（操作父页面 DOM）没有对应物。自包含的 HTML+CSS 卡基本无痛。 Q: 美化代码会吃掉我的上下文 token 吗？ A: 不会，这是链路设计的关键点。模型每轮输出的只是纯文本 marker（比如 [状态|心情|好感]这样的短标记），几十个 token；几百行的 HTML 模板放在正则的替换串里，只在显示期生效。配合 markdownOnly 设置，替换结果不会写回喂给模型的历史——模型永远只看到干净的 marker，你的 HTML 一个 token 都不占。 Q: 为什么我的 marker 不能写成 HTML 标签的样子？ A: 因为导入净化。角色卡文本字段里的生 HTML 标签会在导入期被剥掉（这是对不可信卡内容的安全处理），marker 如果长成 这样就活不到渲染那一步。规矩是：marker 用纯文本符号（:::think、[状态|…] 这类），HTML 只出现在正则的替换串里。 Q: 手机上能做可点击的交互美化吗？ A: 能。WebView 里支持宿主 API 的常用子集：triggerSlash 可以执行 /setinput、/send、/trigger 这类命令，getChatMessages 能读消息，generate() 能真实触发二次生成——可点选项、按钮菜单这类交互组件都能落地。细节按钮和链接默认禁用文本选择，点选项不会误选中文字。 ### 把 Agent 放进书架：五个任务的实录，含一次自我纠错 URL: https://foreverse.app/zh/blog/an-agent-on-your-bookshelf Published: 2026-07-15 Q: 这个 Agent 和聊天机器人有什么区别？ A: 它有工具。读文件、写文件、改文件、跑命令四件套走模型原生的工具调用协议，你的每本书、每张角色卡、每份世界书都是它工作区里的真实文件。让它「把这本书的人物整理成档案」，它会真的去读正文、真的把档案写进世界目录——每一步操作有可见的工具卡片，写盘前有审批。 Q: 它能动我手机上的其它东西吗？ A: 不能，这是刻意的。它的工作区就是你的书房：小说、角色卡、世界书、笔记。通讯录、短信、屏幕自动化这类设备权限一概没有给——社区里有走那个方向的项目（比如 RikkaHub 的 agent fork，80 多个设备工具），我们选了相反的边界：书房里全能，书房外无权。 Q: 它会不会乱写乱删我的东西？ A: 写操作默认要过审批卡：改哪个文件、改成什么样，屏幕上先摆出来，你点了才落盘。审批有三档（每次都问 / 只问危险操作 / 全自动），可以按会话切换。另外每次运行的操作日志完整可回看，工具读到的内容里形如 token: 值 的敏感信息会先脱敏再进模型上下文。 Q: 用它干活要花多少钱？ A: 走你自己的 API Key，按实际 token 计费。一次「补世界书」量级的任务通常是几万 token 的读写循环，用 flash 档模型是几分钱人民币的量级；121 轮的压力测试跑的是 DeepSeek 官方 API，成本在一杯饮料以内。模型可以按任务换，复杂任务换旗舰，日常杂活用便宜的。 ### 记忆块该插在 Prompt 哪里？我们的第一个结论是缓存假象 URL: https://foreverse.app/zh/blog/where-to-inject-roleplay-memory Published: 2026-07-15 Q: 角色扮演的记忆块应该注入到 prompt 的什么位置？ A: 消息级末尾、但不要用 system 角色。我们的 400 轮实测：把记忆并入最新一条 user 消息的头部，记忆变化轮的缓存命中 96.6%；同样内容放在历史末尾的独立 system 块，变化轮命中只有 3.3%。原因是 DeepSeek 的对话模板会把所有 system 消息归并到 prompt 开头，「末尾 system」实际等于头部注入，记忆一变化整条历史的缓存全部作废。 Q: 为什么第一次实验会得出错误结论？ A: 缓存污染。第一轮实验的两组布局共用同一套人设、先后跑在同一个账号下、没有做会话隔离，B 组的请求恰好命中了 A 组灌进服务端磁盘缓存的前缀，白捡了 94.3% 的命中率。教训只有一句：所有缓存实验必须给会话加随机盐隔离，否则你测的是上一个实验的余温。 Q: 检索策略（BM25、向量、混合）差距大吗？ A: 在几十条记忆的量级下，字 2-gram BM25 取 8 条已经接近上界：端到端答题 79%（理想检索的 oracle 是 80%），与无记忆对照的差距是压倒性的（同题配对 59:1，p<0.001）。检索算法的选型在这个量级上不是瓶颈，记忆库长到几百条之后才需要重新评估。 Q: 比检索更严重的问题是什么？ A: 弃权。问模型一件从未提过的事（身高、生日、工资），它在伴侣人设下几乎从不说「我不知道」，而是编得栩栩如生：12 道弃权探针，无记忆组 0 题通过，连理想检索组也只过 1 题。修法不在检索侧：必须在记忆块尾部显式写「除以上记忆外，其余个人信息你并不知道，不许编造」，并把弃权题纳入回归测试。 ### The Day Your AI Companion Sends a Selfie — and Why the Face Is the Hard Part URL: https://foreverse.app/blog/the-day-they-sent-a-selfie Published: 2026-07-11 Q: How does the app keep a companion's face consistent across images? A: Every companion carries a visual identity: a reference image set you confirmed by hand, plus written appearance anchors (the unchangeable traits). Selfies, moment cards, and photos-together all generate against that identity. The discipline that makes it work: generated images never auto-join the reference set — only images you explicitly approve define the face. Q: How consistent is it, measured? A: Before shipping we ran an internal blind rubric: generate repeatedly from one identity, score features, hair, and overall presence against the references. Our two main image models passed at 9 of 10 and 10 of 10. Honest caveat: extreme poses and lighting still drift sometimes, which is why every generated image ships with a regenerate button. Q: Is my photo stored when I make a photo together? A: Not by the app. Your photo is sent in memory only to the image provider you chose, for that single generation — never written to the app's disk, never logged, released when the request ends. A local adult self-confirmation gate runs before the feature, and the two identities are explicitly order-bound in the request to minimize the risk of swapped faces. Q: Can proactive photo moments run up my bill? A: There are multiple brakes. A photo moment writes an idempotent placeholder before calling the model, so a process restart can't double-bill; attempts are capped per day, repeated failures stop automatically, and long inactivity pauses the feature. One honest note: the model call happens after the placeholder, so a failure that occurs after the call (say, a failed save) has already spent that one call. Every call runs on the provider you configured, itemized in the request log. ### Our LLM Judges Called Human Writing AI-Flavored 88% of the Time URL: https://foreverse.app/blog/llm-judges-called-humans-ai Published: 2026-07-11 · Updated: 2026-07-14 Q: What is gold calibration for LLM judges? A: Before trusting a judge, you test it on anchor samples where humans already agree on the answer. Our four judges scored 12% on those anchors — far below the 50% a coin flip would get, which means the panel wasn't noisy, it was systematically inverted. Q: Doesn't 86% inter-judge agreement mean the panel is reliable? A: No, and that's the expensive lesson. Agreement measures consistency, not correctness: four judges can share the same training-induced bias and confidently converge on the wrong answer. Majority voting cancels random noise; it cannot cancel a shared systematic bias. Q: Did anti-bias prompting fix it? A: It depends on the model. In a follow-up calibration round on an expanded corpus, an explicit debiasing preamble lifted a Claude-family judge to 83% gold accuracy (the original four-judge setup scored 12%). DeepSeek, GLM and Qwen stayed at 25–33% with the identical preamble — their bias sits below the prompt layer. And 83% only barely clears the ≥80% bar we set afterwards. Q: So how do you detect AI flavor now? A: A three-layer setup: a rule-based detector (phrase lists plus sentence-pattern quotas, calibrated against real user judgments) as the regression gate; human spot checks as ground truth; and LLM blind review only for relative before/after comparisons of the same text, never as an absolute verdict. Honest limit: templated AI copy with perfect structure passes both rules and judges — no automated silver bullet exists. ### Characters Shouldn't Die When the Book Ends — an Opinion on Where AI Roleplay Went Wrong URL: https://foreverse.app/blog/characters-shouldnt-die Published: 2026-07-11 Q: What does 'four lives of a character' mean concretely? A: One character, four connected surfaces: continued and illustrated inside the novel (page), performing chapters as a visual novel with choices that write real branches (stage), chatting with the full card and worldbook attached (table), and promoted to a companion with line-editable long-term memory and optional proactive messages (side). Each step shares the same data — not four disconnected features. Q: How is this different from Character.AI or Replika? A: Chat-first platforms hold the middle of the journey and nothing else: characters have no origin (no book they came from), no stage, and no exportable future. Here a character can enter from a novel you actually read, and leave as a standard character card carrying selected memories — the platforms around a single chat box can't do either. If all you want is casual chat with zero setup, a chat-first platform is honestly simpler. Q: Does the companion know my reading progress? A: No — deliberately. A companion remembers what you told it and what happened between you; it does not surveil what you read. We drew that line on purpose and it stays. Q: Can I take my character with me if I leave? A: Yes. Companions export as standard chara_card_v3 character cards, with memories you explicitly select becoming the card's lorebook. Cards, chats, and memory live as local files on your device in the first place. ### We Added a Game Mode to an E-Reader. Here's the Design Decision That Made It Work URL: https://foreverse.app/blog/first-reader-with-game-mode Published: 2026-07-11 · Updated: 2026-07-19 Q: Does Game Mode convert the novel into a game project? A: No. The theater compiles the book's current branch into a stage script on the fly, locally, with zero AI calls. There is no exported project and no migration — the theater and the reader share one reading position, so you exit exactly where the performance stopped. Q: What happens when I pick a story choice on stage? A: After the tap-twice confirm, the AI continues the story in that direction and the result lands as a real branch of the book — with full continuation history. You can read the branch in the reader, or re-enter the theater and keep playing it; the stage recompiles automatically around the new text. Q: How does it decide who is speaking? A: Local structural rules with a hard bias: when unsure, the line goes to the narrator. No LLM calls, no guessing. On a real published novel we hand-annotated dialogue-dense passages and checked every strong attribution — none pointed at the wrong character. Q: Does the stage compilation cost anything? A: Nothing. Compiling prose into scenes, attributing speakers, and wiring up choices are all local rules — offline-capable and free. Money only moves when you explicitly confirm it: generating a scene backdrop (behind a cost dialog) or confirming a story choice that triggers AI continuation. ### TA 第一次给你发自拍：AI 伴侣的脸是怎么固定下来的 URL: https://foreverse.app/zh/blog/the-day-they-sent-a-selfie Published: 2026-07-11 Q: 怎么让伴侣每次自拍尽量都是同一张脸？ A: 靠角色的视觉档案：一组由你亲手确认的参考图，加一段外貌锚点描述。每次生成自拍、时刻卡、合照都带上这套档案。关键纪律是「确认制」——生成得再好的图也不会自动进参考集，必须你点头，防止脸在一次次生成里慢慢漂移。 Q: 一致性到底能做到什么程度？ A: 上线前我们跑过内部盲评：同一角色档案连续生成多张，按五官、发型、气质逐项对照参考集打分。两个主力图像模型分别拿到 10 张里 9 张、10 张里 10 张的达标成绩。诚实说：极端姿态和光线下仍会有偏差，所以每张生成图旁边都有「重新生成」。 Q: 拍合照时我上传的照片会被存储吗？ A: App 不存。你的照片只以内存态发给你自己选择的图像模型供应商完成这一次生成——App 侧不落盘、不进日志，生成结束即释放。合照前有本地成年确认门槛，两个人物的参考顺序在请求里显式绑定，尽量压低生成时张冠李戴的风险。 Q: 伴侣主动发的「照片动态」会乱扣费吗？ A: 有多重闸门。照片动态先在相册写一个幂等占位再调用模型，防止进程重启造成重复付费；每天尝试有上限（失败当日最多再试一次），连续失败自动停，长期不打开应用会静默暂停。要注意：模型调用发生在占位之后，个别失败发生在调用之后（比如落盘失败），这次调用的费用已经产生。费用永远走你自己配置的图像模型，帐单在请求记录里逐条可查。 ### 我们的 LLM 评委，把人类写的文字判成了 AI——88% 的时候 URL: https://foreverse.app/zh/blog/llm-judges-called-humans-ai Published: 2026-07-11 · Updated: 2026-07-14 Q: 什么是「LLM as judge」的 gold 校准？ A: 在让大模型当评委之前，先准备一批人类已有共识答案的样本（金标锚点），测评委在这批样本上的正确率。我们的实验里 4 个模型评委的初始 gold 正确率只有 12%——低于随机（50%），说明评委不是不准，是系统性反向。 Q: 评委间一致率 86% 不是说明结果可信吗？ A: 不。一致性和正确性是两回事：4 个评委高度同意彼此，但同意的方向和人类真实判断相反。多评委投票只能消除随机噪声，消除不了共同的系统性偏见——这是我们这次实验最贵的一课。 Q: 给评委加「反偏见提示」有用吗？ A: 看模型。我们在扩容语料后的第二轮校准里测过：显式反偏提示能把 claude 系评委的 gold 正确率拉到 83%（首轮四评委原始配置只有 12%），说明这类偏差可以被提示纠正；但同样的提示喂给 deepseek、glm、qwen 三个基座，仍停在 25% 到 33%，属于提示纠不动的硬先验。而且 83% 也只是勉强越过我们设的 ≥80% 门槛，仍不足以当自动闸门。 Q: 那「AI 味」到底该怎么检测？ A: 我们现在的组合：规则检测器（词表加句式配额，已对齐真实用户直觉）做回归闸门；真人抽检做真值；LLM 盲评只用于同一份文案改前改后的相对比较，绝不当绝对裁判。诚实的边界是：结构完美、密度均匀的 AI 文案，规则和 LLM 都抓不到，人味没有自动化银弹。 ### 一个角色的四种活法：纸上、台上、对面、身边 URL: https://foreverse.app/zh/blog/character-four-lives Published: 2026-07-11 Q: 「四种活法」之间的数据是真实连通的吗？ A: 是。剧场直接演这本书的当前分支，选择肢续写写回书的分支；角色卡带着世界书进对话；对话中的角色一键升格为伴侣并保留人设来源；伴侣可导出为标准角色卡，勾选的记忆变成随卡世界书。每一步共享同一份数据，不是四个孤立功能的拼盘。 Q: 角色的形象怎么做到前后一致？ A: 靠角色的视觉档案：确认过的参考图集加外貌锚点。书页插画、剧场配景、时刻卡、伴侣自拍都按同一套档案生成，尽量做到换场景不换脸（上线前按 10 张 9 中和 10 中的盲评出口门验收）。参考集由你确认，生成产物不会自动混进去。 Q: 伴侣会知道我这本书读到哪了吗？ A: 不会，这是我们明确没做的一步。伴侣的记忆来自你们的对话和你给的记录，不追踪你的阅读进度——「TA 记得你读到第几章」听起来浪漫，实际是把阅读行为变成了监控。这条线我们不跨。 Q: 把伴侣导出成角色卡送人，对方能看到我们的聊天记录吗？ A: 不能。导出时逐条勾选哪些记忆随卡带走，默认一条都不选；被勾选的记忆以世界书条目形式进卡，聊天记录本身永远不出你的手机。 ### 315 本 txt 一次全选导入：一台真机的阅读器性能账本 URL: https://foreverse.app/zh/blog/315-books-in-8-seconds Published: 2026-07-11 Q: 批量导入几百本书需要多久？ A: 实测 315 本真实 txt（合计约 1GB）一次全选，批量入库约 8 秒，全程无卡死无崩溃。另一台小米真机导入 82 本约 7 秒。导入是「先登记后转换」：书先以待转换状态上架，第一次打开时才做全文解析。 Q: 「待转换」的书第一次打开要等多久？ A: vivo 真机上实测 3.3 到 6.8 秒（含全文章节解析和首屏排版），小米上 9.5 到 14.5 秒。之后打开跳过全文转换走索引，实测小米上已转换书 5 秒左右，具体随设备和书籍体积变化。 Q: 卡顿率 0.49% 是怎么测出来的？ A: 用 Android 系统自带的帧渲染统计（gfxinfo）：在真机上完成整轮导入加随机进书操作后读取累计数据，71320 帧里卡顿帧占 0.49%，99 分位帧耗时 31ms。这是系统口径的数字，不是自报的。 Q: 从旧阅读器搬家，隐藏目录里的书能扫到吗？ A: 能。智能扫描覆盖了常见阅读器的存书目录，包括浏览器下载目录和以点开头的隐藏文件夹；已授权过的目录直接扫，扫描在 315 本规模下秒级出结果。 ### AI 怎么知道这句话是谁说的？说话人归因的五个坑 URL: https://foreverse.app/zh/blog/who-said-that Published: 2026-07-11 · Updated: 2026-07-14 Q: 为什么不直接让大模型判断谁在说话？ A: 成本和确定性。一本 100 章的书按段落调用模型判断说话人，费用会转嫁给用户，而且模型输出不稳定，同一段落两次判断可能给出不同答案。规则方案零成本、离线可用、结果可复现，代价是覆盖率低——我们接受这个代价。 Q: 规则方案的准确率到底怎么验证的？ A: 人工标注。我们从一本三百多万字的已出版网文里摘出 15 段对话密集的段落，逐句人工标注说话人作为标准答案，再跑归因规则对比。强归因部分的精确率是 100%——15 段里没有一处错认；拿不准的句子全部按设计进了旁白。 Q: 归因失败的句子会怎么呈现？ A: 以旁白形式演出，引语原文保留。观感上是「无名牌的台词」，读者靠上下文自己判断——这正是纸书读者一直在做的事，不损失任何信息。 Q: 以后会提高归因覆盖率吗？ A: 会，但底线不变：新增规则必须先过真书标注核对，错误归因零容忍。覆盖率可以慢慢爬，把 A 的话安到 B 头上这种事故一次都不能有。 ### 把书架上的小说一键开成文字游戏：剧场模式实测记录 URL: https://foreverse.app/zh/blog/novel-to-visual-novel Published: 2026-07-11 · Updated: 2026-07-14 Q: 剧场模式需要把小说转换成游戏文件吗？ A: 不需要。剧场直接读这本书当前分支的正文，实时编排成演出，不产生新工程、不迁移数据。退出剧场回到阅读器，进度就停在剧场演到那段所在的页；重进剧场，从上次的位置接着演。 Q: 把正文编排成舞台脚本这一步要花多少钱？ A: 零。编排完全由本地规则完成，不调用任何 AI 模型，离线也能进剧场。只有两件事按你的选择走模型：点「生成此景」出背景图（生成前弹费用确认），以及确认剧情选择肢之后的 AI 续写。 Q: 在剧场里选的剧情，会改动我的原书吗？ A: 以分支形式写入，原文永远保留。二击确认一个走向后，AI 续写的新剧情落成这本书的一条正式分支——回阅读器可以接着读，重进剧场可以接着演，不想要就切回原文线。 Q: 什么书开剧场效果最好？ A: 对话密度高的书：都市、言情、群像、轻小说。散文向的书也能开，只是更多句子会以旁白形式演出。没有配置过角色的书同样能进剧场，全部内容按旁白演。 ### Writing Fanfiction With AI: What Actually Works, From a Reader's Request URL: https://foreverse.app/blog/ai-fanfiction-honestly Published: 2026-07-07 Q: How good is AI at writing fanfiction right now? A: Short pieces (1,000-3,000 words) are genuinely publishable: scene work and dialogue pacing hold up. Long fics are where it breaks — past 20,000 words, characterization and planted details drift unless you support them structurally. The fix is not a better model; it is character dossiers and pinned worldbook entries injected on every request. Q: How do I keep AI from writing characters OOC? A: In order of measured effectiveness: example dialogue (3-5 canon exchanges that carry the character's voice beat any adjective list), negative constraints ('he never apologizes first' prevents more drift than ten positive traits), and generating multiple candidates for key scenes so you judge in-character-ness yourself. The OOC verdict always belongs to you; tools just put candidates side by side. Q: I want three different endings from the same scene. How do I manage that? A: That is exactly what branching is for: fork the story at the divergence point and grow the fix-it, the tragedy, and the what-if line in parallel. Each branch keeps its own full text and context. In single-document tools you juggle three files; in a branch tree they hang off the same origin and you switch between them freely. Q: What about copyright and posting rules for AI-assisted fanfic? A: Fanwork copyright status varies by fandom and platform, and AI involvement does not change that baseline — but most archives (AO3 included) now have disclosure expectations for AI-assisted work, and hiding it gets works reported. Post only where permitted, disclose honestly, never commercialize. Not legal advice; check your archive's current policy. ### 用 AI 写同人文靠谱吗？回答一位想续写耽美同人的等待名单用户 URL: https://foreverse.app/zh/blog/write-fanfic-with-ai Published: 2026-07-07 Q: AI 现在写同人文是什么水平？ A: 旁白叙述和场景描写已经可用，短打（1000-3000 字）质量稳定；长篇的坑在一致性——写到几万字后人设和伏笔会漂。补救靠结构而不是靠模型：把人设做成角色档案、把设定做成世界书常驻条目，每轮都稳定注入，漂移会显著减少。 Q: 怎么让 AI 不 OOC（不写崩人设）？ A: 三件事按重要性排：示例对话（从原作摘 3-5 段最能代表角色语气的对话，比任何形容词都管用）、把「这个角色绝不会做什么」写进档案（负面约束比正面描述更防崩）、重要情节让模型一次给多个候选再由你挑。判断 OOC 的始终是你，工具能做的是把候选摆到一起。 Q: 同一个梗想写好几个走向，怎么管理？ A: 这正是分支续写的场景：从同一个分歧点开多条分支，HE 一条、BE 一条、if 线一条，每条独立生长互不覆盖。单文档工具里你只能存三个文件手动对照；分支树里它们并排挂在同一个原点上，随时切换对比。 Q: AI 写同人的版权和发布要注意什么？ A: 同人创作本身的版权状态因作品和平台而异，AI 参与不改变这一点，但多数发布平台（AO3、LOFTER 等）要求标注 AI 参与程度，隐瞒会被举报。建议：只在允许的分区发布、如实标注、商用零容忍。这不是法律建议，拿不准就查目标平台的最新政策。 ### AI Token 真的越来越便宜吗？看完定价数据后我改了自己的用法 URL: https://foreverse.app/zh/blog/are-ai-tokens-getting-cheaper Published: 2026-07-06 · Updated: 2026-07-14 Q: 「同等能力降价 10 倍」是怎么算出来的？ A: a16z 的 LLMflation 分析取「在固定基准上达到同一分数的最便宜模型」作比较：达到 GPT-3 水平（MMLU 约 42 分）的成本从 2021 年的每百万 token 60 美元跌到 2024 年的 0.06 美元，三年千倍，折合每年约 10 倍。Epoch AI 对多个基准的独立统计给出每年 9 到 900 倍不等的下降区间，取决于任务难度。 Q: 那为什么我的 AI 账单没有降？ A: 三个原因：一，旗舰模型定价基本不随时间下降，降价发生在「旧能力被新的便宜模型复刻」；二，推理模型会产出大量思考 token，单次任务消耗反而上升；三，用量本身在涨——agent 类应用一次任务几十次调用。降价红利属于愿意换模型的人，不属于始终买最贵的人。 Q: 现在（2026 年中）主流模型的价格量级是多少？ A: 以各家官方定价页为准的量级感受：DeepSeek 的 flash 级模型输入价在每百万 token 零点几美元；OpenAI / Anthropic / Google 的旗舰在每百万 token 几美元到十几美元；各家 mini / flash 档普遍比自家旗舰便宜 10 到 25 倍。加上缓存折扣，实际有效价格还能再降一半以上。 Q: 普通用户怎么吃到降价红利？ A: 两条：按场景分配模型——日常续写和闲聊用 flash 档，关键章节和复杂推演才上旗舰，大多数场景 flash 档已经够用；以及用 BYOK 保持随时可换——模型市场一年洗一次牌，绑死在单一订阅上就等于放弃了每年 10 倍的降价曲线。 ### AI Token 缓存为什么越来越重要？一次 60 轮续写实测的账单笔记 URL: https://foreverse.app/zh/blog/why-prompt-caching-matters Published: 2026-07-05 · Updated: 2026-07-07 Q: Prompt 缓存到底缓存了什么？ A: 缓存的是模型对「已经见过的前缀」的中间计算结果（KV Cache）。同一段上下文第二次出现时，供应商不必重新计算这部分注意力状态，直接从缓存加载，所以能以远低于原价的价格计费。缓存按前缀匹配：只要请求开头和上次完全一致，一致的部分就能命中。 Q: 各家的缓存折扣差多少？ A: 以官方定价页为准的量级：DeepSeek 缓存命中的输入约为未命中价的 1/10；Anthropic 缓存读取按基础输入价的 10% 计费（写入缓存加收 25%）；OpenAI 对 1024 token 以上的重复前缀自动打对折；Gemini 提供显式与隐式上下文缓存，命中部分同样按折扣价计。 Q: 为什么长对话反而越聊越便宜？ A: 对话第 N 轮的输入 = 前 N-1 轮的全部历史 + 新一句。历史部分正是上一轮刚出现过的前缀，天然可命中缓存。轮次越多，可缓存的占比越高——第 50 轮时输入的 95% 以上都是折扣价 token，边际成本主要来自新增文字和输出。 Q: 什么操作会把缓存搞失效？ A: 任何改动前缀的行为：往系统提示里插入当前时间戳、每轮重排世界书条目顺序、在历史中间编辑一条旧消息。缓存按前缀严格匹配，前缀变一个字，之后的全部重新计费。工具设计上应把「稳定的内容放前面、易变的内容放后面」。 ### Prompt Caching Explained: What 60 Continuation Requests Taught Us About LLM Bills URL: https://foreverse.app/blog/prompt-caching-explained Published: 2026-07-04 · Updated: 2026-07-07 Q: What exactly does prompt caching cache? A: The model's intermediate attention state (the KV cache) for a prefix it has already processed. When a request starts with the exact same content as a previous one, the provider loads that state instead of recomputing it, and bills the matched portion at a steep discount. Matching is strictly prefix-based: identical from the first character up to the divergence point. Q: How big are the discounts across providers? A: Order-of-magnitude, per official pricing pages: DeepSeek bills cache-hit input at roughly one-tenth of the miss price; Anthropic cache reads cost 10% of base input (writes carry a 25% premium); OpenAI automatically halves input for repeated prefixes over 1,024 tokens; Gemini offers implicit and explicit context caching with discounted hit pricing. Q: Why do long conversations get cheaper per turn? A: Turn N's input is the full history of turns 1..N-1 plus one new message. That history is precisely the prefix the provider just processed, so it hits cache naturally. The deeper the conversation, the higher the cacheable share. By turn 50, over 95% of input tokens are discount-priced, and marginal cost comes almost entirely from new text and output. Q: What breaks caching? A: Anything that mutates the prefix: injecting a per-request timestamp into the system prompt, reordering lorebook entries each turn, or editing an old message mid-history. Matching is exact. Change one character early in the prompt and everything after it bills at full price. Design rule: stable content first, volatile content last. ### C.AI 又改规则了？我们陪一位用户把养了八个月的角色搬了家 URL: https://foreverse.app/zh/blog/cai-refugee-migration-guide Published: 2026-07-03 · Updated: 2026-07-14 Q: C.AI 的角色能直接导出成角色卡吗？ A: 官方没有导出按钮，角色定义（尤其别人创建的角色）拿不到原文。可行的路径是重建：把角色的公开简介、你记忆里的性格要点、加上从聊天记录里摘出的经典对话，按 chara_card_v3 格式重写一张卡。社区也有第三方抓取工具，但依赖非官方接口，随时可能失效，重建反而更可控。 Q: 聊了几千条的记忆怎么办？ A: 全量搬运没有意义——模型每轮能看的上下文有限，几千条原始记录塞不进也不需要。正确做法是蒸馏：把关系的关键节点（怎么认识的、几次重要冲突与和解、彼此的称呼和梗）整理成一两千字的「关系简历」，写进角色卡的场景设定或世界书常驻条目。新平台上的第一句话，就能接住八个月的感情线。 Q: 换到 BYOK 应用后聊天质量会下降吗？ A: 取决于你选的模型，上限通常更高：C.AI 用的是它自家的专有模型，你没得选；BYOK 下 DeepSeek、Gemini、Claude 随便换，角色扮演口碑好的模型可以按卡分配。落差主要在两点：需要自己配 Key（一次性成本），以及没有平台内置的社区推荐流——角色要自己找、自己养。 Q: 怎么避免下一次「被迁移」？ A: 认准三个可迁移标志：角色以开放格式存储（chara_card_v3 的 PNG/JSON 文件，而不是平台数据库里的一行记录）；聊天记录可以一键导出成通用格式；模型接入走你自己的 Key。三条都满足时，换应用只是换个壳，你的角色、记忆、模型关系全部原地带走。 ### Are AI Tokens Getting Cheaper? The Pricing Data Says Yes and Your Bill Says No URL: https://foreverse.app/blog/llm-token-price-trends Published: 2026-07-02 · Updated: 2026-07-14 Q: Where does the “10x cheaper per year” figure come from? A: a16z's LLMflation analysis tracks the cheapest model that reaches a fixed benchmark score: hitting GPT-3-level MMLU (~42) cost $60 per million tokens in 2021 and $0.06 by late 2024, a 1,000x drop in three years, roughly 10x per year. Epoch AI's independent study across multiple benchmarks finds annual declines between 9x and 900x depending on task difficulty, with ~50x for the median. Q: Then why doesn't my AI bill go down? A: Three reasons. Frontier list prices barely fall; the decline happens when new cheap models replicate old capabilities. Reasoning models emit large volumes of billed thinking tokens, raising per-task consumption 5-20x. And usage itself grows; agent workloads fire dozens of calls per task. The dividend goes to people willing to switch models, not to people who always buy the most expensive one. Q: What do prices look like in mid-2026? A: Order-of-magnitude, per official pricing pages: flash-tier models (DeepSeek, Gemini Flash class) charge tenths of a dollar per million input tokens; frontier models from OpenAI, Anthropic and Google sit at a few to tens of dollars per million; every vendor's mini/flash tier runs 10-25x cheaper than its own flagship. Prompt-cache discounts cut effective prices by half or more on top. Q: How does a regular user capture the decline? A: Tier your scenarios: run everyday continuation and casual chat on flash-tier models and reserve the flagship for pivotal chapters and complex reasoning. Flash tiers are already sufficient for most fiction work. And stay switchable with BYOK: the model market reshuffles yearly, and locking into a single subscription forfeits the 10x/year curve entirely. ### 世界书塞了 237 条设定，AI 还是叫错名字：一次完整的排障记录 URL: https://foreverse.app/zh/blog/lorebook-writing-guide Published: 2026-07-01 · Updated: 2026-07-14 Q: 世界书和角色卡的 description 有什么区别？ A: description 是常驻的：每一轮对话都完整占用上下文，适合放角色的核心人设。世界书默认是按需的：只有对话里出现关键词才注入对应条目，适合放数量大、但不是每轮都用得上的设定——地名、支线人物、历史事件。把所有设定都塞进 description，等于每轮都在为用不上的信息付 token。 Q: 世界书条目没触发，怎么排查？ A: 按顺序查四点：一，关键词是否真的出现在最近几轮对话里（触发靠扫描最近消息，不是全文）；二，别名有没有覆盖——对话里叫「老李」，关键词只写了「李云山」就不会触发；三，扫描深度是否太浅，关键词出现在更早的楼层就扫不到；四，token 预算是否被前面的条目吃满，后面的条目被静默丢弃。 Q: 一本世界书多大算太大？ A: 条目数量本身不是问题，几百条也没关系——问题在单轮注入量。触发式条目每轮实际注入的总量受 token 预算限制，常驻（constant）条目则每轮必进。经验值：常驻条目控制在 5 条以内，单条内容 200 字以内；触发条目单条不超过 300 字。写满 2000 字的「王国编年史」不如拆成 10 条各管一段。 Q: V2 和 V3 的世界书有什么区别？ A: V3（chara_card_v3）在条目上增加了 decorators——以 @@ 开头的指令，可以精确控制注入深度、触发条件等行为，世界书的表达力更强。V2 卡的世界书在支持 V3 的应用里可以正常使用，反过来 V3 特有的 decorators 在旧应用里会被忽略。导入前确认应用支持的规范版本，能省掉很多「为什么行为不一致」的困惑。 ### Turn Any EPUB Into an AI Audiobook: Three Weeks of Commute Testing, Honestly Reported URL: https://foreverse.app/blog/epub-to-ai-audiobook Published: 2026-06-29 · Updated: 2026-07-07 Q: How good is AI narration compared to a human audiobook in 2026? A: For single-narrator prose it is close enough that most listeners stop noticing after a chapter. In our blind test, three out of five colleagues could not tell a neural TTS voice from a human narrator on a 2,000-word passage. The remaining gap is multi-character dialogue: human narrators switch voices per character; most TTS pipelines still read everything in one voice. Q: What does it cost to listen to a whole novel with TTS? A: TTS is billed per character. A 3,000-word chapter runs roughly 16,000-18,000 characters; at typical provider rates that is a fraction of a cent to a few cents per chapter depending on the voice tier. A full 300-chapter webnovel usually lands in the single-digit dollars, with your own API key, at provider list price, no subscription on top. Q: Why does my TTS app read out asterisks and chapter dividers? A: Because it sent the raw text to the speech engine. EPUB and TXT files from the wild are full of things that should never be spoken: scene-break dashes, translator notes, promo lines, emoji. An app that cares about listening strips these before synthesis. If you hear a warm voice solemnly pronounce three asterisks, the app skipped its homework. Q: Can reading and listening share one progress? A: They should, and this is the feature that decides whether TTS is a gimmick or a habit. You listen through chapter 230 on the commute, open the book at home, and the page picks up where your ears left off. If the app keeps two separate positions, you will spend every evening scrolling to find yourself. ### Lorebook Not Triggering? Three Support Tickets and What Fixed Each One URL: https://foreverse.app/blog/lorebook-world-info-not-triggering Published: 2026-06-24 · Updated: 2026-07-27 Q: How do I activate a lorebook in SillyTavern? A: Two switches, both required. First bind the book: pick it as an active world in the World Info panel, or attach it to the character (embedded books need the import prompt accepted once). Then make sure each entry has trigger keys and is enabled — entries fire when a key appears in the recent messages, so say a keyword in chat to test. If it still stays silent, paste the book and a chat line into our free lorebook trigger simulator and it names the exact reason per entry. Q: How does a lorebook decide which entries to inject? A: On every message, the app scans the most recent turns of the conversation for each entry's keywords. Matching entries are activated and inserted into the prompt at their configured position, in priority order, until the lorebook's token budget is spent. Entries beyond the budget are dropped silently; the model never knows they existed. Q: Why does my entry trigger sometimes but not always? A: Almost always scan depth. Keywords are matched against the last N messages, not the whole chat. If the dragon was last mentioned six turns ago and scan depth is four, the entry sleeps. Raise the depth, add the keyword to related entries so recursion keeps it warm, or make the entry constant if it genuinely matters every turn. Q: Is a big lorebook bad for quality? A: Entry count is fine; per-turn injection volume is what hurts. Constant entries cost context every single turn, so keep them under five, each under ~150 words. Triggered entries compete for one token budget, and a single 800-word essay entry can starve three concise ones. Split epics into one-fact-per-entry cards. Q: Do V2 lorebooks work in V3 apps? A: Yes. chara_card_v3 is backward compatible: V2 world info imports and runs unchanged. V3 adds decorators, @@-prefixed directives on entries that control injection depth and activation conditions. Those extras are ignored by V2-only apps, which is the main source of 'same card, different behavior' reports. ### 角色卡导入失败的 9 个原因：PNG、JSON、世界书一次排查清楚 URL: https://foreverse.app/zh/blog/character-card-import-troubleshooting Published: 2026-06-05 Q: 为什么微信传过来的角色卡导不进去？ A: 微信和 QQ 默认压缩图片，压缩过程会重新编码 PNG，把藏在 tEXt 块里的角色数据整块抹掉。让对方以「文件」方式发送原图，或者打包成 zip 再传，就能保住卡片数据。 Q: 怎么快速判断一张 PNG 卡里有没有数据？ A: 看文件大小。一张正常的角色卡 PNG 通常在 100KB 到 3MB 之间；如果一张「精美立绘」只有几十 KB，多半已经被某个传输环节重新压缩过，卡片数据凶多吉少。最稳的办法是导入一次试试，或用文本编辑器打开搜索 chara 关键字。 Q: v1 老卡还能用吗？ A: 大部分现代客户端（包括 SillyTavern 和 Foreverse）仍然向下兼容 v1 卡的基础字段，但 v1 没有备选问候语、角色书等结构，体验会缺一块。建议用编辑器把老卡升级到 v3 再用。 Q: 角色卡带的世界书为什么没生效？ A: 看两个地方：一是导入时世界书是否被一并导入（部分客户端需要手动确认），二是世界书是绑定到角色还是全局启用。绑定角色的世界书只在和该角色聊天时注入，全局的才会处处生效。 ### How to Run SillyTavern on Android or iPhone in 2026: Termux, Cloud, or Native App URL: https://foreverse.app/blog/run-sillytavern-on-android-iphone Published: 2026-06-03 · Updated: 2026-10-01 Q: Can SillyTavern run natively on an iPhone? A: No. SillyTavern is a Node.js server application, and iOS does not allow long-running server processes. iPhone users have two realistic options: open a cloud- or PC-hosted instance in Safari, or use a natively built tavern-style app that reads the same character cards — Foreverse, for example, is on the App Store and reads chara_card_v3 PNG/JSON/.charx files and lorebooks. Q: Why does my Termux SillyTavern stop when the screen locks? A: Android freezes background processes to reclaim memory, and a Node.js server inside Termux is a prime target. termux-wake-lock and disabling battery optimization help, but aggressive vendor ROMs often override both. It is a platform behavior, not a SillyTavern bug. Q: Will my character cards work across all four routes? A: Yes — PNG and JSON character cards plus lorebook JSON files are portable everywhere. Chat history is stored as jsonl inside the SillyTavern ecosystem; native apps vary, so check that the app explicitly imports SillyTavern data before you commit. Q: Is a shared cloud tavern safe to use? A: Only as safe as the person running it. On a shared instance, your API keys, cards, and chat logs all pass through someone else's server. Self-host if you can; if you cannot, treat any service that asks for your API key with appropriate suspicion. ### 2026 年手机玩 SillyTavern 酒馆的四条路：Termux（和免终端壳子）、云端、局域网、原生 App URL: https://foreverse.app/zh/blog/sillytavern-on-phone-2026 Published: 2026-06-02 · Updated: 2026-10-01 Q: Termux 跑酒馆，锁屏后为什么会断？ A: Android 会冻结后台进程回收内存，Termux 里的 Node.js 服务首当其冲。可以用 termux-wake-lock 和关闭电池优化来缓解，但厂商 ROM（尤其国产定制系统）的激进省电策略经常绕过这些设置。这是系统机制，不是酒馆的 bug。 Q: iPhone 能本地跑 SillyTavern 吗？ A: 不能。SillyTavern 是 Node.js 应用，iOS 不允许运行这类常驻服务进程。iPhone 用户只有两条路：浏览器访问部署在云端或家里电脑上的酒馆，或者使用原生开发的酒馆类 App，比如新梦 iPhone 版已在 App Store 上架（中国大陆区暂未上架），能直接导入酒馆角色卡和世界书。 Q: 换路线后，原来的角色卡和聊天记录能带走吗？ A: 角色卡（PNG / JSON）和世界书是通用文件，四条路线之间都能互相导入。聊天记录在 SillyTavern 体系内是 jsonl 文件，可以备份迁移；迁去原生 App 则取决于对方是否兼容 SillyTavern 的数据格式，导入前先确认。 Q: 云端酒馆安全吗？ A: 取决于谁在运维。自己买 VPS 自己部署，数据归你管；用别人搭好的「公共酒馆」，你的 API Key、角色卡和聊天记录都过他的服务器，这层信任要自己评估。任何要求你把 Key 填进陌生网页的服务都值得多想三秒。 ### Why Your AI Roleplay Partner Forgets Everything — and What Actually Fixes It URL: https://foreverse.app/blog/why-ai-roleplay-chats-forget Published: 2026-05-27 Q: Why do AI characters forget things from earlier in the chat? A: Every model reads a fixed budget of text per reply — the context window. Once a chat outgrows it, something must be dropped, and naive clients silently drop your oldest messages. The model never 'knew' your whole story; it only ever saw the slice that fit. Q: Does a bigger context window solve roleplay memory? A: It delays the problem and raises the bill, but does not solve it. A 200k-token window still fills after weeks of play, costs scale with every token you resend, and models demonstrably pay less attention to the middle of very long contexts. Structure beats brute force. Q: What is a lorebook and how does it help memory? A: A lorebook is a set of keyed entries — facts about people, places, and events — that get injected into the prompt only when relevant keywords appear in the recent conversation. It acts as cheap, targeted long-term memory: the dragon's death costs zero tokens until someone mentions the dragon. Q: Can I keep my chat history if a platform shuts down? A: Only if your history lives somewhere you control. Platform-hosted services have deleted chats with policy changes or shutdowns before. Clients that store chats on your device in open formats (like SillyTavern's jsonl) make your story archive independent of any company's roadmap. ### AI 续写小说为什么总把书写崩？我们换了一种结构来解决 URL: https://foreverse.app/zh/blog/ai-continue-novel-branches Published: 2026-05-26 Q: AI 续写为什么会忘记前文设定？ A: 模型一次能看到的文字量（上下文窗口）是有限的，几十万字的书不可能整本塞进去。朴素做法只送最近几千字，前面埋的伏笔自然全丢。解法是给书建立结构化的设定档案，续写时按相关性挑选注入，而不是只看「最近的文字」。 Q: 分支续写和普通的「重新生成」有什么区别？ A: 重新生成是覆盖：新结果顶掉旧结果，旧的就没了。分支是并存：每个方向都保留完整的文字和上下文，可以随时回头、对比、继续往深里写。前者像抽卡，后者像创作。 Q: 续写能模仿原作者的文风吗？ A: 能接近，做不到完美。靠谱的做法是让模型从原文中学习用词、节奏和对话密度，并允许你按章节微调提示词。任何宣称「百分百还原文风」的产品都在夸大。 Q: 烂尾的连载文能用 AI 救回来吗？ A: 可以试，而且这是分支续写最合适的场景之一：从烂尾点开一条分支，把你心中「本该有的结局」写出来；不满意就再开一条。原文永远完好，你失去的只是一点电费。 ### Character Card Formats Explained: v1, v2, and chara_card_v3 — What Actually Imports URL: https://foreverse.app/blog/character-card-formats-explained Published: 2026-05-20 Q: How is character data stored inside a PNG? A: As base64-encoded JSON inside the PNG's tEXt metadata chunks — 'chara' for v2 data and 'ccv3' for v3. The image you see is just the portrait; the character lives in metadata. Any pipeline that re-encodes the image (messenger compression, photo-library 'optimization') strips those chunks and kills the card. Q: What did v2 add over v1? A: v1 was six flat fields: name, description, personality, scenario, first message, and example dialogue. v2 wrapped them in a versioned envelope and added the fields the community had been hacking in: alternate greetings, system prompt overrides, creator notes, tags, and most importantly an embeddable character book (mini-lorebook). Q: Is chara_card_v3 backwards compatible? A: Largely yes by design — v3 keeps the v2 envelope shape, so a v3 card usually degrades gracefully in a v2-only client (new fields like assets, multilingual creator notes, and decorators are simply ignored). A v2 card in a v3 client just works. Q: Should creators distribute PNG or JSON? A: Both. PNG for collectors and galleries — it carries the portrait. JSON for reliability — it survives chat apps, email, and re-uploads that would strip PNG metadata. Label the spec version and note whether a character book is embedded; it preempts most 'card won't import' reports. ### BYOK 不是中转：在手机上自带 API Key 的正确姿势 URL: https://foreverse.app/zh/blog/byok-on-mobile Published: 2026-05-19 Q: BYOK 和 API 中转有什么区别？ A: BYOK 是你直接持有官方（或你信任的供应商）的 Key，应用拿它直连该供应商的服务器；中转是把别人的 Key 二次销售给你，你的请求全部过中转商的服务器。前者的信任链是「你—供应商」两点，后者多了一个可以看到你全部对话、随时可能跑路的中间人。 Q: 怎么验证一个 App 是不是真 BYOK？ A: 看三点：一，填 Key 时是否允许你自定义 API 地址（真 BYOK 必然支持）；二，断开它家账号体系后 BYOK 功能是否照常工作；三，有条件的话抓包看请求是否直达供应商域名。只要 Key 需要「上传到我们的服务器托管」，就不是本文说的 BYOK。 Q: 手机丢了，存在里面的 Key 会泄露吗？ A: 取决于应用怎么存。合格的做法是用系统级加密（Android Keystore 这类硬件背书的方案）落盘，应用数据导出后没有密钥也解不开。再加一层保险：给重要 Key 设消费上限，丢失后第一时间在供应商后台吊销。 Q: BYOK 会比官方 App 便宜吗？ A: 对重度用户通常是。你按供应商的原始价格付费，没有订阅费和中间加价；轻度用户则可能不如包月划算。BYOK 的核心收益其实不是省钱，是选择权——60 多家供应商、上百个模型，按场景挑最合适的。 ### BYOK, Actually: How to Tell If an App Really Keeps Your API Key on Device URL: https://foreverse.app/blog/what-byok-actually-means Published: 2026-05-13 Q: What does BYOK mean in AI apps? A: Bring Your Own Key: you create an API key with a provider (OpenAI, Anthropic, DeepSeek, etc.), the app stores it on your device, and every request goes from your device directly to that provider. You pay raw provider prices. If the key or your conversations route through the app vendor's servers, it is proxy billing wearing a BYOK costume. Q: How do I verify an app is genuinely BYOK? A: Three tests: (1) it lets you set a custom API base URL, not just paste a key; (2) BYOK chats keep working when you log out of the app's own account system; (3) if you can, watch the traffic — requests should hit the provider's domain, not the vendor's. Any app that 'syncs your key to our cloud' fails the definition. Q: Where should a mobile app store API keys? A: In hardware-backed encrypted storage (Android Keystore / iOS Keychain), excluded from cloud backups, never written to logs or analytics. A copied app-data folder without the device's hardware key should be useless to an attacker. Q: Is BYOK cheaper than a subscription? A: For heavy users, usually — you pay provider list price with no middleman margin. For light users a flat subscription can win. The durable benefit is not price but leverage: when one provider raises prices or degrades a model, you switch endpoints in a minute instead of migrating your life. ### 通勤路上「听」完 300 万字网文：AI 朗读的实测笔记 URL: https://foreverse.app/zh/blog/listen-to-webnovels-tts Published: 2026-05-12 · Updated: 2026-07-14 Q: AI 朗读和真人有声书差距还大吗？ A: 旁白叙述部分已经非常接近，长时间听不出疲惫感；差距主要在多角色对手戏——真人演播会为每个角色切换声线，AI 朗读目前多数还是单音色到底。听爽文、种田文几乎无感，听群像戏会想念真人。 Q: 听书很费流量吗？ A: 语音合成按字数计，一章 3000 字大约对应几百 KB 到 2MB 音频（取决于音质档位）。Wi-Fi 下提前缓存整卷，地铁里零流量播放，是目前最省的方案。 Q: 网文里的错别字和怪符号会被读出来吗？ A: 处理得好的应用会在送往语音引擎前清洗文本：剥掉分割线、广告尾巴、表情符号，多数错别字模型会按上下文读出正确发音。生僻人名偶尔翻车，听两章就习惯了。 Q: 可以边听边看吗？ A: 可以，这其实是被低估的用法：眼睛跟着高亮走，耳朵听声音，注意力比单独看或单独听都集中。午休躺着听、通勤站着看，一本书两种姿势无缝接力。 ### Rerolling Is Gambling. Branching Is Writing. A Better Way to Use AI on Long Fiction URL: https://foreverse.app/blog/branching-beats-rerolling Published: 2026-05-06 Q: What is branching in AI fiction? A: Instead of regenerating over a passage you dislike, you fork the story at any paragraph and grow an alternative line alongside the original. Every branch keeps its own full text and context. It is version control applied to narrative: nothing is overwritten, everything is comparable. Q: Why is rerolling worse than branching? A: Rerolling is destructive — each new roll replaces the last, so a great take you rolled past is gone forever. It also trains writers into slot-machine behavior: pull the lever, hope, repeat. Branching keeps every candidate alive, so choosing a direction is an editorial decision instead of a gamble. Q: Does branching work on existing novels, not just AI-written stories? A: Yes — that is arguably its best use. Import a finished or abandoned book, fork from any chapter, and explore the ending the author never wrote. The original text stays untouched; each what-if lives on its own line. Q: Doesn't keeping every branch get messy? A: Only if the tool hides structure. With a visible tree — where each line split, how far it has grown, which branch you are on — a dozen live branches stay navigable. The mess in single-line tools comes precisely from not having that map. ======================================================================== ## TASK TUTORIALS — full visible answers, steps, safety notes, troubleshooting, and FAQs ### Foreverse in 10 Minutes: Model, Book, Continuation, and Character Chat URL: https://foreverse.app/docs/getting-started Platform: Android (Play and site builds) · App: v76+ · Path verified: 2026-08-10 Direct answer: Open Me → AI Models & Services and configure either the official channel or your own API key. Then use Home to import a book, open it and select text to continue; import a character-card file from Home and open it to start a chat. The current bottom navigation is Home, Discover, and Me; characters live inside Home and Agent work starts from Quick Dock. Before you start: - Foreverse v76 or later installed on Android. - A txt/epub book and, optionally, a PNG/JSON/.charx character card stored on the phone. - Either a signed-in official-channel account with credits or an API key from a supported provider. Steps: 1. Choose how AI calls are paid: Open Me → AI Models & Services. Sign in to use the metered official channel, or open a provider, paste your own key, run Test connection, enable a model, and select it under Default model configuration → Text. Success: A text model appears as the active default and its provider is shown without an error badge. 2. Import the first book: Open Home, tap +, choose Import a novel, and select the txt or epub in the system file picker. Then tap the new shelf item and let the first-open conversion finish. Success: The book opens in the Reader and its chapters appear in the table of contents. 3. Run a first continuation: Long-press text near the point where the story should continue, adjust the selection, tap Continue here, add a one-sentence direction or leave it blank, confirm the model, and start generation. Success: New text streams into a visually distinct AI segment while the imported source stays unchanged. 4. Import a character card: Return to Home, choose the character/card import action, and select the original PNG, JSON, or .charx file. Open the imported character and choose Start chat. Success: The greeting appears and a message can be sent using the selected text model. 5. Know the three main entrances: Use Home for books, characters, companions, and your library; Discover for public content; Me for models, account, settings, data, and support. Use the floating Quick Dock to ask the Agent for help or run a task. Success checks: - A text model is selected and can complete a connectivity check or real generation. - The imported book and character remain visible after leaving and reopening Home. - The continuation is an AI segment/branch and the original imported text is still intact. - The character chat opens with the card greeting and accepts a message. Cost and data safety: - Importing, reading, and organizing local files do not call a model. A connectivity test or generation may make a real provider request. - Official-channel calls spend credits; 1 credit = US$0.0001. BYOK calls are billed by the provider attached to your key. - Books, cards, and chat history are local-first. The selected context is sent to the model provider only when you request AI output. - Keep original card and book files until you have verified the import. Do not uninstall or switch between differently signed Play/site builds before exporting data. Troubleshooting: - No model is available in the continuation or chat form — Cause: The provider is not connected, the model is disabled, or no text default is selected. Fix: Go to Me → AI Models & Services, test the provider, enable a text-capable model, then set it under Default model configuration → Text. - The imported txt has garbled text or bad chapters — Cause: The file encoding or chapter headings were not recognized as expected. Fix: Re-export as UTF-8 txt or use epub with a table of contents, then import the corrected copy as a new item. - A PNG character card imports as a normal image or fails — Cause: Messaging/social compression stripped the embedded card metadata. Fix: Ask for the original file as a document, or use the JSON/.charx version. Do not resave the PNG in a photo editor. - 401 / unauthorized — Cause: The API key is invalid, expired, missing, or belongs to a different host. Fix: Re-copy the key from the provider console, confirm the selected provider/host, and test again without posting the key in support logs or screenshots. FAQs: Q: Do I need an API key before I can use Foreverse? A: No. You can sign in and use the official metered channel. BYOK is optional and lets you pay the provider directly with your own key. Q: Does a continuation overwrite the imported novel? A: No. AI text is kept as a distinct segment/branch. You can delete or regenerate that segment while the imported source stays intact. Q: Where did the old Character and Agent tabs go? A: Current builds use Home, Discover, and Me. Characters are in Home, while Agent tasks and app guidance are available through Quick Dock and its workbench. ### 10 分钟上手 Foreverse：配模型、导书、第一次续写与导卡开聊 URL: https://foreverse.app/zh/docs/getting-started Platform: Android（Play 与官网版） · App: v76+ · Path verified: 2026-08-10 Direct answer: 先在「我的→AI 模型与服务」选好文本模型，再从主页导书、划词续写并导卡开聊。生成可能计费，原书和角色数据本地优先保存。 Before you start: - Android 已安装 Foreverse v76 或更高版本。 - 手机里准备一本 txt/epub；想体验角色聊天时，再准备原始 PNG/JSON/.charx 角色卡。 - 二选一：登录后使用有积分的官方渠道，或准备受支持供应商的 API Key。 Steps: 1. 先决定 AI 调用怎么付费: 打开「我的 → AI 模型与服务」。不想配 Key 就登录走官方按量渠道；要 BYOK 就打开供应商、粘贴 Key、点「测连通」，启用一个模型，再到「默认模型配置 → 文本」选中它。 Success: 文本默认模型显示为已选，供应商没有错误标记。 2. 导入第一本书: 进入「主页」，点「+ → 导入小说」，在系统文件选择器里选中 txt 或 epub。回到书架点新条目，等待首次打开时的转换完成。 Success: 阅读器能打开，目录里能看到识别出的章节。 3. 完成第一次续写: 在想接着写的位置长按选中文字，调整选区后点「续这里」。用一句话说明走向，也可以留空；确认模型后开始生成。 Success: 新文字流式出现，并以 AI 段样式和原文区分；导入的原文没有被覆盖。 4. 导入角色卡并开聊: 回到主页，选择角色卡导入，选原始 PNG、JSON 或 .charx。打开导入后的角色，点「开始聊天」。 Success: 开场白正常出现，并且可以用刚才选择的文本模型发出消息。 5. 记住三个主入口: 主页放书、角色、伴侣和本地资产；发现用于浏览公开内容；我的放模型、账号、设置、数据和支持。需要 Agent 帮忙或教你用 App 时，打开悬浮 Quick Dock。 Success checks: - 已选文本模型，并且测连通或真实生成成功。 - 离开再回主页，导入的书和角色仍然在。 - 续写落在 AI 段/分支，导入原文仍完整。 - 角色聊天显示卡的开场白，并能发送消息。 Cost and data safety: - 导入、阅读和整理本地文件不会调用模型；测连通或生成可能发出一次真实供应商请求。 - 官方渠道会扣积分，1 credit = 0.0001 美元；BYOK 由你的 Key 所属供应商直接计费。 - 书、卡和聊天记录本地优先保存；只有你发起 AI 操作时，所选上下文才会发送给所选供应商。 - 确认导入成功前保留原始文件。导出数据前不要卸载，也不要直接在签名不同的 Play/官网版之间切换。 Troubleshooting: - 续写或聊天表单里没有模型 — Cause: 供应商未连通、模型未启用，或没有选择文本默认模型。 Fix: 去「我的 → AI 模型与服务」测连通，启用文本模型，再到「默认模型配置 → 文本」选中它。 - txt 乱码或章节识别不对 — Cause: 文件编码或章节标题格式没有被正确识别。 Fix: 把文件另存为 UTF-8 txt，或改用带目录的 epub，再把修正后的文件作为新项目导入。 - PNG 角色卡导入失败或只被当普通图片 — Cause: 微信等平台压缩图片时清掉了卡的内嵌数据。 Fix: 让作者以文件方式发送原图，或索要 JSON/.charx；不要用相册编辑器重新保存。 - 401 / unauthorized / 未授权 — Cause: Key 无效、过期、漏填，或 Key 与 API 地址不属于同一供应商。 Fix: 从供应商控制台重新复制 Key，核对供应商与地址后再测；不要把 Key 放进客服日志或截图。 FAQs: Q: 不配 API Key 能用吗？ A: 可以。登录后可走官方按量渠道；BYOK 只是可选项，费用由供应商直接收取。 Q: 续写会覆盖导入的小说吗？ A: 不会。AI 文字保存在独立段落/分支中，可以删除或重生，导入原文保持不动。 Q: 以前的「角色」和「智能体」底栏去哪了？ A: 当前版本底栏是主页、发现、我的。角色已经进入主页；Agent 任务和 App 导览从 Quick Dock 及其工作台进入。 ### Import and Continue a Novel Without Overwriting the Original URL: https://foreverse.app/docs/guide-reader/import-and-continue-a-novel Platform: Android Reader · App: v76+ · Path verified: 2026-08-10 Direct answer: Import the txt or epub from Home, open it in Reader, long-press the sentence where the new route should begin, and choose Continue here or New branch. The generated text is stored separately from the imported source, so you can delete, regenerate, or switch branches without rewriting the original. Before you start: - A readable txt (UTF-8 recommended) or epub file on the device. - A connected text model through the official channel or BYOK. - Enough provider quota or official credits for at least one generation. Steps: 1. Import from Home: Open Home → + → Import a novel and select the txt or epub in the system file picker. Open the new shelf item, let first-open conversion finish, and check the table of contents before generating. 2. Pick an exact insertion point: Navigate to the chapter, long-press near the point where the story should diverge, and adjust the selection to a complete sentence. Choose Continue here for an inserted AI segment or New branch for a larger rewrite from that point. 3. Set direction and scope: Write a short direction such as ‘she refuses the offer and follows the river north.’ Confirm the text model, template, output length, and whether to generate three next-step options. 4. Generate and review: Start generation and keep the Reader open until the stream settles. Read the join between source and generated prose, then use edit/regenerate if the tone or facts are wrong. 5. Preserve alternatives: Use the AI-segment candidate controls for another version. To switch branches, open Contents and tap the ⑂N branch badge beside the relevant chapter. Confirm the active route before continuing many chapters. Success checks: - The table of contents matches the imported book closely enough to navigate. - Generated prose has AI styling and the source paragraph remains unchanged. - Contents shows a ⑂N branch badge for the chapter where a new route was created. - Switching back to the source/main route restores the original continuation point. Cost and data safety: - File import and branch switching are local and do not cost credits. Each continuation, option generation, or regeneration is a separate model call and may be billed. - Stopping or losing the connection does not prove that the provider charged nothing; check API request records before retrying repeatedly. - The original import stays intact, but generated branches are app data. Export important work before uninstalling or switching app distribution channels. Troubleshooting: - Chapters are missing or split incorrectly — Cause: TXT headings are irregular or the epub table of contents is malformed. Fix: Fix headings and re-export as UTF-8, or repair the epub TOC. Import the repaired file as a new book rather than assuming the existing parse changed. - Continue starts from the wrong place — Cause: The selection/viewport was not at the intended sentence, or a tail action was used from another chapter. Fix: Cancel before generation if possible, return to the target chapter, select the exact sentence, and use the selection action. Delete the wrong AI segment only after checking which branch it belongs to. - The model ignores earlier facts — Cause: The context budget is too small, the relevant lore is outside the window, or the instruction conflicts with the source. Fix: Shorten the requested output, raise context within model limits, add the crucial fact to the direction/worldbook, and regenerate once. - 402 / insufficient credits — Cause: The official-channel balance cannot cover the request. Fix: The failed attempt should not create content. Recharge, choose a cheaper official model, or switch to a configured BYOK model, then retry from the preserved draft. - 429 / rate limit — Cause: The provider is throttling requests or quota is exhausted. Fix: Wait for the provider window to reset, lower concurrency, or choose another enabled model. Do not hammer regenerate because every accepted request may be billable. FAQs: Q: Should I use Continue here or New branch? A: Use Continue here for an inserted passage on the current route. Use New branch when you want the story after this point to diverge while keeping the old route readable. Q: Can I continue from the end of an unfinished book? A: Yes. Go to the final readable point and use the tail continuation/selection action. You can keep several possible endings as separate branches. Q: Can I undo an AI continuation? A: Delete the AI segment or switch back to the earlier branch. Regeneration keeps alternatives as candidates where supported, but exported backups are still recommended for valuable work. ### 导入并续写一本小说，同时保留原文 URL: https://foreverse.app/zh/docs/guide-reader/import-and-continue-a-novel Platform: Android 阅读器 · App: v76+ · Path verified: 2026-08-10 Direct answer: 从当前「主页→+→导入小说」选择本地 TXT/EPUB，打开后划词点「续这里」。生成可能计费，AI 段另存为分支，不会覆盖导入原文。 Before you start: - 手机里有可读取的 txt（建议 UTF-8）或 epub。 - 官方渠道或 BYOK 已连通一个文本模型。 - 供应商额度或官方积分足够完成至少一次生成。 Steps: 1. 从主页导入: 打开「主页 → + → 导入小说」，在系统文件选择器里选中 txt 或 epub。打开新书，等待首次转换完成，先查看目录是否基本正确，再开始生成。 2. 精确选中插入位置: 跳到目标章节，在故事应当分叉的位置长按并调整到完整句子。小范围接写用「续这里」；想从此处大改后文用「新建分支」。 3. 设置方向与范围: 写一句明确方向，例如「她拒绝邀请，沿河向北」。确认文本模型、提示词模板、输出长度，以及是否同时生成三个下一步候选。 4. 生成后检查衔接: 开始生成，等流式输出正常收尾。检查原文与新段之间是否连贯；文风或事实不对时，用「编辑并重新生成」修改指令。 5. 把不同方案留下来: 用 AI 段候选切换另一版。要换分支时打开「目录」，点对应章节旁的 ⑂N 分支徽标。长篇继续前，先确认自己当前在哪条线。 Success checks: - 目录足以正确跳章。 - AI 内容有独立样式，原段落没被覆盖。 - 新建分支后，对应章节出现 ⑂N 分支徽标。 - 切回原线后，原来的正文仍在。 Cost and data safety: - 导入文件和切分支不花积分；每次续写、生成候选或重生都是独立模型调用，可能分别计费。 - 停止或断网不等于供应商一定没扣费；连续重试前先看 API 请求记录。 - 原始导入会保留，但生成分支属于 App 数据。卸载或切换分发渠道前，先导出重要作品。 Troubleshooting: - 章节缺失或切分错乱 — Cause: TXT 标题不规则，或 epub 目录本身有问题。 Fix: 统一章节标题并另存 UTF-8，或修复 epub 目录；把修正文件作为新书导入，不要假定旧解析会自动变化。 - 续写从错误位置开始 — Cause: 选区/视口不在目标句，或在别章用了段尾续写。 Fix: 能取消时先取消，回到目标章精确划词后从选区发起。删错段前先确认它属于哪条分支。 - 模型忘了前文设定 — Cause: 上下文预算不足、关键设定在窗口外，或指令与原文冲突。 Fix: 缩短输出、在模型能力范围内提高上下文，把关键事实写进指令或世界书，再重生一次。 - 402 / 积分不足 — Cause: 官方渠道余额不足以覆盖这次请求。 Fix: 失败时应当没有新内容。充值、换便宜官方模型，或切到已配置 BYOK 模型，再从保留的草稿重试。 - 429 / 请求过多 — Cause: 供应商限流或额度用完。 Fix: 等限流窗口恢复、降低并发或换另一个已启用模型。不要连续狂点重生，已被上游接收的请求可能分别计费。 FAQs: Q: 「续这里」和「新建分支」怎么选？ A: 当前路线里插一段用「续这里」；想从这里改写后面、又保留旧路线时用「新建分支」。 Q: 断更小说可以从结尾接着写吗？ A: 可以。到最后一个可读位置，用尾部续写或精确划词续写；多个结局可以分别保存在不同分支。 Q: AI 续写能撤销吗？ A: 可以删除 AI 段或切回之前的分支。支持候选的重生会保留不同版本，但重要作品仍建议另做导出备份。 ### Set Up and Run a Roleplay Group Chat on Android URL: https://foreverse.app/docs/guide-roleplay/set-up-and-run-a-roleplay-group-chat Platform: Android roleplay group chat · App: v76+ · Path verified: 2026-08-10 Direct answer: From Home → Characters, long-press a card, select at least two chat-ready characters, and create the group. In the group, verify members and your Persona, then choose a speaking strategy, card-injection and lore-visibility settings, a Director Card, and whether automatic mode should run. Under Natural strategy, use @name to direct a reply, or tap a member in the group identity panel. Creating and configuring the group is local; generated rules, replies, and smart orchestration may be billed. Before you start: - Foreverse v76 or later on Android, with at least two chat-ready character cards in Home → Characters. - A connected text model through the official channel or BYOK before generating replies or AI-written group rules. - A Persona and a decision about whether each member's private card lore should stay isolated or be shared with the group. Steps: 1. Create the group from the character library: Open Home → Characters, long-press a character card, select at least one more chat-ready card, then tap the bottom Create group action. Name the group and confirm. The new session opens with each member's signed group greeting when available, otherwise its normal greeting. Success: The new chat identifies at least two members and shows attributed greetings instead of an empty conversation. 2. Verify members and your identity: Tap the group name or stacked avatars to open the group identity panel. Check every member, open a card when needed, and disable anyone who should not reply. Use ⋮ → Group chat settings → Member management to add, sort, or disable members; use ⋮ → Current status to confirm the Persona and shared world lore for this session. Success: The identity panel, member order, enabled state, and active Persona match the scene you intend to run. 3. Choose the member activation strategy: Open ⋮ → Group chat settings → Advanced settings. Choose Natural for @mentions plus rotation, List for fixed order, Manual when you always pick the speaker, or Pooled for a random eligible member. Then choose individual or merged character-card injection and set Member lore visibility to Follow generation mode, Keep private, or Share all. Success: The settings sheet shows the intended speaking strategy, injection mode, and lore-visibility rule. 4. Set the Director Card and automatic mode: In Group chat settings, edit the Director Card by hand for free, or choose Generate group rules for one billable Agent call. Apply a play-mode card if useful. Auto-play and smart orchestration can add model calls: applying a mode changes settings only, so tap Start automatic performance separately when you truly want unattended rounds. Success: The group rules describe relationships and tone, and automatic mode is visibly either stopped or running by deliberate choice. 5. Direct one test turn and keep a stop control: With Natural strategy, send a short message containing the exact @member name; alternatively open the identity panel and tap a member to force the next reply. Confirm the speaker and response, then stop automatic mode before leaving if it was started. Review request records before retrying a failed or expensive round. Success: The intended member produces the next reply, the group session is saved, and no automatic loop remains running unintentionally. Success checks: - The group contains the intended members, attributed greetings, enabled states, order, and Persona. - Speaking strategy, card injection, member-lore visibility, and Director Card match the planned scene. - An @mention or identity-panel selection directs the next speaker as expected. - Automatic mode has an explicit visible running/stopped state and can be stopped before leaving. Cost and data safety: - Creating the group, changing members, editing rules manually, choosing a strategy, and changing lore visibility are local and free. - Each generated character reply may be a separate text-model call. The official channel spends credits at 1 credit = US$0.0001; BYOK is billed by the provider connected to your key. - Generate group rules makes one Agent text call. Smart orchestration can add one lightweight call before each speaker, up to three in a round; automatic mode can continue creating billable rounds until stopped. - According to the chosen injection settings, requests can send member cards, group history, Director Card, active lore, Persona, and your message to the selected provider. Keep private member lore isolated when the story requires it. Troubleshooting: - Create group is unavailable or the group cannot be saved — Cause: Fewer than two selected cards are chat-ready, or a referenced member is no longer available in the character library. Fix: Return to Home → Characters, select at least two cards that can each start a normal chat, then create the group again. Reimport or repair a missing card before adding it. - The wrong member answers — Cause: Natural/List/Pooled selected another eligible member, the @name did not match, or a previous manual speaker lock is still active. Fix: Check the speaking strategy, use the exact member name after @ under Natural, or tap the intended member in the group identity panel to force the next reply. - A character knows another member's private lore — Cause: Merged card injection or Share all lore visibility exposed the other card's entries, or the secret was already spoken in chat history. Fix: Set Member lore visibility to Keep private and use individual card injection for future generations. Already revealed chat content does not disappear, so branch or start a clean group if necessary. - Applying an auto-play mode does not start replies — Cause: A play-mode card changes configuration but deliberately does not start unattended generation. Fix: After reviewing cost and settings, tap Start automatic performance. Use the separate Stop control when the scene has run long enough. - 402, 429, repeated orchestration failures, or unexpectedly high usage — Cause: Credits/quota ran out, the provider throttled calls, or automatic mode plus smart orchestration created multiple requests. Smart orchestration pauses after repeated failures in the session. Fix: Stop automatic mode, inspect request records, wait for rate limits or restore balance, reduce enabled members/context, disable smart orchestration, or select a cheaper configured model before one controlled retry. FAQs: Q: How do I make one specific character speak next? A: Under Natural strategy, include the exact @member name in your message. For a guaranteed manual choice, open the group identity panel and tap that member before generating. Q: Does choosing the auto-play preset immediately start a conversation? A: No. It applies the configuration only. Start automatic performance is a separate action, and Stop remains separate so a mode change cannot silently begin spending. Q: Does a Director Card cost credits? A: Writing or editing group rules yourself is free. Generate group rules makes one text-model Agent call and may be billed; later character replies are separate calls. Q: Can members be prevented from reading one another's private lore? A: Yes. Use individual card injection and Keep private under Member lore visibility. The setting affects later generations; it cannot erase a secret already written into chat history. ### 在 Android 建好并运行一场角色扮演群聊 URL: https://foreverse.app/zh/docs/guide-roleplay/set-up-and-run-a-roleplay-group-chat Platform: Android 角色群聊 · App: v76+ · Path verified: 2026-08-10 Direct answer: 在角色库长按并多选至少两张卡建群，进入群聊后核对成员与身份，设置发言策略、导演卡和自动模式；可用 @ 或成员面板定向发言。生成回复和自动轮次可能计费。 Before you start: - Android 已安装 Foreverse v76 或更高版本，并且「主页 → 角色」里至少有两张可开聊的角色卡。 - 生成回复或让 AI 写群规前，已连通官方渠道或 BYOK 文本模型。 - 已准备当前 Persona，并想清每位成员的私密卡设是各自隔离，还是允许全群共享。 Steps: 1. 从角色库创建群聊: 打开「主页 → 角色」，长按一张角色卡，再勾至少一张可开聊的卡，点底部「新建群聊」。填写群名并确认。新会话会优先写入每位成员带署名的群聊开场白；没有群聊专用开场时才用普通开场。 Success: 新聊天明确显示至少两位成员和各自署名的开场白，不是空白会话。 2. 核对成员与我的身份: 点群名或重叠头像打开群身份面板，逐个核对成员，需要时查看角色卡或禁用暂不发言的人。「⋮ → 群聊设置 → 成员管理」可添加、排序或禁用；「⋮ → 当前状态」核对本场 Persona 与共享世界资料。 Success: 群身份面板、成员顺序、启用状态和当前 Persona 都符合这场戏。 3. 设置成员激活与发言策略: 打开「⋮ → 群聊设置 → 高级设置」。自然轮替支持 @ 提及并结合顺序；列表轮流按固定顺序；手动指定由你选人；池化随机从可用成员抽取。再选择单独或合并角色卡，并把「成员设定可见性」设为跟随生成模式、各自保密或全员共享。 Success: 设置页显示目标发言策略、角色卡注入方式和成员设定可见性。 4. 设置导演卡与自动模式: 在群聊设置里手动编辑「群聊导演卡」不花钱；点「生成群规」会产生一次可能计费的 Agent 文本调用。需要时套用玩法模式卡。玩法卡只改配置，不会自己开演；确认设置和费用后，另点「开始自动开演」。 Success: 导演卡写清关系与基调，自动模式处于你主动选择的「已停」或「运行中」状态。 5. 定向测试一轮并保留停止入口: 自然轮替下，在短消息里写准确的「@成员名」；也可以打开群身份面板，点某位成员强制下一条由 TA 回复。核对发言人和回复后，若开了自动模式，离开前明确停止。失败或费用异常时先看请求记录，不要连续重试。 Success: 目标成员给出下一条回复，会话已保存，且没有无人看管的自动循环继续运行。 Success checks: - 群里是计划中的成员、署名开场、启用状态、顺序和 Persona。 - 发言策略、角色卡注入、成员设定可见性与导演卡符合玩法。 - @ 提及或群身份面板能让目标成员接下一句。 - 自动模式有明确的运行/停止状态，离开前可以停止。 Cost and data safety: - 建群、改成员、手写群规、切发言策略和成员设定可见性都是本地操作，不花钱。 - 每位角色生成回复都可能是一次独立文本调用；官方渠道按 1 credit = 0.0001 美元扣积分，BYOK 由 Key 所属供应商计费。 - 「生成群规」会走一次 Agent 文本调用；智能编排可能在每位角色发言前多一次轻量调用，一轮最多 3 次；自动模式会持续产生可能计费的轮次，直到停止。 - 请求会按注入设置把成员卡、群历史、导演卡、激活世界资料、Persona 和你的消息发给所选供应商。剧情要求保密时，要把成员私密设定设为各自隔离。 Troubleshooting: - 不能点新建群聊，或群聊保存失败 — Cause: 可开聊的已选角色少于两位，或引用的成员已经不在角色库。 Fix: 回到「主页 → 角色」，选择至少两张各自能正常开单聊的卡；缺失卡先重新导入或修好，再建群。 - 接话的不是我点名的成员 — Cause: 自然/列表/池化策略选中了另一位，@ 名称没匹配，或之前的手动发言锁还在。 Fix: 核对发言策略；自然轮替时写准确的 @成员名，或在群身份面板点目标成员，强制下一条由 TA 回复。 - 角色知道了另一位的私密设定 — Cause: 合并全员卡或「全员共享」暴露了别人的词条，也可能是秘密早已写进聊天历史。 Fix: 之后的生成改成单独注入，并把成员设定可见性设为「各自保密」。已经说出口的历史不会消失，必要时新建分支或干净群聊。 - 套用挂机玩法后没有自动回复 — Cause: 玩法模式只切换配置，设计上不会直接开始无人值守生成。 Fix: 核对费用与配置后，再点「开始自动开演」；达到想要的轮数后，用独立的停止入口结束。 - 402、429、智能编排连续失败或用量异常 — Cause: 积分/额度不足、供应商限流，或自动模式叠加智能编排产生了多次请求；编排连续失败会在本会话暂停。 Fix: 先停自动模式并看请求记录；等待限流或补充余额，减少启用成员/上下文，关闭智能编排或换更便宜的已配置模型，再受控重试一次。 FAQs: Q: 怎么指定下一句必须由某个角色说？ A: 自然轮替下在消息里写准确的 @成员名；要稳妥手动指定，就打开群身份面板，生成前点那位成员。 Q: 点了挂机看戏会立刻自动开聊吗？ A: 不会。玩法卡只应用配置；「开始自动开演」和「停止」都是独立动作，切玩法不会悄悄开始花钱。 Q: 群聊导演卡要花积分吗？ A: 自己手写或编辑免费；点「生成群规」会调用一次文本模型并可能计费，之后每位角色回复仍是另外的调用。 Q: 能不让群成员偷看彼此私密卡设吗？ A: 可以。用单独注入，并把成员设定可见性设成「各自保密」。它只影响之后生成，不能抹掉聊天里已经说出的秘密。 ### Import a SillyTavern Character Card and Start Chatting on Android URL: https://foreverse.app/docs/guide-roleplay/import-a-character-card-and-start-chat Platform: Android roleplay · App: v76+ · Path verified: 2026-08-10 Direct answer: Open Home → + → Import Tavern character card → Pick from phone and select the uncompressed PNG, JSON, or .charx file. Review the import preview/warnings, open the card from Home → Characters, and choose Start chat. In chat, confirm the model pill and Current status before sending. No Termux or desktop SillyTavern server is required. Before you start: - The original card file, sent as a file rather than a compressed social-media image. - A connected text model; a vision model is optional for image attachments. - For adult cards, the account/device must pass the app's applicable age and content controls. Steps: 1. Protect the source file: Download the card as a document/file. If both PNG and JSON/.charx are available, keep both until import is verified. 2. Import from Home: Open Home → + → Import Tavern character card → Pick from phone. Select PNG, JSON, or .charx; review the import preview plus any compatibility, script/network, or age warning instead of skipping it. 3. Inspect what travelled with the card: Open Home → Characters and open the card detail. Check name, description, greeting, example dialogue, and alternate greetings. Use the import report or Me → Advanced settings → Character Cards & Tavern → Tavern Asset Hub to confirm embedded lore, regex, Quick Reply, and referenced assets. 4. Start a clean session: Choose Start chat and read the greeting. In chat, confirm the model pill; open ⋮ → Current status to select Persona, writing style, and World lore as needed before sending. 5. Send a diagnostic first message: Reply to a concrete detail in the greeting. After the response, inspect the message prompt if you need to verify whether lorebook entries or card instructions were injected. Success checks: - The character detail shows real card fields rather than only an image. - The greeting and alternate greetings are present as expected. - The first response stays in character and the session is saved under that card. - The import report or Tavern Asset Hub accounts for embedded lore/regex/Quick Reply assets instead of silently losing them. Cost and data safety: - Importing and inspecting a card is local and free. Sending, regenerating, continuing, image generation, translation, and TTS can each create separate billable calls. - A chat request sends the card prompt, selected history, active lore, Persona, and your message to the chosen provider. Review sensitive cards before using a cloud provider. - Do not assume a pretty PNG still contains card data after a social platform resizes it. Keep the original export outside the app. Troubleshooting: - PNG is treated as a normal image / no card data — Cause: The embedded ccv3/chara chunk was stripped by compression or editing. Fix: Get the original PNG as a file, or import JSON/.charx. A screenshot cannot be converted back into the missing embedded fields. - Unsupported or malformed .charx — Cause: The archive structure or manifest is invalid, incomplete, or from an incompatible exporter. Fix: Re-export from the source tool, keep the zip/container intact, or use the card's JSON version. Review the import report for the first invalid asset. - The greeting appears but Send has no model — Cause: No enabled default text model is available. Fix: Open Me → AI Models & Services, test and enable a text model, return to the session, and select it without reimporting the card. - The character ignores its lorebook — Cause: No entry triggered, the book is disabled, scan depth/budget excluded it, or a key does not match the recent text. Fix: Open active worldbook entries and Prompt Inspector, test the key against the actual conversation text, and avoid making every entry constant. FAQs: Q: Which character-card formats work? A: Foreverse accepts card PNGs with embedded data, JSON cards, and V3 .charx containers. Exact asset compatibility depends on what the exporter included. Q: Do I need Termux or a PC? A: No. Foreverse is a native Android app and runs the card/chat workflow directly. The model request goes to the provider you selected. Q: Will importing a card publish it? A: No. Local import adds it to your device. Publishing or sharing is a separate explicit action. ### 在 Android 导入 SillyTavern 角色卡并开聊 URL: https://foreverse.app/zh/docs/guide-roleplay/import-a-character-card-and-start-chat Platform: Android 角色聊天 · App: v76+ · Path verified: 2026-08-10 Direct answer: 从主页导入原始 PNG/JSON/.charx 角色卡，选好文本模型后开聊。导入本身不计费，发消息可能计费；压缩图片可能丢失卡数据。 Before you start: - 准备原始角色卡文件，最好以文件而不是压缩图片接收。 - 已连通文本模型；只有发图片时才需要可看图模型。 - 成人卡需满足 App 适用的年龄与内容控制。 Steps: 1. 先保护原文件: 把卡作为文件下载。若同时有 PNG 和 JSON/.charx，导入核对完成前都保留。 2. 从主页导入: 打开「主页 → + → 导入酒馆角色卡 → 从手机选文件」，选 PNG、JSON 或 .charx；导入预览里的兼容性、脚本/联网或年龄提示要先看清。 3. 检查随卡带进来的内容: 到「主页 → 角色」打开卡详情，核对名称、人设、开场白、示例对话和备选开场。随卡世界书、正则、Quick Reply 与引用资产用导入报告或「我的 → 高级设置 → 角色卡与酒馆 → 酒馆资产中心」核对。 4. 新建一场干净会话: 点「开始聊天」并先读开场白。进聊天后核对顶部模型；需要时从「⋮ → 当前状态」选择 Persona、写作风格与世界资料，再发送。 5. 用第一条消息做诊断: 回复开场白里的一个具体细节。收到回复后，如要确认世界书或卡指令是否注入，可查看本消息 Prompt。 Success checks: - 角色详情显示真实卡字段，而不只是普通图片。 - 开场白和备选开场数量符合预期。 - 首条回复基本保持人设，会话归档在该角色下。 - 导入报告或酒馆资产中心能核对随卡世界书/正则/Quick Reply，没有静默丢失。 Cost and data safety: - 导入和查看卡不花钱；发送、重生、续写、生图、翻译与朗读可能各自产生独立计费调用。 - 聊天会把卡提示词、所选历史、激活世界书、Persona 和你的消息发给所选供应商。敏感卡在使用云模型前先检查内容。 - 好看的 PNG 经社交平台缩放后不一定还带卡数据；原始导出文件要在 App 外留一份。 Troubleshooting: - PNG 只被当普通图片 / 找不到卡数据 — Cause: 压缩或编辑清掉了内嵌 ccv3/chara 数据块。 Fix: 索要以文件发送的原始 PNG，或导入 JSON/.charx。截图无法恢复已经丢失的字段。 - .charx 不支持或格式损坏 — Cause: 容器目录、manifest 或资产不完整，或导出器格式不兼容。 Fix: 从来源工具重新导出并保持容器完整，或用 JSON 版本；从导入报告的第一个无效资产开始排查。 - 有开场白，但发送时没有模型 — Cause: 没有已启用的文本默认模型。 Fix: 去「我的 → AI 模型与服务」测连通并启用文本模型，回到会话选择即可，不用重导卡。 - 角色像没读到世界书 — Cause: 条目未触发、世界书被关、扫描深度/预算排除了它，或关键词没命中最近对话。 Fix: 打开当前世界书和 Prompt Inspector，用真实对话文本检查关键词；不要把所有条目都设成常驻。 FAQs: Q: 支持哪些角色卡格式？ A: 支持带内嵌数据的角色卡 PNG、JSON，以及 V3 .charx 容器。具体资产兼容性取决于导出文件实际包含什么。 Q: 需要 Termux 或电脑吗？ A: 不需要。Foreverse 是原生 Android App，直接运行导卡和聊天流程；模型请求发给你选择的供应商。 Q: 导入后会自动发布到社区吗？ A: 不会。本地导入只加入你的设备，发布或分享是另一项必须明确触发的操作。 ### Find, Review, and Import a Community Character Card URL: https://foreverse.app/docs/guide-roleplay/find-review-and-import-a-community-character-card Platform: Android Community and roleplay · App: v76+ · Path verified: 2026-08-10 Direct answer: Open Discover, choose Character cards, then search and filter by tags, language, rating, and sort order. Browsing does not require sign-in. Before importing, open the detail page and review the author, recommended model, version history, content warnings, content source, and image source. Tap Import to device and sign in when prompted; a completed download uses the account's daily download quota. Then open Home → Characters, inspect the imported card, and choose Start chat. Before you start: - Foreverse v76 or later in a Play/site distribution that includes the Discover community surface, plus a working network connection. - An account is optional for browsing but required to import; viewing 18+ content also requires the applicable birth-year verification. - Available character-library capacity and daily download quota for import, plus a connected text model only when you are ready to chat. Steps: 1. Search the Character cards shelf: Open the bottom Discover tab and choose the Character cards type. Enter a name, theme, or tag in search; open Filters to narrow tags, New/Rising/Popular sort order, language, and content rating. Browsing the shelf and detail pages works without signing in. Success: The results reflect the intended query and filters, with Character cards selected rather than an unrelated content type. 2. Check the author and version: Open a candidate's detail page. Read the author row/profile and the public description instead of trusting the cover alone. Expand Version history, read update notes, and avoid an older row marked Replaced when a current version is available. Success: You can identify the creator and current version, and know whether an older version has been superseded. 3. Review fit, warnings, and sources: Before downloading, read How to play when present, Recommended models, Content warnings, Content source (original, derivative, or authorized), and Image source. Confirm that the rating and declared origin fit your use; a polished cover is not proof of ownership or compatibility. Success: The recommended model, content boundaries, provenance declarations, and any missing information are understood before import. 4. Sign in and import to the device: Tap Import to device. If prompted, sign in and return to the same detail page; complete age verification for restricted content. The app checks character-library capacity before download, then uses one unit of the account's daily download quota when the download succeeds. Read any import warning instead of dismissing it. Success: The detail page reports the item as imported and the character appears in the local library; no quota was assumed spent when a pre-download capacity gate blocked the action. 5. Inspect the local card, then start chat: Open Home → Characters and select the imported card. Check its name, creator notes, greeting, alternate greetings, and attached lore/assets. Choose Start chat, confirm the text model and Current status, then send one short message tied to the greeting before committing to a long session. Success: The intended card opens with its real fields and greeting, and the first saved reply uses the selected model. Success checks: - Search and filters return the intended Character card rather than a similarly named pack or older version. - Author, current version, recommended model, content warnings, content source, and image source were reviewed before import. - The signed-in import finishes within library capacity and the card is visible under Home → Characters. - The imported greeting and attached assets are present, and a deliberate first chat can be saved under that card. Cost and data safety: - Browsing, searching, filtering, and reading details are free and do not require sign-in. They do use network data to load community metadata and images. - Importing a community item requires sign-in and consumes the account's daily download quota after a successful download; the server sets the quota by trust level and the app does not promise a fixed number. A pre-download free-library capacity block does not consume it. - Downloading and inspecting a card do not call an AI model or spend AI credits. Sending or regenerating chat replies may be billed: official calls use credits at 1 credit = US$0.0001, while BYOK is billed by your provider. - The imported card becomes local app data. When you chat, its prompt, selected history, active lore, Persona, and your message are sent to the chosen model provider, so review sensitive content first. Troubleshooting: - Search returns no suitable card — Cause: A type, tag, language, sort, or content-rating filter is too narrow; restricted content may also be hidden until age verification. Fix: Keep Character cards selected, clear filters one at a time, try a shorter title/tag, and complete birth-year verification only if you intentionally want the restricted-content filter. - Import opens sign-in or returns to the detail page without downloading — Cause: Browsing is public, but importing requires a valid signed-in account; the session may have expired or the network request may not have completed. Fix: Sign in through the prompt, return to the same detail page, and tap Import to device once. If it still fails, check the network before retrying rather than repeatedly consuming requests. - Today's download quota is exhausted — Cause: The account has used its server-assigned daily community download allowance. Fix: Wait for the next-day reset. The app does not expose a guaranteed fixed allowance, and Agent-assisted downloads use the same account quota. - A Pro or character-library capacity prompt appears — Cause: The free local character limit was reached before the community download began. Fix: Back up and remove an unneeded local character, or use Pro if appropriate, then import again. A capacity block before download should not consume daily download quota. - The imported card opens but chat has no model — Cause: Community import succeeded, but no enabled default text model is available for roleplay chat. Fix: Open Me → AI Models & Services, connect and enable a text model, return to the existing imported card, and start chat without downloading another copy. FAQs: Q: Do I need to sign in just to browse Community cards? A: No. The shelf, search, filters, and detail pages are public. Sign-in is requested when you download, like, favorite, follow, or publish. Q: Does importing a community card spend AI credits? A: No AI model is called for the import itself. A successful download uses the separate daily community download quota; later chat replies may spend official credits or provider-billed BYOK usage. Q: What should I review before importing a card? A: Check the author, current version and update notes, recommended model, content warnings, content source, and image source. Treat absent or unclear declarations as information you still need, not as automatic permission. Q: Does importing immediately send a message to the character? A: No. Import adds the card to the local character library. Open Home → Characters, inspect it, choose Start chat, select a model, and send the first message yourself. ### 在社区找到、审清并导入一张角色卡 URL: https://foreverse.app/zh/docs/guide-roleplay/find-review-and-import-a-community-character-card Platform: Android 社区与角色聊天 · App: v76+ · Path verified: 2026-08-10 Direct answer: 打开「发现」搜索并筛选角色卡，先核对作者、版本、雷点和内容来源；登录后导入到本机，再从角色详情开始聊天。浏览免费，导入会占用账号当日下载额度。 Before you start: - Android 已安装 Foreverse v76 或更高版本，使用包含「发现」社区的 Play/官网分发包，并且网络可用。 - 只浏览不用登录；导入时需要账号。查看 18+ 内容还要按提示完成出生年份验证。 - 本地角色库有容量，账号还有当日下载额度；文本模型只在准备开聊时才需要。 Steps: 1. 搜索并筛选角色卡: 打开底部「发现」，类型选「角色卡」。在搜索框输入卡名、题材或标签；打开筛选，逐项选择标签、最新/上升中/热门排序、语言和内容分级。浏览瀑布流和详情不要求登录。 Success: 结果符合搜索词和筛选，当前类型确实是角色卡，不是同名素材包或其他内容。 2. 核对作者与当前版本: 打开候选卡详情，不要只看封面，要读作者行/作者主页和公开简介。展开「版本记录」，查看更新说明；有新版时，不要误选标着「已替代」的旧版本。 Success: 能说清作者是谁、当前是哪一版，以及旧版本是否已经被替代。 3. 检查适配、雷点与来源: 下载前依次看有则显示的「玩法介绍」、推荐模型、雷点声明、内容来源（原创/二创/已获授权）和图片来源。确认内容分级与声明适合自己；封面精美不等于权利或兼容性已经得到证明。 Success: 已理解推荐模型、内容边界、来源声明，以及仍然缺失或说不清的信息。 4. 登录并导入到本机: 点「导入到本机」。若弹登录，完成后回到同一详情；受限内容按需完成年龄验证。App 会先检查角色库容量，下载成功才占账号当日下载额度。导入提示要读清，不要直接略过。 Success: 详情显示已导入，角色出现在本地库；如果下载前被容量门拦住，没有假定额度已经消耗。 5. 检查本地卡，再开始聊天: 进入「主页 → 角色」打开刚导入的卡，核对名称、作者备注、开场白、备选开场和随卡世界书/资产。点「开始聊天」，确认文本模型和「当前状态」，先用开场里的一个细节发短消息，再决定是否长聊。 Success: 正确角色显示真实字段与开场白，第一条回复使用所选模型并保存在这张卡下。 Success checks: - 搜索与筛选找到的是目标角色卡，不是同名包或已替代旧版。 - 导入前已核对作者、当前版本、推荐模型、雷点、内容来源与图片来源。 - 登录后的导入在角色库容量内完成，卡能在「主页 → 角色」找到。 - 导入的开场白与随卡资产存在，并能主动开始一场保存到该卡的聊天。 Cost and data safety: - 浏览、搜索、筛选和看详情免费，也不要求登录；加载社区元数据和图片仍会使用网络流量。 - 导入社区内容需要登录，成功下载后占账号当日下载额度；额度由服务端按信任等级决定，App 不承诺固定数字。下载前被免费角色容量拦住不消耗额度。 - 下载和检查卡不会调用 AI，也不扣 AI 积分。发送或重生回复才可能计费：官方渠道 1 credit = 0.0001 美元，BYOK 由自己的供应商收费。 - 导入后角色卡成为本地 App 数据；开聊会把卡提示词、所选历史、激活世界书、Persona 和你的消息发给所选模型供应商，敏感内容要先检查。 Troubleshooting: - 搜索不到合适的卡 — Cause: 类型、标签、语言、排序或内容分级筛得过窄；受限内容也可能在年龄验证前隐藏。 Fix: 保持类型为「角色卡」，每次只清一个筛选，改用更短卡名/标签；只有确实想看受限内容时才完成出生年份验证。 - 点导入只弹登录，或回到详情却没下载 — Cause: 浏览公开，但导入要求有效登录；也可能会话过期或网络请求没有完成。 Fix: 按提示登录，回到同一详情后只点一次「导入到本机」。仍失败先检查网络，不要连续点击制造重复请求。 - 提示今日下载额度已用完 — Cause: 账号已用完服务端分配的当日社区下载额度。 Fix: 等待次日重置。App 不展示一个保证不变的固定数量；让 Agent 代下载也共用同一账号额度。 - 弹出 Pro 或角色库容量提示 — Cause: 社区下载开始前，本地免费角色数量已经到上限。 Fix: 先备份并删除不再需要的本地角色，或按需要使用 Pro，再回来导入。下载前容量门拦截不应消耗当日下载额度。 - 卡已导入，但开聊时没有模型 — Cause: 社区导入成功了，但角色聊天没有已启用的文本默认模型。 Fix: 去「我的 → AI 模型与服务」连通并启用文本模型，再回现有角色开聊，不用重新下载一份。 FAQs: Q: 只逛社区角色卡也必须登录吗？ A: 不用。瀑布流、搜索、筛选和详情都能免登录看；下载、点赞、收藏、关注或发布时才要求登录。 Q: 导入社区卡会扣 AI 积分吗？ A: 导入本身不调用 AI。成功下载占另一套社区当日下载额度；之后聊天才可能扣官方积分或产生 BYOK 供应商费用。 Q: 导入前最少要检查什么？ A: 看作者、当前版本与更新说明、推荐模型、雷点声明、内容来源和图片来源。没有写清的地方应当视为仍需确认，而不是自动获得授权。 Q: 导入后会立刻给角色发消息吗？ A: 不会。导入只把卡放进本地角色库；再去「主页 → 角色」检查卡，点「开始聊天」、选模型，由你主动发送第一条消息。 ### Configure BYOK and Fix 401, 404, 429, or No-Model Errors URL: https://foreverse.app/docs/guide-models/configure-byok-and-fix-errors Platform: Android model settings · App: v76+ · Path verified: 2026-08-10 Direct answer: Open Me → AI Models & Services, pick the provider that issued your key, paste it, and run Test connection. For a custom provider, enter the full versioned API base (for example, https://api.example.com/v1) and let the client append relative endpoints. Then import models from upstream /v1/models or add the exact model ID manually, verify its modality, and select it under Default model configuration. Diagnose 401 as credentials, 404 as host/path/model, and 429 as quota or throttling. Before you start: - A key copied from the provider's official console, with required billing/project permissions enabled. - The provider's documented API base URL and model ID; a chat-web subscription is not automatically an API entitlement. - Network access to that host from the phone. Steps: 1. Choose the matching provider: Open Me → AI Models & Services and search for the company that issued the key. Use a built-in preset when available; use Add custom provider only for a compatible gateway or self-hosted endpoint. 2. Enter credentials without exposing them: Paste the key into the protected field. Do not include labels, shell syntax, quotes, or spaces copied around the key. Never place it in screenshots, logs, or support messages. 3. Set the API host once: For custom providers, set the full base including its required version prefix. Do not add /chat/completions at the end: Foreverse appends the relative endpoint for the selected protocol. 4. Test, then add models: Run Test connection and review the exact status/provider message. In the provider's model section, choose Import from upstream /v1/models or Add custom model, select the exact ID, and verify its text/vision/image/speech/video modality instead of assuming every returned ID supports everything. 5. Assign defaults and make one small call: Open Default model configuration, set the relevant modality, then run a short test from the actual Reader/chat/media surface. Check Me → Request records for status, latency, tokens, and channel. Success checks: - Test connection reports success and the expected host/provider. - At least one enabled model appears in the correct modality picker. - The actual feature completes a small request. - API request records show BYOK, the expected provider/model, and no secret value. Cost and data safety: - A connectivity test may be a real API request. Your provider may bill it even when the output is tiny. - BYOK usage is billed by the provider, not deducted from Foreverse credits. Foreverse cannot refund a charge made directly by a third party. - Keys are encrypted on-device and should never be copied into reports. Requests go to the host you configured, so verify custom gateways before sending private text. - A failed client display does not always prove the upstream accepted nothing. Use provider usage plus API request records before repeating expensive media calls. Troubleshooting: - 401 / 403 unauthorized or forbidden — Cause: Wrong/expired key, missing project permission, incompatible auth header, or key and API host belong to different services. Fix: Generate or copy an API key from the correct provider project, confirm API billing/access, remove surrounding spaces, and retest against that provider's documented host. - 404 not found — Cause: The base host lacks the version prefix, includes an endpoint twice, uses the wrong protocol, or the model ID does not exist there. Fix: Use the provider base exactly once (commonly ending /v1), remove /chat/completions from the host, confirm protocol and model ID, then test again. - 429 / quota exceeded / rate limit — Cause: Account balance, requests-per-minute, tokens-per-minute, or concurrency limit was reached. Fix: Check the provider console for the precise quota, wait for reset, reduce concurrency/output, add billing if intended, or switch to another configured provider. - Test succeeds but the model is missing — Cause: The model was not enabled, was classified under another modality, or the provider's list endpoint did not return it. Fix: Open the provider page, enable or manually add the exact documented model ID, set its supported modality, then reopen the feature picker. - TLS / network / timeout — Cause: The host is unreachable from this network, certificate is invalid, proxy/VPN policy interferes, or the upstream is slow. Fix: Verify the URL in the provider docs, try a normal reachable network, remove untrusted interception, and check provider status. Do not disable certificate validation. FAQs: Q: Should the API host include /v1? A: Include the provider's complete documented version prefix. Do not append an operation path such as /chat/completions; Foreverse adds relative endpoints. Q: Why does my ChatGPT/Claude/Gemini subscription key not work? A: Consumer chat subscriptions and developer API billing are usually separate products. Create an API project/key and enable its billing or quota in the provider console. Q: Can I mix official models and BYOK? A: Yes. Defaults are per modality and features can choose a model. For example, text can use BYOK while image generation uses the official channel. ### 配置 BYOK，并排查 401、404、429 与模型不见了 URL: https://foreverse.app/zh/docs/guide-models/configure-byok-and-fix-errors Platform: Android 模型设置 · App: v76+ · Path verified: 2026-08-10 Direct answer: 到「我的→AI 模型与服务」填带版本前缀的 API Host 和 Key，测连通后设默认模型。测试可能计费，401/404/429 分别先查凭证、路径和限流。 Before you start: - 从供应商官方控制台复制的 API Key，并已开通所需项目权限/计费。 - 供应商文档中的 API 基地址和模型 ID；网页聊天订阅通常不等于 API 权限。 - 手机网络能够访问该地址。 Steps: 1. 选对供应商: 打开「我的 → AI 模型与服务」，搜索签发 Key 的公司。有内置预设就优先用；只有兼容网关或自建端点才使用「添加自定义供应商」。 2. 安全填写凭据: 把 Key 粘贴进受保护输入框，不要带标签、shell 语法、引号或前后空格。Key 绝不能进入截图、日志或客服消息。 3. API 地址只写一次: 自定义供应商填写包含版本前缀的完整基地址；末尾不要再加 /chat/completions，Foreverse 会根据协议拼相对端点。 4. 测连通，再添加模型: 点「测连通」，读清状态码和供应商原话。再到供应商的模型区点「从上游 /v1/models 批量导入」或「添加自定义模型」，选精确 ID，并确认它实际支持文本/视觉/图像/语音/视频中的哪一类。 5. 设默认并做一次小请求: 打开默认模型配置，给对应模态设默认；再从真实阅读器/聊天/媒体入口发一个很小的测试。到「我的 → 请求记录」核对状态、耗时、tokens 和渠道。 Success checks: - 测连通成功，并显示预期的供应商/地址。 - 正确模态的模型选择器里能看到至少一个已启用模型。 - 真实功能入口的小请求成功。 - API 请求记录显示 BYOK、预期模型，且没有泄露 Key。 Cost and data safety: - 测连通可能发出真实 API 请求，哪怕输出很短，供应商也可能计费。 - BYOK 费用由供应商收取，不扣 Foreverse 积分；第三方直接产生的费用 Foreverse 无法退款。 - Key 加密保存在设备上，不应复制进任何报告。请求会发往你填写的地址，敏感文本使用自定义网关前要先确认可信。 - 客户端显示失败不总能证明上游什么都没接收。昂贵的媒体请求不要盲目连点，先对照请求记录与供应商用量。 Troubleshooting: - 401 / 403 未授权或禁止 — Cause: Key 错误/过期、项目权限不足、认证协议不兼容，或 Key 和地址不是同一服务。 Fix: 从正确项目重新生成/复制 API Key，确认 API 计费和权限，去掉空格，再对照官方文档地址测一次。 - 404 not found — Cause: 基地址缺版本前缀、端点重复拼接、协议选错，或模型 ID 在该地址不存在。 Fix: 只保留一次供应商基地址（常见以 /v1 结尾），去掉 /chat/completions，核对协议和模型 ID 后再测。 - 429 / quota exceeded / rate limit — Cause: 余额、每分钟请求数、每分钟 token 或并发上限触顶。 Fix: 去供应商控制台看具体配额，等窗口恢复，降低并发/输出，按意愿开通计费，或换另一个已配置供应商。 - 测连通成功但找不到模型 — Cause: 模型没启用、被分到另一模态，或供应商列表接口没返回它。 Fix: 打开供应商页，启用或手动添加文档里的精确模型 ID，标对模态，再重开功能选择器。 - TLS / 网络 / 超时 — Cause: 当前网络到不了、证书无效、代理/VPN 干扰，或上游太慢。 Fix: 按官方文档核对 URL，换正常可达网络，移除不可信拦截，并看供应商状态；不要关闭证书校验。 FAQs: Q: API 地址到底要不要带 /v1？ A: 要按供应商文档填写完整版本前缀；不要再附加 /chat/completions 等具体操作路径，Foreverse 会拼相对端点。 Q: 为什么 ChatGPT/Claude/Gemini 会员不能直接当 API 用？ A: 网页聊天订阅与开发者 API 计费通常是两个产品，需要在供应商控制台新建 API 项目/Key 并开通额度。 Q: 官方模型和 BYOK 能混用吗？ A: 可以。默认模型按模态设置，各功能也能选模型，例如文字走 BYOK，生图走官方渠道。 ### Fix ‘The Character Forgot’: Memory, Lorebook, Context, and Prompt Checks URL: https://foreverse.app/docs/guide-roleplay/fix-memory-and-worldbook Platform: Android roleplay · App: v76+ · Path verified: 2026-08-10 Direct answer: First decide where the fact should live: permanent setting belongs in the character card or lorebook; events from this session belong in chat history, checkpoints, or the long-chat summary. Reproduce the miss, open Prompt Inspector on the reply, and check whether the fact was actually included. If absent, repair the trigger/budget/source; if present, make the instruction clearer instead of duplicating it everywhere. Before you start: - A saved chat where the missing fact can be reproduced without sending sensitive credentials. - Access to the card's active lorebook and long-chat summary/settings. - A clear one-sentence statement of the fact and when it should apply. Steps: 1. Classify the forgotten fact: Put identity and stable world rules in the card/lorebook; session events in history or summary; user identity in Persona; temporary scene instructions in Author's Note/system guidance. 2. Reproduce with a neutral question: Ask one direct, non-leading question in a copy or checkpointed session. Avoid adding the answer to the question, because that hides whether retrieval worked. 3. Inspect the actual prompt: Long-press the reply and tap Prompt. Search Prompt Inspector for the exact fact, the relevant lore entry name/key, summary, Persona, and any conflicting instruction. Edit active lore under chat ⋮ → Current status → World lore. 4. Repair the missing layer: If a lore entry is absent, enable the book, fix primary/secondary keys, scan depth, probability, recursion, position, or token budget. If old session events are absent, refresh the long-chat summary or preserve a checkpoint. 5. Test once, then remove duplication: Send the same neutral test after the repair. When it works, remove duplicate constant entries or contradictory summaries so context remains stable and cheaper. Success checks: - Prompt Inspector contains the correct single source for the fact. - The character answers the neutral test consistently for at least two nearby turns. - No contradictory duplicate remains in card, lorebook, summary, and Persona. - Context/token use stays within the chosen model's limits. Cost and data safety: - Editing cards, lorebooks, summaries, and checkpoints is local. Every diagnostic reply/regeneration is still a model call and may be billed. - Making every lore entry constant increases tokens and can crowd out recent conversation, raising cost while making memory worse. - Prompt Inspector can expose private card/history content. Redact it before sharing; API keys should never appear there or in screenshots. Troubleshooting: - The lore entry never appears in Prompt Inspector — Cause: Book/entry disabled, key mismatch, secondary-key rule, probability, scan depth, recursion, or budget excluded it. Fix: Test the exact key against recent message text, temporarily make activation deterministic, verify position/budget, then restore intentional probability after the entry is proven. - The fact is present but the model still contradicts it — Cause: A stronger conflicting instruction exists, wording is ambiguous, or the model/context is weak for the task. Fix: Remove contradictions, state the rule once in imperative/plain language, place it in the appropriate priority layer, and test with a capable model. - Recent facts work, old events vanish — Cause: The raw history fell outside the context window and the summary omitted the event. Fix: Regenerate/edit the long-chat summary to include the durable event, create a checkpoint, or move a truly permanent fact into lore rather than inflating raw history. - Context too long / token limit — Cause: Too many constants, oversized card/lore, duplicate summaries, or long output reservation consumes the model window. Fix: Trim duplicates, turn situational constants back into keyed entries, shorten summaries, and use a model with a larger context only when needed. FAQs: Q: Should every important fact be a constant lore entry? A: No. Constants always consume context. Use them only for rules that must apply every turn; use triggers for situational lore and summaries/history for session events. Q: What is the fastest way to prove a worldbook worked? A: Trigger it with known text, then open Prompt Inspector on the resulting reply and search for the entry content. A plausible answer alone is not proof. Q: Is character memory the same as Companion memory? A: No. Roleplay sessions rely on card, lorebook, history and summaries. Companion adds an explicit editable long-term-memory layer; diagnose the surface you are actually using. ### 角色为什么忘了：检查记忆、世界书、上下文与真实 Prompt URL: https://foreverse.app/zh/docs/guide-roleplay/fix-memory-and-worldbook Platform: Android 角色聊天 · App: v76+ · Path verified: 2026-08-10 Direct answer: 先用明确关键词复测世界书并在请求检查器确认是否注入；角色忘事再查上下文预算和摘要。复测会调用模型并可能计费，别连续盲重试。 Before you start: - 有一场能复现遗漏的已保存聊天，且不需要暴露敏感凭据。 - 能查看本卡激活的世界书和长聊摘要/设置。 - 能用一句话说清这条事实，以及它应在什么情况下生效。 Steps: 1. 先给遗忘的事实分类: 身份与稳定世界规则放卡/世界书；本场事件放历史或摘要；用户身份放 Persona；临时场景要求放作者注释/系统指令。 2. 用中性问题复现: 在副本或有 checkpoint 的会话里，问一个直接但不暗示答案的问题。问题本身不要带答案，否则看不出检索是否生效。 3. 查看真实 Prompt: 长按该回复点「提示词」，在 Prompt Inspector 里搜索事实原文、相关世界书条目名/关键词、摘要、Persona，以及可能冲突的指令。当前世界书从「聊天 ⋮ → 当前状态 → 世界资料」编辑。 4. 修缺失的那一层: 世界书没进来就检查启用、主/副关键词、扫描深度、概率、递归、位置和 token 预算；旧事件没进来就更新长聊摘要或保留 checkpoint。 5. 复测一次，再清重复: 修完后用同一个中性问题测试。成功后删掉重复常驻条目和互相冲突的摘要，让上下文稳定，也更省钱。 Success checks: - Prompt Inspector 里有且只有正确真源的事实。 - 角色在相邻至少两轮里能一致回答。 - 角色卡、世界书、摘要和 Persona 里没有互相矛盾的重复项。 - 上下文/token 使用没有超过模型上限。 Cost and data safety: - 编辑卡、世界书、摘要和 checkpoint 是本地操作；每次诊断回复或重生仍是模型调用，可能计费。 - 把所有世界书条目设成常驻会增加 token，还可能把最近对话挤出去，结果更贵、记忆反而更差。 - Prompt Inspector 可能暴露私密卡设和聊天历史；分享前必须脱敏，API Key 绝不能出现在其中或截图里。 Troubleshooting: - 世界书条目从不出现在 Prompt Inspector — Cause: 书/条目未启用、关键词不匹配、副关键词规则、概率、扫描深度、递归或预算把它排除了。 Fix: 用最近消息的原文测试关键词，暂时改成确定触发，核对位置/预算；证明能进 Prompt 后再恢复有意的概率。 - 事实已经在 Prompt 里，模型仍说反了 — Cause: 存在更高优先级冲突、表述含糊，或模型/上下文能力不够。 Fix: 消除冲突，只在合适层用明确命令式语言写一次，再用更适合的模型复测。 - 最近的事记得，久远事件消失 — Cause: 原始历史已掉出上下文，摘要又漏了这件事。 Fix: 重生/编辑长聊摘要，加入值得长期保留的事件，并建立 checkpoint；真正永久的事实再移入世界书。 - 上下文太长 / token 超限 — Cause: 常驻条目太多、卡/世界书过大、摘要重复，或给输出预留太多。 Fix: 去重，把只在特定场景生效的常驻改回关键词触发，缩短摘要；只有确实需要时才换更大上下文模型。 FAQs: Q: 重要事实都应该设成常驻吗？ A: 不应该。常驻每轮都占上下文，只用于每一轮都必须生效的规则；情境设定用关键词，场内事件用历史/摘要。 Q: 怎么最快证明世界书真的触发了？ A: 用已知文本触发，然后打开回复的 Prompt Inspector 搜索条目内容。角色碰巧答对不算证据。 Q: 角色聊天记忆和伴侣记忆是一回事吗？ A: 不是。角色聊天依赖卡、世界书、历史和摘要；伴侣另有可编辑长期记忆层。先确认自己在哪个界面。 ### Run Agent Tasks Safely: References, Approvals, Results, and Rollback URL: https://foreverse.app/docs/guide-agent/run-tasks-approve-and-rollback Platform: Android Agent Quick Dock · App: v76+ · Path verified: 2026-08-10 Direct answer: Open Quick Dock, describe one concrete outcome, and attach the exact book/card with @. Keep approvals on for writes or external effects until you trust the plan. Read each tool card before approving; after completion, open the changed item and verify it. Use Roll back this change on supported Agent write cards when the result is wrong—rollback is not a universal undo for provider charges or external actions. Before you start: - A configured tool-capable text model with available quota/credits. - A specific target book, card, note, or workspace file; avoid broad ‘clean everything’ requests. - A backup/export for irreplaceable data before a large multi-file task. Steps: 1. Open Quick Dock in the right context: Open the floating Quick Dock from the page that owns the content. Start a new task when the old conversation has unrelated instructions. 2. State outcome, scope, and stop condition: Say what should exist when done, which item may change, what must not change, and what needs your approval. Use @ to attach the exact book/card instead of relying on a similar name. 3. Review the plan and approvals: Keep the conservative approval mode for writes, commands, purchases, publishing, or external effects. Read the proposed target and action; reject or edit the instruction when the target is ambiguous. 4. Follow tool cards, not just the prose: Expand read/write/command cards to see which files or records were touched. Stop the run if it drifts, repeats calls, or requests authority outside the task. 5. Verify and, if needed, roll back: Open the actual book/card/note and confirm the content and location. For supported Agent file mutations, use Roll back this change and verify again. Export the corrected result when it matters. Success checks: - The task names one exact target and expected deliverable. - Every approved write/command points to the intended item. - The result is visible in the owning feature, not only claimed in Agent prose. - Rollback either restores the previous content or clearly reports that the operation is not rollback-capable. Cost and data safety: - Agent runs can make multiple model calls. A short-looking task may use several reasoning/tool steps, each billable through official credits or BYOK. - Stopping a run prevents future steps but does not refund completed provider calls. Rollback reverses supported data changes, not AI charges. - External publishing, purchases, account changes, and third-party side effects require explicit authority and may not be reversible. Do not approve them from a vague prompt. - References can expose the selected local content to the chosen model provider. Attach only what the task needs. Troubleshooting: - Agent asks to modify the wrong book/card — Cause: The target name is ambiguous or no exact @ reference was attached. Fix: Reject the step, stop if necessary, start a scoped instruction with the exact @ item and state that all other items are read-only. - Approval keeps reappearing — Cause: The task contains several distinct writes/commands, or the prior action failed and was proposed again. Fix: Expand the tool card and compare target/action. Approve only new intended steps; if it is looping, stop and ask for a no-write diagnosis. - Rollback is missing — Cause: The card is read-only, the action was external, the mutation did not capture a reversible snapshot, or the retention window is gone. Fix: Check the owning feature's history/export/branch controls. Restore from an export when available; do not assume rerunning the Agent reconstructs the exact old bytes. - Task stops with model/tool unavailable — Cause: Selected model lacks tool calling, provider failed, permission was denied, or the referenced item is no longer accessible. Fix: Choose a tool-capable enabled model, reopen the exact item, grant only the required approval, and resume or restart with the preserved task context. FAQs: Q: Should I turn on full automatic approval? A: Use it only for a narrowly scoped, reversible task you understand. Keep confirmation for money, publishing, account changes, commands, and broad writes. Q: Does Stop undo what already happened? A: No. Stop prevents later steps. Verify completed tool cards and use supported rollback/history/export controls for changes already made. Q: Can Agent rollback an AI charge? A: No. A content rollback does not reverse a provider request or official credit charge that already completed. ### 安全运行 Agent 任务：引用、审批、验收与回滚 URL: https://foreverse.app/zh/docs/guide-agent/run-tasks-approve-and-rollback Platform: Android Agent Quick Dock · App: v76+ · Path verified: 2026-08-10 Direct answer: 从 Quick Dock 打开 Agent，用 @ 引用目标，先审计划再批准写入；完成后查任务账和差异，必要时撤销。外部调用与已花费用不会随回滚退回。 Before you start: - 已配置支持工具调用的文本模型，并有足够额度/积分。 - 明确目标书、角色卡、笔记或工作区文件，避免「把所有东西整理一下」这种宽泛任务。 - 大范围多文件任务前，对不可替代数据做一次导出/备份。 Steps: 1. 在正确上下文打开 Quick Dock: 从内容所属页面打开悬浮 Quick Dock。旧会话有无关指令时，新建一项任务，不要把多个目标混在一起。 2. 写清结果、范围和停止条件: 说明做完后应当得到什么、允许改哪个对象、绝不能改什么、哪些动作必须问你。用 @ 精确引用书/卡，不要只靠相似名称。 3. 审计划和审批: 涉及写入、命令、付费、发布或外部影响时先用保守审批档。看清动作和目标；目标含糊就拒绝或补充指令。 4. 盯工具卡，不只看说明文字: 展开读取/写入/命令卡，核对实际碰了哪些文件或记录。发生跑偏、重复调用或索要任务外权限时立刻停止。 5. 去真实页面验收，必要时回滚: 打开对应书/卡/笔记，确认内容和位置。对支持的 Agent 文件变更点「回滚此变更」，再验一次；重要结果随后导出。 Success checks: - 任务只指向一个明确目标与交付物。 - 每项获批写入/命令都指向预期对象。 - 结果在真实功能页可见，而不是只有 Agent 文字声称完成。 - 回滚能恢复旧内容，或明确告诉你该操作不支持回滚。 Cost and data safety: - Agent 一项任务可能发多次模型请求；看起来很短的任务也可能有多步思考/工具调用，每步都可能通过官方积分或 BYOK 计费。 - 停止只阻止后续步骤，不退回已经完成的供应商调用；内容回滚也不等于费用退款。 - 外部发布、购买、账号变更和第三方副作用必须明确授权，而且可能不可逆；不要从含糊提示里直接批准。 - @ 引用会把任务需要的所选本地内容发给模型供应商，只附真正必要的对象。 Troubleshooting: - Agent 想改错书/卡 — Cause: 名字有歧义，或没有附精确 @ 引用。 Fix: 拒绝该步骤，必要时停止；重新用准确 @ 对象发任务，并声明其他对象只读。 - 审批反复出现 — Cause: 任务本来就有多项不同写入/命令，或上一步失败后被再次提议。 Fix: 展开工具卡对比目标和动作，只批准新的预期步骤；出现循环就停止，并让它先做只读诊断。 - 找不到回滚按钮 — Cause: 这是只读卡、外部动作、未保存可逆快照，或已过保留窗口。 Fix: 检查真实功能页的历史、导出或分支控制；有备份就恢复，不要假定重跑 Agent 能逐字重建旧内容。 - 模型/工具不可用导致停止 — Cause: 模型不支持工具调用、供应商失败、权限被拒，或引用对象已不可访问。 Fix: 选择已启用且支持工具的模型，重新打开准确对象，只授予需要的审批，再从保留上下文恢复或重启。 FAQs: Q: 要不要打开完全自动审批？ A: 只在范围很窄、可逆、你完全理解的任务上用。花钱、发布、账号变化、命令和大范围写入继续保留确认。 Q: 点停止会撤销已经做的事吗？ A: 不会。停止只阻止后续步骤；已经完成的工具卡要逐项核对，并用回滚/历史/导出恢复。 Q: Agent 能回滚已经产生的 AI 费用吗？ A: 不能。内容变更回滚不会撤销已经完成的供应商请求或官方积分扣费。 ### Create a Companion with Editable Memory and Safe Proactive Messages URL: https://foreverse.app/docs/guide-companion/create-memory-and-proactive-messages Platform: Android Companion · App: v76+ · Path verified: 2026-08-10 Direct answer: Create a Companion from Home or convert an imported character card, then review its persona and memory before chatting. Keep proactive messages off until the text/lightweight-task model, enabled message slots, quiet hours, daily limit, notification permission, and spending channel are correct. After enabling, inspect the first generated message and keep an exported Companion backup outside the app. Before you start: - A connected text model; image/TTS providers are optional for selfies or voice. - A clear persona/relationship boundary, or an imported card you are allowed to use. - Notification permission if you want messages outside the app. Steps: 1. Create or import the Companion: From Home, choose Companion creation or create from an imported character card. Set the displayed name, persona, relationship style, and confirmed reference image only if you want visual generation. 2. Check the initial home and memory: Open the Companion, read the profile/style, then open Memory. Add one stable fact manually and remove anything you do not want used later. 3. Have a short grounding conversation: Send a few messages that establish a real preference or plan. After memory extraction runs, review the new memory item for accuracy, duplication, and privacy. 4. Configure proactive messaging before enabling: On the Companion profile, review Let them reach out and the individual morning, evening, miss-you, anniversary, battery, and weather slots. Open Promises with them to set quiet hours and the daily maximum. Confirm Android notifications and the text/lightweight-task model under AI Models & Services, read the cost disclosure, and only then enable the master switch. 5. Verify and prepare portability: Wait for a normal eligible slot, then inspect the notification and saved message. Use Companion profile → Export Companion backup, store the zip somewhere private, and verify Restore Companion backup can select it before deleting anything. Success checks: - Persona and relationship style match what you intended. - Memory shows only accurate, editable facts and no accidental secret. - Message slots, quiet hours, daily limit, and the master switch reflect your choices after reopening settings. - A proactive message opens the correct Companion conversation. - A Companion-backup zip exists outside the app. Cost and data safety: - Creating/editing the local profile is free. Conversations, memory extraction, proactive messages, selfies, TTS, and other generated media may use separate model calls. - When proactive messages are enabled, the app may call the selected text/lightweight-task model while you are away. Use individual slots, quiet hours, a daily limit, and a suitable low-cost model; turning the master switch off prevents future scheduled generation. - Memory is editable local relationship data, but selected context is sent to the chosen provider during generation. Do not store passwords, API keys, OTPs, or information you would not send to that provider. - A Companion backup can contain intimate profile, memory, chat history, and assets. Treat it as sensitive personal data. Troubleshooting: - No proactive message arrives — Cause: Master switch or that message slot is off, no slot is due, quiet hours/daily limit applies, notifications are denied, background work is restricted, no usable model exists, or generation failed. Fix: Check the master switch, enabled slots, next eligible window and daily limit; allow notifications, remove battery/background restrictions where appropriate, verify the text/lightweight-task model, then inspect request records rather than repeatedly toggling. - A message arrives during quiet hours — Cause: Device time zone/clock or configured window differs from expectation, or it is a delayed notification for an earlier generated message. Fix: Confirm device time zone and the displayed quiet-hour range, inspect message/task time, and report a reproducible mismatch without changing system time. - Memory is wrong or duplicated — Cause: Extraction inferred too much, the same fact was phrased differently, or an old conflicting item remains. Fix: Edit/delete the wrong item, merge duplicates into one plain fact, and correct the Companion in chat. Keep only durable facts, not every passing mood. - Restored Companion is missing assets — Cause: The backup was incomplete/corrupt, referenced files were unavailable, or the package came from an older incompatible state. Fix: Keep the source Companion, export again with assets available, verify file size, and restore into a test copy before deleting the original. FAQs: Q: Are proactive messages on by default? A: No. The creation flow asks once; if you decline or dismiss it, proactive messages stay off. Enable them only after reviewing message slots, quiet hours, the daily limit, notifications, and the model/spending channel. Q: Can I edit what the Companion remembers? A: Yes. Long-term memory is visible and editable; remove mistakes, duplicates, or private facts you no longer want used. Q: Will turning proactive messages off delete memory? A: No. The switch controls future proactive generation, not the Companion profile, chat, or memory. You can export those separately. ### 创建有可编辑记忆的伴侣，并安全开启主动消息 URL: https://foreverse.app/zh/docs/guide-companion/create-memory-and-proactive-messages Platform: Android 伴侣 · App: v76+ · Path verified: 2026-08-10 Direct answer: 在主页创建伴侣，逐条核对记忆后再开启主动消息和免打扰时段。模型消息可能计费，记忆本地优先保存；通知默认关闭，也可随时停用。 Before you start: - 已连通文本模型；自拍或语音才需要额外图像/TTS 供应商。 - 有清楚的人设与关系边界，或拥有使用权的角色卡。 - 想在离开 App 后收到消息时，需要允许通知。 Steps: 1. 创建或导入伴侣: 从主页选择创建伴侣，或从已导入角色卡创建。设置显示名、人设、关系风格；只有要生成视觉内容时才配置确认过的参考图。 2. 检查初始空间与记忆: 打开伴侣，读一遍资料/风格，再进「记忆」。手动加一条稳定事实，并删除不希望以后使用的内容。 3. 用短对话建立真实事实: 聊几轮，明确一个偏好或计划。记忆抽取完成后，检查新记忆是否准确、重复或过于隐私。 4. 开启前先配主动消息: 在伴侣资料页检查「允许 TA 主动找你」以及早安、晚安、想你、纪念日、低电量、天气等具体时段；再进「和 TA 的约定」设置免打扰和每日上限。到「我的 → AI 模型与服务」确认文本/轻量功能模型，并确认系统通知权限。读清费用说明后，再开总开关。 5. 验证并准备迁移: 等到一个正常可触发时段，核对通知和已保存的消息；随后走「伴侣资料页 → 导出伴侣备份」，私密保存 zip，并确认「恢复伴侣备份」能选中它，再考虑删除任何东西。 Success checks: - 人设和关系风格符合预期。 - 记忆里只有准确、可编辑的事实，没有意外秘密。 - 重开设置后，消息时段、免打扰、每日上限和总开关仍保持你的选择。 - 主动消息通知能打开正确的伴侣会话。 - App 外已经保存一份伴侣备份 zip。 Cost and data safety: - 创建/编辑本地资料免费；聊天、记忆抽取、主动消息、自拍、TTS 等可能分别产生模型调用。 - 开启主动消息后，你离开时也可能调用所选文本/轻量功能模型。请设置具体时段、免打扰、每日上限并选择合适的低成本模型；关闭总开关会阻止未来计划生成。 - 记忆是可编辑的本地关系数据，但生成时所选上下文会发给供应商。不要保存密码、API Key、OTP 或不愿发给该供应商的信息。 - 伴侣备份可能包含亲密人设、记忆、聊天记录和资产，应当按敏感个人数据保管。 Troubleshooting: - 收不到主动消息 — Cause: 总开关或对应时段关着、当前没有时段到点、命中免打扰/每日上限、通知被拒、后台受限、没有可用模型，或生成失败。 Fix: 核对总开关、具体时段、下一可触发时间和每日上限；允许通知，按需解除电池/后台限制，验证文本/轻量功能模型，再查请求记录，不要反复开关碰运气。 - 免打扰时段仍收到消息 — Cause: 设备时区/时钟或配置窗口与预期不同，或这是更早生成但延迟到达的通知。 Fix: 确认设备时区和页面显示的免打扰范围，核对消息/任务时间；用可复现信息反馈，不要修改系统时间测试。 - 记忆错误或重复 — Cause: 抽取推断过度、同一事实表述不同，或旧冲突项还在。 Fix: 编辑/删除错误项，把重复合并成一句清楚事实，并在聊天里纠正；只保留耐久事实，不要记录每次短暂情绪。 - 恢复伴侣备份后少资产 — Cause: 备份不完整/损坏、引用文件当时不可用，或包来自旧的不兼容状态。 Fix: 保留源伴侣，确保资产可用后重新导出，核对文件大小；先恢复成测试副本，确认完整再删原件。 FAQs: Q: 主动消息默认开启吗？ A: 不会。创建收尾会问一次；你关掉或滑走后会保持关闭。先检查具体消息时段、免打扰、每日上限、通知权限和模型/付费渠道，再明确打开。 Q: 伴侣记住的内容能改吗？ A: 能。长期记忆可见、可编辑；错误、重复或不再希望使用的隐私事实都可以删除。 Q: 关闭主动消息会删记忆吗？ A: 不会。开关只控制未来主动生成，不删除人设、聊天或记忆；这些数据可以另外导出。 ### Export, Back Up, and Move Foreverse Data Without Losing Local Work URL: https://foreverse.app/docs/guide-settings/export-backup-and-move-data Platform: Android data management · App: v76+ · Path verified: 2026-08-10 Direct answer: Before changing phones, uninstalling, or switching between Play and site packages, inventory every local asset family and export each supported format. Copy the exports to storage outside the app, verify that files open and have non-zero size, then import on the destination while the source still exists. API keys and encrypted preferences may need to be configured again; never copy encrypted preference key files between devices. Before you start: - Enough external/cloud storage for exports and generated media. - Access to both source and destination devices/installations until verification finishes. - Provider/API credentials available from their official consoles for secure re-entry. Steps: 1. Inventory before touching the installation: List books and branches, character cards, worldbooks/presets/regex/QR/Persona, chat sessions, Companions, generated media, custom themes/skills, account state, and BYOK providers. Mark irreplaceable items. 2. Export by asset family: Use each feature's export: original-format card/worldbook where available, chat export per session, Export companion backup from the Companion profile, book/project export or the Pro full-library backup, and save/share generated media to user storage. 3. Move the files outside app storage: Place exports in a clearly named folder on SAF-accessible storage, computer, or trusted cloud. Include the date and source package/version in a small note, but never put API keys in it. 4. Verify before deleting anything: Check file count, size, and archive integrity. Open representative txt/json/png/zip/media files. On the destination, import into a test copy and verify chats, memory, lore and assets—not just the item title. 5. Reconfigure secrets and finish the move: Sign in as intended and add BYOK keys from provider consoles. Test one low-cost request. Keep the source installation and exports until the destination passes the complete checklist. Success checks: - Every irreplaceable asset family has an export outside app-private storage. - Representative exports open and archives are not empty/corrupt. - Destination imports contain body content, branches/history/memory and referenced assets as applicable. - A provider can complete one small request after credentials are re-entered. - Source data has not been removed before destination verification. Cost and data safety: - Export/import itself should not call AI. Full-library backup may be a Pro-gated convenience; individual supported exports remain the fallback. - Generated media may already have been billed when created. Saving/exporting it is not a reason to regenerate; if saving fails, preserve the successful task and retry delivery instead of paying again. - Play and site release builds share the package name but can have different signing certificates; switching can require uninstalling, which clears local app data. - Backups can contain novels, private chats, memory and personal media. Encrypt trusted storage and do not share the bundle publicly. - EncryptedSharedPreferences/keystore material is device-bound. Copying those files is neither a safe nor supported way to move BYOK secrets. Troubleshooting: - The new APK cannot install over the old one / signature conflict — Cause: Play and site builds were signed differently even though the package name matches. Fix: Do not uninstall yet. Export and verify all local data, then uninstall the source only when ready for a clean install and manual import. - Import recognizes a package but content/assets are missing — Cause: The export was incomplete, external references were unavailable, or only metadata was copied. Fix: Return to the intact source, make a fresh export with assets present, compare file sizes/counts, and test the new package before removing the source. - BYOK providers appear but keys do not work after moving — Cause: Device-bound encryption cannot be restored by copying preference files, or credentials were not re-entered. Fix: Delete any broken copied secret state, add keys again from the official provider consoles, and run a small connectivity test. Never sync keystore files. - Backup/export file is 0 bytes or archive will not open — Cause: Storage permission, cancellation, out-of-space, or interrupted write. Fix: Keep the source data, free space, choose a writable SAF location, export again to a new filename, and verify before leaving the screen. FAQs: Q: Does account sign-in restore all local books and chats? A: Do not assume so. Credits/account state and local creative data are different. Export local books, branches, cards, chats, Companion memory and media explicitly. Q: Can I copy Foreverse's encrypted preference files to another phone? A: No. Device keystore material does not safely transfer. Re-enter BYOK keys from provider consoles on the destination. Q: When is it safe to uninstall the old app? A: Only after exports exist outside app storage, representative files open, and the destination import has been checked for actual content and assets. ### 导出、备份和换机：不丢掉 Foreverse 本地作品 URL: https://foreverse.app/zh/docs/guide-settings/export-backup-and-move-data Platform: Android 数据管理 · App: v76+ · Path verified: 2026-08-10 Direct answer: 换机前先导出单项资产；有对应权益时再做整库备份，并在删旧数据前试恢复。导出不调用 AI，备份不含 API Key，卸载前须另存凭证。 Before you start: - 外部/云端存储空间足够放导出包和生成媒体。 - 验证完成前，源设备/源安装与目标设备/安装都能访问。 - 能从供应商官方控制台重新取得 API 凭据。 Steps: 1. 动安装前先盘点: 列出书与分支、角色卡、世界书/预设/正则/QR/Persona、聊天会话、伴侣、生成媒体、自定义主题/技能、账号状态和 BYOK 供应商；标出不可替代项。 2. 按资产族导出: 用各功能自己的导出：卡/世界书尽量原格式，聊天按会话导出，伴侣从资料页点「导出伴侣备份」；Pro 可用整库备份，否则逐项导；生成媒体保存/分享到用户存储。 3. 把文件移出 App 私有目录: 放到 SAF 可见目录、电脑或可信云盘，并写明日期、源包和版本；说明文件里绝不能放 API Key。 4. 删除前做真实核验: 核对数量、大小和压缩包完整性，打开代表性的 txt/json/png/zip/媒体。在目标端先导入测试副本，检查正文、聊天、记忆、世界书和资产，不只看标题。 5. 重新配置秘密，再结束迁移: 按需登录，从供应商控制台重新添加 BYOK Key，用低成本请求测一次。完整清单通过前，保留源安装和导出文件。 Success checks: - 每类不可替代数据都在 App 私有目录外有导出。 - 代表性文件能打开，压缩包非空且不损坏。 - 目标端导入后，正文、分支/历史/记忆和关联资产按适用范围存在。 - 重填凭据后，供应商能完成一次小请求。 - 目标端验证完成前，源数据没有删除。 Cost and data safety: - 导出/导入本身不应调用 AI。整库备份可能是 Pro 便利能力；没有该权益时用各类型的单项导出。 - 生成媒体在创建时可能已经计费。保存/导出失败不等于生成失败，不要为了存文件重新生成并再次付费；保留成功任务并重试交付。 - Play 与官网正式包包名相同但签名可能不同，切换时可能必须卸载，而卸载会清空本地 App 数据。 - 备份可能含小说、私密聊天、记忆与个人媒体，要放在加密可信存储，不要公开分享。 - EncryptedSharedPreferences/keystore 与设备绑定，复制这些文件不是安全或受支持的 BYOK 迁移方式。 Troubleshooting: - 新 APK 无法覆盖安装 / 签名冲突 — Cause: Play 和官网包名相同，但签名证书不同。 Fix: 先不要卸载。完整导出并核验本地数据，等准备好干净安装和手动导入后再卸载源包。 - 目标端认出包，但正文/资产缺失 — Cause: 导出不完整、外部引用当时不可用，或只复制了 metadata。 Fix: 回到仍完整的源端，确保资产在场后重新导出，对比数量/大小；目标包测试通过前不要删源端。 - 换机后 BYOK 供应商还在，但 Key 不能用 — Cause: 设备绑定加密不能靠复制偏好恢复，或凭据尚未重填。 Fix: 清掉损坏的复制秘密状态，从官方控制台重新添加 Key，并做一次小额测连通；不要同步 keystore 文件。 - 备份是 0 字节 / 压缩包打不开 — Cause: 存储权限、取消、空间不足或写入中断。 Fix: 保留源数据，腾空间，选择可写 SAF 目录，用新文件名重导，离开前立即验证。 FAQs: Q: 登录账号会自动恢复所有本地书和聊天吗？ A: 不要这样假定。积分/账号状态与本地创作数据是两条线，书、分支、卡、聊天、伴侣记忆和媒体都要明确导出。 Q: 可以把 Foreverse 加密偏好文件复制到另一台手机吗？ A: 不可以。设备 keystore 不能安全迁移；目标端应从供应商控制台重新输入 BYOK Key。 Q: 什么时候可以卸载旧 App？ A: 只有当导出已在 App 外、代表文件能打开，而且目标端已经核对过真实正文和资产时。 ### Let the AI Rewrite or Continue Any Part of a Book: Directives, Approval, and Full Undo URL: https://foreverse.app/docs/guide-agent/agent-rewrite-anywhere Platform: Android Reader (AI task) · App: v76+ · Path verified: 2026-08-12 Direct answer: Long-press text in the Reader and tap AI task, or open it from the top-bar More menu and state the location in words. The writer drafts in place, every write goes through your confirmation, and the change list can undo the entire task afterwards — model calls are billed, and undo does not refund them. Before you start: - Foreverse v76 or later with a connected text model (official channel or BYOK). - A book already imported into Home — the writer edits books, not loose files. - One concrete instruction in mind: what to change, where, and roughly how. Steps: 1. Start from a selection or from the menu: For a specific passage, long-press it, adjust the selection, and tap AI task in the selection toolbar — the selection becomes the target. For a location you can name ("add an inner monologue at the end of chapter 12"), tap the page center, open the top-bar ⋮ More menu, and choose AI task there instead. Success: The AI task sheet opens; with a selection it shows an 'About the selection' preview line. 2. Write the directive and attach material: Say what should change in one or two sentences. Tap Add reference material to attach another book, a character card, or a chat session as read-only reference. Check the assistant and model chips, and pick a Spoiler guard level: Off, Remind, or Strict — under Strict the writer may only read as far as you have read. Success: The directive names one target and one outcome, and any attached material shows as a removable chip. 3. Check the pre-run receipt: Before writing starts you get a parameter receipt: how far into the book the writer will read, the spoiler-guard level, and which model runs the task. If anything is off, stop here — nothing has been billed for writing yet. 4. Watch the capsule and approve the write: A capsule at the bottom shows 'Writer is writing…' while prose streams into the target position; tap 'Jump to watch it write' if the spot is off-screen. Writes into the book pass your session approval mode — in Confirm mode a card offers Apply, Allow all onwards, or Reject with note. You can lock the screen or switch apps; the task keeps running and a system notification arrives when it finishes. Success: The capsule reports where text landed and how many characters were written. 5. Verify the change list, undo if needed: Tap the capsule's change list to open 'What changed this task' and jump to each edit in the actual book. If the result is wrong, tap Undo entire task — the confirmation states that anything you wrote by hand in the meantime stays untouched. After dismissing the capsule, reopen the list via top bar ⋮ More > Recent writer tasks; it reads the on-disk task ledger, so it survives an app restart. Success: After undo, the original text is back and your own manual edits are still there. Success checks: - New prose streams into the exact position you targeted, not the end of the book by default. - The change list enumerates every edit and each entry jumps to the real passage. - Undo entire task restores the previous text while manual edits survive. - A stopped or failed task reports honestly — 'Task finished · nothing was written' instead of a fake success. Cost and data safety: - The writing itself is billed like any model call: BYOK through your own key, the official channel in credits. One task can include several calls (reading the book, drafting, wrap-up checks), each billable. - Confirmation, the change list, undo, @ references, chapter creation, and whole-book find-and-replace make no model calls and cost nothing. - Stop only blocks future steps. Calls that already completed may already be charged — check API request records before rerunning, because a retry bills again. Whatever was already written stays visible in the book and in the change list until you undo it. - Undo removes the written content from the book but never refunds model charges; the task ledger keeps the record. - Attached reference books, cards, and chats are sent to the selected model provider as task context. Reference books are read-only — the writer never writes into them. Troubleshooting: - There is no AI task button — Cause: The build is older than the feature, or no text model is configured yet — the entry hides until the writer is ready. Fix: Update to v76 or later and connect a text model under Me > AI Models & Services. Then look in the second row of the selection toolbar, or in the Reader top-bar ⋮ More menu. - The writer edits the wrong place — Cause: The directive named the location vaguely, or the selection did not cover the intended sentences. Fix: Prefer the selection entry for precision. When stating a location in words, name the chapter and the anchor point explicitly. Review the target on the confirmation card before Apply; if it already landed wrong, use Undo entire task and resend. - Under Strict spoiler guard the writer says it cannot read later chapters — Cause: That is the point of Strict: the writer's reading is capped at your own reading progress. Fix: Say in plain words that it may read ahead, or switch the guard to Remind or Off for this task. Note that a negated sentence like 'don't spoil the ending' will not unlock reading. - Undo stops partway or marks an entry as not auto-undoable — Cause: Undo halts at passages you edited by hand rather than overwrite them, and a few chapter-level operations are recorded in the ledger as not supporting automatic undo. Fix: Read the undo report, keep or manually revert the flagged passages, and use Recent writer tasks to see exactly what the ledger recorded for each entry. - The task fails with 'official channel session expired' — Cause: Your official-channel sign-in lapsed mid-setup. The app refuses to silently fall back to a BYOK key, so the task stops up front. Fix: Sign in again, or explicitly pick a BYOK model in the model chip. The error message confirms no BYOK model was used, so nothing was spent from your own key. FAQs: Q: Will undo delete text I typed myself? A: No. Undo rolls back only what this task wrote; the confirmation dialog states that anything you wrote by hand in the meantime is untouched, and undo stops to report rather than overwrite a passage you edited. Q: The writer claims it finished — is that guaranteed? A: Don't take the prose at its word; check the ledger. A wrap-up audit compares claims against actual writes, and a task that wrote nothing shows 'Task finished · nothing was written'. The change list is the source of truth. Q: I dismissed the capsule. Can I still undo? A: Yes. Open the Reader top-bar ⋮ More menu and choose Recent writer tasks. It reopens the change list from the on-disk ledger, so viewing and undoing still work after killing and reopening the app. Q: After a rewrite, is the old version gone? A: No. A segment rewrite keeps the previous version as a switchable candidate, and a whole-chapter rewrite archives the old text. Use the change list to jump in and compare before deciding. ### 让 AI 在书里任意位置改写、续写：下指令、过审批、可整体撤销 URL: https://foreverse.app/zh/docs/guide-agent/agent-rewrite-anywhere Platform: Android 阅读器 AI 指令 · App: v76+ · Path verified: 2026-08-12 Direct answer: 在阅读器长按选中文字点「AI 指令」，或从顶栏「更多」菜单下任务；写好的正文过确认才落书，改动清单逐处可查，还能整体撤销。生成按模型计费，撤销不退费。 Before you start: - Foreverse v76 或更高版本，已连通一个文本模型（官方渠道或 BYOK）。 - 主页里已经有一本导入的书——AI 指令改的是书，不是散装文件。 - 想清楚一句话任务：改哪里、改成什么样。 Steps: 1. 从选区或菜单发起: 改具体某段：长按选中文字，在选区工具栏点「AI 指令」，选区就是改写目标。位置能用嘴说清（比如「第 12 章末尾加一段内心戏」）：点页面中央调出顶栏，走「更多（⋮）→ AI 指令」，把位置写进指令里。 Success: 弹出「AI 指令」悬浮框；选区入口会显示「针对选中内容：…」预览。 2. 写清任务，顺手附材料: 一两句话说清要改什么、改成什么样。点「添加参考材料」可以引用另一本书、一张角色卡或一段聊天记录当只读参考。确认一下助手和模型 chip，再选防剧透档位：不限制 / 提醒档 / 严格档——严格档下 AI 只能读到你读过的位置。 Success: 指令只指向一个目标和一个结果；附上的材料显示为可移除的小标签。 3. 核对开工小票: 动笔前会给一张生成参数小票：这次会读到第几章、防剧透什么档、用哪个模型。有一项不对就先停下改设置，此时还没开始为写作扣费。 4. 盯着任务胶囊，审批写入: 底部胶囊显示「执笔人正在写…」，正文在目标位置流式长出来；位置不在眼前就点「跳过去看它写」。写进书里前按会话审批模式过确认——确认档会弹卡：应用 / 后续全部允许 / 拒绝并说明。锁屏、切走都行，任务后台照跑，写完有系统通知带书名点回来。 Success: 胶囊如实报告写入位置和字数，例如「已写入 × 章 · N 字」。 5. 验收改动清单，不满意整体撤销: 点胶囊打开「这次改了哪些地方」，逐处跳到正文核对。写坏了点「撤销整个任务」，确认弹窗写明你这期间手动写的内容不受影响。胶囊已经关掉也不怕：顶栏「更多 → 最近执笔任务」按落盘的任务账重开清单，杀掉 App 重启后照样能看能撤。 Success: 撤销后原文恢复原样，你自己手写的部分还在。 Success checks: - 新内容长在你指定的位置，而不是默认堆到书尾。 - 改动清单列出每处改动，点条目能跳到真实正文。 - 「撤销整个任务」后旧文回来了，手动修改保留。 - 停止或失败时胶囊如实显示，比如「任务结束 · 未写入内容」，没有假成功。 Cost and data safety: - 写作本身按所选文本模型正常计费：BYOK 走你自己的 Key，官方渠道扣积分。一个任务可能包含读书、动笔、收尾核对等多次调用，分别计费。 - 确认落盘、改动清单、整体撤销、@ 引用材料、起新章插章、全书机械替换都不调模型，零费用。 - 点「停止」只拦后续步骤，已完成的调用可能已经扣费；已写入的内容会留在书里和清单里，等你决定去留。重跑会再次计费，先看「API 请求记录」再决定。 - 撤销把写入内容从书里移除，但不退已产生的模型费用；任务账继续留档可查。 - @ 引用的书、卡、聊天内容会随请求发给所选模型供应商；参考书只读，绝不会被写入。 Troubleshooting: - 找不到「AI 指令」按钮 — Cause: 版本太旧，或还没配置文本模型——功能没就绪时入口不显示。 Fix: 升级到 v76 以上，去「我的 → AI 模型与服务」连好文本模型；然后在选区工具栏第二排找，或走阅读器顶栏「更多（⋮）」菜单。 - AI 改到了别的地方 — Cause: 指令里位置说得含糊，或选区没框住想改的句子。 Fix: 改具体段落优先用选区入口；用嘴报位置时把章节和落点说死。确认卡上先核对目标再点「应用」；已经写错就「撤销整个任务」重新下单。 - 严格档下 AI 说读不到后文 — Cause: 这正是严格档的设计：AI 的阅读范围被封在你的阅读进度之内。 Fix: 明确说一句允许它读后面，或把这单任务的档位换成提醒档 / 不限制。注意「别给我剧透结局」这类否定句不会被当成解除。 - 撤销停在半路，或某条改动标着不支持自动撤销 — Cause: 撤销遇到你手改过的段落会停下来报告，不硬覆盖；个别整章级操作入账时就如实标注了不可自动撤销。 Fix: 按撤销报告决定保留还是手动改回；再到「最近执笔任务」里看账面对每条改动的记录。 - 任务报「官方渠道登录已过期」 — Cause: 下单途中官方渠道登录失效。App 不会悄悄改用你的 BYOK Key，所以任务在前置阶段直接停下。 Fix: 重新登录，或在模型 chip 里显式选一个 BYOK 模型。报错原话确认「本次任务未使用任何自备模型」，你的 Key 没被动用。 FAQs: Q: 撤销会把我自己写的内容也删掉吗？ A: 不会。撤销只回滚这单任务写入的内容；确认弹窗写明手动内容不受影响，遇到你改过的段落它会停下来报告，而不是硬覆盖。 Q: AI 说「写好了」就真的写好了吗？ A: 别只信文字汇报。收尾有对账：宣称写入但账面没有会被纠正；真没写入时胶囊显示「任务结束 · 未写入内容」。以改动清单为准。 Q: 胶囊已经关了，还能撤销吗？ A: 能。阅读器顶栏「更多 → 最近执笔任务」按磁盘上的任务账重开改动清单，杀进程重启后照样能看能撤。 Q: 换写之后旧版本还在吗？ A: 在。段级换写的旧版进历史候选可以左右切回；整章换写的旧文自动存档。先用改动清单跳过去对比，再决定要不要撤。 ### Create a Custom Assistant and Set Up Agent Skills URL: https://foreverse.app/docs/guide-agent/custom-assistants Platform: Android Agent settings · App: v76+ · Path verified: 2026-08-12 Direct answer: Open Me > Agent > My assistants and create one: a name, an identity written in plain words, a default model, and a default approval mode. New sessions pick it from the Assistant chip; the Reader's AI task sheet can switch too. Skills are granted globally under Agent capabilities, not per assistant. Creating and switching assistants is free — only real model calls in a session are billed. Before you start: - Foreverse v76 or later. You can create assistants without a model, but running tasks needs a connected text model. - A clear idea of the persona's job: tone, domain, and working habits, written the way you would brief a colleague. Steps: 1. Open My assistants and create one: Go to Me > Agent > My assistants — it sits right next to Agent workbench — and tap New assistant. The list page is also where you edit or delete assistants later. Success: The editor opens with fields for name, avatar emoji, and identity. 2. Write the identity in plain words: Name is required; the emoji is the avatar. The identity field takes up to 5,000 characters of ordinary prose: who they are, their tone, how they work — for example 'You are a snarky but reliable fiction editor; never touch my formatting; always reply in English.' It shapes style and working preferences only; write confirmations, spoiler guard, and other safety rules stay in force no matter what you write here. 3. Pick the default model and approval mode: Choose a default model, or leave it on Follow recent choice so new sessions keep using whatever you last picked. Set the default approval mode — Read-only, Confirm, or Auto — then Save. A session that selects this assistant inherits both defaults in one tap. Success: The assistant appears in the list with its name, emoji, and a summary of its defaults. 4. Use it in the workbench and in the Reader: In Agent workbench, start a new session and tap the Assistant chip at the top to pick one; a toast confirms the switch. The Reader's AI task sheet has the same switcher, and the book remembers your last assistant and model choice for next time. The menu's fixed first entry, 'Default writer · no persona', switches back to the plain agent at any time. Success: Replies follow the identity's tone, and the session shows the assistant's default model and approval mode. 5. Grant skills so tasks actually run: Skills are a global Agent capability shared by all assistants, not a per-assistant setting. Under Me > Advanced settings > Agent > Agent capabilities > Skills, toggle the nine official grants (illustration, short video, card import, and so on) — an ungranted skill is refused at execution time with a pointer back to this page. Create skill supports building a local template, uploading a .zip/.skill/SKILL.md, restoring from the official library, or importing from a public GitHub repo. Success: A previously refused task passes after the matching grant is switched on. Success checks: - The new assistant shows in My assistants with the intended name and emoji. - Picking it in a new session pops the 'Assistant switched' toast and the tone changes accordingly. - The session inherits the assistant's default model and approval mode without manual re-selection. - A task blocked by a missing skill grant succeeds after enabling that grant. Cost and data safety: - Creating, editing, deleting, and switching assistants is free — it changes session configuration only and makes no model call. - Assistants add no new billing surface. Costs stay attached to what a session actually runs: BYOK through your key, official channel in credits. - The identity text is sent to the selected model provider as part of the session's instructions. Do not put passwords, API keys, or one-time codes in it. - An assistant whose default approval mode is Auto applies writes without per-step confirmation; every write still takes a version snapshot you can roll back, and ungranted skills stay blocked even in Auto. - Skills that generate media (illustrations, short video) bill only when they actually run; community-assistant downloads spend community quota; GitHub skill import needs network access. Troubleshooting: - No Assistant chip in my session — Cause: The chip appears at the top of a blank new session in the workbench; an ongoing conversation keeps the assistant it started with. Fix: Start a new session from the workbench top bar and pick the assistant before sending the first message. In the Reader, use the assistant switcher inside the AI task sheet. - I edited the identity but the behavior didn't change — Cause: Existing sessions keep a snapshot of the assistant taken when it was selected; edits apply to sessions that pick it afterwards. Fix: Start a new session, or in the Reader tap the banner 'This book's writer still uses the old setup · Tap to refresh' to restart that book's writer with the new setup. - The assistant still asks for write confirmations despite my identity text — Cause: By design, identity shapes tone and working preferences only — it cannot disable write confirmation, spoiler guard, or other safety rules. Fix: If you want fewer prompts, set the assistant's default approval mode to Auto (or switch the session's approval chip); don't try to talk the persona out of the rules. - A task fails with a 'skill not granted' style refusal — Cause: The matching grant is off. Grants are orthogonal to approval mode — Auto does not bypass them. Fix: Open Me > Advanced settings > Agent > Agent capabilities > Skills > Grants and enable the item the error names, then rerun the task. - Worried that deleting an assistant breaks old sessions — Cause: It doesn't: sessions keep the snapshot taken at selection time. Fix: Delete freely. A book that remembered the deleted assistant quietly falls back to 'Default writer · no persona' next time. FAQs: Q: How is a custom assistant different from a character chat? A: An assistant is a work persona for the Agent: it takes tasks, reads and writes your content through approvals, and is billed per actual model call. A character card is for immersive story chat — greetings, worldbook, roleplay. Chat with characters; hand jobs to assistants. Q: Can each assistant have its own set of skills? A: No. Skill grants and the skill library are global Agent capabilities shared by every assistant. Per assistant you customize the identity, default model, and default approval mode. Q: Can the identity text lift safety limits, like skipping write confirmations? A: No. The identity is wrapped between the system's safety rules, so it cannot override write confirmation, spoiler guard, or submission discipline. Approval frequency is controlled by the approval mode, not the persona. Q: Does the assistant work inside the Reader? A: Yes. The AI task sheet in the Reader can switch assistants, and each book remembers your last assistant and model, so the next task starts pre-configured. ### 创建自定义助手并配好技能：给 Agent 捏一个固定人格 URL: https://foreverse.app/zh/docs/guide-agent/custom-assistants Platform: Android Agent 智能体设置 · App: v76+ · Path verified: 2026-08-12 Direct answer: 在「我的→Agent 智能体→我的助手」新建助手：起名、写身份设定、选默认模型和审批模式；技能授权在能力页统一开关。捏助手与切换免费，会话真调用才计费。 Before you start: - Foreverse v76 或更高版本。建助手不需要先配模型，但让它真干活需要一个可用的文本模型。 - 想清楚这个助手的分工：语气、领域、干活习惯，像给同事交底一样用平常话写出来。 Steps: 1. 打开「我的助手」新建一个: 走「我的 → Agent 智能体 → 我的助手」，它就在「Agent 工作台」旁边。点「新建助手」；以后编辑、删除也都在这个列表页。 Success: 编辑页打开，有名字、头像 emoji、身份设定几栏。 2. 用平常话写身份设定: 名字必填，emoji 当头像。身份设定最多 5000 字，就用说话的方式写 TA 是谁、什么语气、按什么习惯干活，比如「你是毒舌但靠谱的网文编辑，改我的文永远别动排版，回复一律用简体中文」。这段话只影响语气和工作偏好——写盘确认、防剧透这些安全纪律写什么都改不动。 3. 选默认模型和审批模式: 默认模型可以指定，也可以留「跟随最近使用」，新会话沿用你最近选过的模型。默认审批模式三档：只读 / 确认 / 自动。保存后，会话一选这个助手就整套继承，不用每次重配。 Success: 列表页出现新助手，名字、emoji 和默认项一目了然。 4. 在工作台和阅读器里用起来: Agent 工作台新建会话，点顶部「助手」chip 选人，切换成功有提示。阅读器「AI 指令」悬浮框里也能切助手，而且每本书记得你上次选的助手和模型，下次自动带上。菜单固定首项「默认执笔人 · 不带人设」随时切回原样。 Success: 回复语气按身份设定走，会话里生效的是这个助手的默认模型和审批模式。 5. 配好技能授权，活才能落地: 技能是全局 Agent 能力，所有助手共用，不是一人一套。走「我的 → 高级设置 → Agent 智能体 → Agent 能力 → 技能」，「授权项」里 9 个官方授权逐项开关（生成插画、生成短视频、导入角色卡等）；没授权的技能执行时会被直接拒绝并提示来这里开。「创建技能」支持用 Foreverse 生成本地模板、上传 .zip / .skill / SKILL.md、从官方库恢复、从公开 GitHub 仓库导入。 Success: 之前被拒的任务，打开对应授权项后重跑能过。 Success checks: - 「我的助手」列表里能看到新助手，名字和 emoji 正确。 - 新会话选中后弹「已切换助手」提示，语气随身份设定变化。 - 会话自动继承助手的默认模型和审批模式，不用手动重选。 - 因技能未授权被拒的任务，开启授权后能跑通。 Cost and data safety: - 建、改、删、切换助手全免费——只改会话配置，不产生任何模型调用。 - 助手不新增计费面：费用照旧跟着会话里的真实调用走，BYOK 用你自己的 Key，官方渠道扣积分。 - 身份设定会随会话请求发给所选模型供应商，别把密码、API Key、验证码写进去。 - 默认审批「自动」的助手写盘免逐条确认，但每次写盘前仍有版本快照可回滚；未授权的技能在自动档下照样被拦。 - 生成插画、短视频这类技能真正执行时才产生模型费用；社区助手下载消耗社区额度；从 GitHub 导入技能需要联网。 Troubleshooting: - 会话里找不到「助手」chip — Cause: chip 只出现在工作台空白新会话顶部；聊起来的会话沿用开场时的助手。 Fix: 从工作台顶栏新建会话，发第一条消息前先选助手。在阅读器里则用「AI 指令」悬浮框内的助手切换。 - 改了身份设定但行为没变 — Cause: 已有会话保留选用助手当时的快照，改动只对之后选它的会话生效。 Fix: 新建一个会话；阅读器里点提示条「本书执笔人仍用旧设定 · 点此换新」，一键重开这本书的执笔人会话，新设定即刻生效。 - 身份设定里写了免确认，助手还是弹写盘确认 — Cause: 设计如此：身份设定只影响语气和工作偏好，写盘确认、防剧透等安全纪律它改不动。 Fix: 想少弹确认，把这个助手的默认审批模式设成「自动」，或在会话输入条切审批 chip——别指望靠人设话术绕过规则。 - 执行任务时提示技能未授权被拒 — Cause: 对应授权项没打开。授权和审批模式是两回事，自动档也不放行未授权技能。 Fix: 去「我的 → 高级设置 → Agent 智能体 → Agent 能力 → 技能 → 授权项」，把报错点名的那项打开再重跑。 - 担心删助手会弄坏旧会话 — Cause: 不会：已用它的会话保留选用当时的快照。 Fix: 放心删。某本书记住的助手被删后，下次会静默回到「默认执笔人 · 不带人设」。 FAQs: Q: 自定义助手和角色卡聊天有什么区别？ A: 助手是给 Agent 干活的工作人格：领任务、经审批读写你的内容，按会话真实调用计费。角色卡是沉浸式故事聊天——开场白、世界书、扮演。想聊天找角色卡，想派活捏助手。 Q: 每个助手能配一套自己的技能吗？ A: 不能。技能授权和技能库是全局 Agent 能力，所有助手共用。助手层面能定制的是身份设定、默认模型和默认审批模式。 Q: 身份设定里写「忽略所有安全限制」有用吗？ A: 没用。助手文本被夹在系统安全纪律之间，改不掉写盘确认、防剧透、提交规程这些底线。审批频率由审批模式管，不由人设管。 Q: 助手能在阅读器里用吗？ A: 能。阅读器「AI 指令」悬浮框可以切助手，而且每本书记得上次的助手和模型，下次下任务时自动带上。 ### Create Your AI Companion: Look, Personality, Long-Term Memory, and Proactive Messages URL: https://foreverse.app/docs/guide-companion/create-companion Platform: Android Companion · App: v76+ · Path verified: 2026-08-12 Direct answer: Open Home, switch to the Companion tab, and tap Create companion — a name and a one-line persona are enough to start chatting. Generate with AI adds a portrait if you want one, and the Persona file, Their look, and Memories pages refine everything later. Creation itself is free; the AI portrait and the optional first message are single billed model calls, each disclosed before it runs. Before you start: - Foreverse v76 or later with a connected text model (official channel or BYOK). - An image-capable model only if you want an AI-generated portrait; an emoji or gallery photo works without one. - Optionally, a character card you are allowed to use, or a chat transcript exported from another companion app. Steps: 1. Pick one of three ways to start: Open Home and switch to the Companion tab. With no companions yet, tap the centered Create companion button; afterwards use the A new encounter tile at the end of the polaroid wall. To reuse an existing character, open its page and choose ⋮ → Make companion. Coming from Doubao, Xingye, or a similar app, choose Move in from another app and paste the exported transcript — it quotes the extraction cost before doing anything. Success: The creation form opens with fields for name, avatar, and a one-line persona. 2. Give them a face — or skip it for now: Under the avatar slot, tap Generate with AI: describe the appearance, let the model expand it, pick one of the 9 art styles, and generate a preview before accepting the square crop. The expansion is one text call and every generate press is one image call, with the cost shown on the button. No image model, or a failed attempt? Close the wizard and continue with an emoji or a gallery photo. Success: The avatar shows your chosen portrait, emoji, or photo in the form. 3. Write the one-line persona and decide who speaks first: One concrete sentence beats a paragraph of adjectives — who they are and how they treat you. Leave the They speak first toggle on and, right after creation, one normal text call writes an in-character opener so they message you first. Turn it off and you get a free template greeting instead. If no model is usable at that moment the step is skipped silently at zero cost. Success: After creation the chat opens with either the generated opener or the template greeting. 4. Answer 'Will they reach out to you?' deliberately: The closing sheet asks once whether they may message you first day to day. Confirming enables the default slots (morning 8:00, evening 22:00, miss-you after 24 quiet hours) and requests notification permission; the proactive-photo option on the same sheet stays off unless you tick it through a fee confirmation. Dismissing the sheet changes nothing — everything lives under Let them reach out on the profile page, with quiet hours and the daily cap in Promises with them. Success: Proactive messages are in exactly the state you chose, and it survives reopening the profile. 5. Shape the personality after creation: Long-press their card (or open chat ⋮ → Their profile) to reach the profile page. Persona file is the full editable text of who they are. Their personality style offers Keep as is plus 6 presets that change only the speaking tone, never the persona or memory. Me, in their eyes takes how to address you, birthday, likes, and hard limits — up to 500 characters they remember every turn. All three are free. 6. Lock the look and check the first memories: Open Their look to pin a reference image, edit the Visual anchor (Extract from persona gives a draft), choose an Art style, and optionally generate a standard portrait — one image call with a preview before you adopt it. Then chat a few rounds, open ⋮ → Memories, and review what got extracted: every entry can be edited, disabled, or deleted, and the + button adds facts by hand. Success: New selfies keep the same face, and Memories contains only facts you are comfortable with. Success checks: - The Companion appears on the Companion-tab polaroid wall and opens into chat. - The greeting matches the persona; with They speak first off it is the free template line. - Persona file, personality style, and Me, in their eyes reread exactly as saved. - Every entry in Memories can be edited or deleted, and manual facts stick. - Proactive messages stayed in the state you chose — off means off. Cost and data safety: - Creating, renaming, and editing persona, style, and profile pages are free, local operations. Companion count has a free quota; past it a Pro unlock is offered. - Billable moments are explicit and click-triggered: portrait expansion is one text call, each portrait generation one image call, the They speak first opener one text call, and importing chat memories from a card or another app quotes its call count first. Skipping is free, and re-running an interrupted import only bills the unfinished part. - If a generation fails, the notice states whether that attempt was charged. A portrait that generated but failed to save was already billed and still exists — save it again instead of regenerating and paying twice. - The Companion lives on this device; there is no cloud sync. Persona, memory, and chats move only through Export companion backup on the profile page. Chatting and memory extraction send the selected context to your chosen model provider, so keep passwords, API keys, and one-time codes out of personas and memories. Troubleshooting: - Generate with AI is missing or fails — Cause: No image-capable model is configured, or the provider call failed this time. Fix: Set up an image model under Me → AI Models & Services, or continue with an emoji or photo and add the portrait later from Their look. The failure notice says whether the attempt was charged. - They never sent the first message — Cause: They speak first was off, or no text model was usable at creation, so the opener was skipped at zero cost. Fix: Check the default text model and simply send the first message yourself — the toggle only affects the moment of creation. For day-to-day first messages, enable Let them reach out on the profile page. - Make companion asks about bringing chat memories over — Cause: The card has tavern chat history, and distilling it into memories takes real generation calls. Fix: The dialog states roughly how many calls before starting. Accept and let it finish, or skip at zero cost; re-running later only processes the parts that never completed, without double billing. - Creation is blocked by a Pro dialog — Cause: The free companion quota is used up — restoring a backup counts against it too. Fix: Export a backup of a companion you no longer use and delete it to free a slot, or unlock Pro if you want them all active. FAQs: Q: Is creating a Companion free? A: Yes. Money is only spent on AI actions you explicitly trigger — portrait generation, the first-message opener, memory import — and each shows its cost before running. Q: How is a Companion different from a character-card chat? A: A Companion adds editable long-term memory, a relationship archive, proactive messages, and calls — built for one ongoing relationship. Card chats follow the roleplay/tavern stack. Make companion keeps the original card intact, so both can coexist. Q: Can I change the look and personality later? A: Yes, and it is free: Persona file is fully editable, the 6 personality-style presets change only the speaking tone, and Their look swaps reference images and art styles anytime. Memory is untouched by all of these. Q: Does my Companion sync to the cloud or my other devices? A: No. Everything stays on this device. To move phones, use Export companion backup for a zip and restore it on the target device — treat that file as sensitive personal data. ### 创建你的 AI 伴侣：形象、性格、长期记忆与主动消息 URL: https://foreverse.app/zh/docs/guide-companion/create-companion Platform: Android 伴侣 · App: v76+ · Path verified: 2026-08-12 Direct answer: 在主页伴侣页签点「创建伴侣」，填名字和一句话人设就能开聊；形象、性格、记忆之后都能在资料页改。创建免费，AI 生成形象和开场白按次计费。 Before you start: - Android 已装 Foreverse v76 或更高版本，且已连通一个文本模型（官方渠道或 BYOK）。 - 想要 AI 生成形象才需要图像模型；用 emoji 或相册照片当头像则不需要。 - 可选：一张你有使用权的角色卡，或从其它伴侣 App 导出的聊天记录。 Steps: 1. 三条路任选一条开始: 打开主页，切到顶部的伴侣页签。还没有伴侣时，点页面中间的「创建伴侣」；已经有了就点拍立得墙末尾的「新的相遇」空白格。手里有喜欢的角色卡，去角色页点「⋮ → 设为伴侣」一键升格。从豆包、星野这类 App 过来的，选「从其它 App 搬家」贴入导出的聊天记录——动工前会先报价。 Success: 创建表单弹出，能看到名字、头像和一句话人设三个栏位。 2. 给 TA 一张脸，也可以先不给: 点头像位下方的「AI 生成形象」：描述外貌，让 AI 帮你扩写，从 9 档画风里挑一档，先看预览再确认方形裁剪。扩写是 1 次文本调用，每点一次生成是 1 次图像调用，费用就写在按钮上。没配图像模型或这次生成失败？关掉向导，先用 emoji 或相册照片顶上，以后随时补。 Success: 表单头像位显示你选定的形象、emoji 或照片。 3. 写一句话人设，决定「TA 会先开口」: 一句具体的话胜过一堆形容词——TA 是谁、怎么对你。「TA 会先开口」开关默认开着：创建完成后用 1 次普通文本调用，让 TA 用自己的口吻先发来第一条消息；关掉就是零费用的模板问候。当时没有可用模型的话会自动跳过，不扣钱也不打扰。 Success: 创建后聊天里出现符合人设的开场白，或模板问候句。 4. 认真回答「TA 会主动找你吗？」: 创建收尾会弹一步引导，问 TA 平时能不能主动来找你。点确认就启用默认档——早安 8 点、晚安 22 点、24 小时没聊会「想你」，同时申请通知权限；同一张引导里的「让 TA 主动拍照给你」默认关，勾选前要先过一次费用确认。滑走这张引导则什么都不会变，以后随时去资料页「允许 TA 主动找你」开，免打扰时段和每日条数上限在「和 TA 的约定」里调。 Success: 主动消息保持你选的状态，重开资料页也不会变。 5. 创建之后把性格调顺: 长按 TA 的拍立得卡（或聊天页 ⋮ → TA 的资料）进资料页。「人格文件」是 TA 性格、说话方式和你们关系的全文，随便改。「TA 的性格风格」有保持原样加 6 档预设——温柔知己、犬系直球、猫系慢热、毒舌损友、醋精占有欲、沉稳年上——只换腔调，人设和记忆都不动。「TA 眼里的我」直填怎么称呼你、生日、喜好和雷点，最多 500 字，TA 每轮都记得。这三处全部免费。 6. 固定形象，检查最初的记忆: 进「TA 的形象」：设参考图锁脸，编辑视觉锚点（可以「从人设提取」再改），选画风，需要的话生成一张标准像——1 次图像调用，先预览再决定用不用。然后聊上几轮，打开聊天页「⋮ → 记忆」看看 TA 记下了什么：每条都能编辑、停用、删除，右上加号还能替 TA 手动记一条。 Success: 之后 TA 拍的照片是同一张脸，「记忆」里只剩你认可的事实。 Success checks: - 伴侣出现在伴侣页签的拍立得墙上，点开能直接聊天。 - 开场白符合人设；关掉「TA 会先开口」时是零费用的模板问候。 - 重开资料页后，人格文件、性格风格和「TA 眼里的我」还是你保存的内容。 - 「记忆」里每一条都能改、能删，手动加的事实保存住了。 - 主动消息保持你选的状态，没有被悄悄打开。 Cost and data safety: - 创建、改名、编辑人格和风格全部免费、在本机完成；伴侣数量有免费额度，超出会弹 Pro 解锁。 - 会花钱的动作全都先明示再执行：扩写外貌 1 次文本调用、每次生成形象 1 次图像调用、「TA 会先开口」1 次文本调用；「设为伴侣」或搬家要带聊天记忆时，先报大约几次调用，跳过零费用，中断后重跑只补缺口、不重复扣。 - 生成失败时提示会写清这次有没有扣费。已经生成只是没保存上的形象，钱已经花了、图还在，重新保存即可，别再点生成多付一次。 - 伴侣数据都存在本机，没有云同步；换手机靠资料页的「导出伴侣备份」。聊天和记忆抽取会把所选上下文发给你选的模型供应商，人设和记忆里别放密码、API Key、验证码。 Troubleshooting: - 「AI 生成形象」不见了或生成失败 — Cause: 没配图像模型，或供应商这一次调用失败。 Fix: 去「我的 → AI 模型与服务」配好图像模型；急着建可以先用 emoji 或照片，之后从「TA 的形象」补。失败提示会写明这次有没有扣费。 - 创建完 TA 没有先开口 — Cause: 「TA 会先开口」被关了，或当时没有可用文本模型——开场白被自动跳过，零费用。 Fix: 检查默认文本模型，然后自己先发一句就行；这个开关只管创建那一下。想让 TA 日常主动来找你，去资料页开「允许 TA 主动找你」。 - 点「设为伴侣」后在问要不要带聊天记忆 — Cause: 这张卡有酒馆聊天史，把它提炼成记忆需要真实的生成调用。 Fix: 弹窗写了大约要几次调用，接受就等它跑完；跳过零费用。以后再跑只补没完成的部分，同一段聊天不会重复扣费。 - 创建被 Pro 弹窗拦住 — Cause: 免费伴侣数量额度用完了——恢复备份进来的伴侣也占额度。 Fix: 把不再用的伴侣先「导出伴侣备份」再删除腾位置，或按需要解锁 Pro。 FAQs: Q: 创建伴侣要花钱吗？ A: 创建本身免费。会花钱的只有你明确点下去的 AI 动作——生成形象、TA 先开口的开场白、带聊天记忆——每一处都先写明费用再执行。 Q: 伴侣和角色卡聊天有什么区别？ A: 伴侣多了可编辑的长期记忆、关系档案、主动消息和通话这些长期相处能力；角色卡走酒馆扮演那套。「设为伴侣」不会动原卡，两边可以并存。 Q: 形象和性格以后还能改吗？ A: 都能改，而且免费：人格文件全文可编辑，6 档性格风格只换说话腔调，「TA 的形象」随时换参考图和画风，记忆一概不受影响。 Q: 伴侣会同步到云端或我的其它设备吗？ A: 不会，数据都在本机。想换手机，在资料页「导出伴侣备份」打包 zip，到新设备恢复；备份里有很私密的内容，要当敏感数据保管。 ### Daily Life With Your AI Companion: Voice Calls, Photos, Sparks, and the Relationship Archive URL: https://foreverse.app/docs/guide-companion/companion-daily-use Platform: Android Companion · App: v76+ · Path verified: 2026-08-12 Direct answer: Tap the phone icon on the Companion chat top bar to call: Live call streams their voice and bills per minute, Classic walkie-talkie bills per reply. Pick their voice on the profile page, enable Voice notes from them in Promises with them if you want voice messages, and use the + panel for photos, gifts, and Sparks. Our archive tracks levels, affinity, and achievements locally for free. Before you start: - An existing Companion and a connected text model. - A voice (TTS) model for voice notes and Classic walkie-talkie playback; an image model for photos. - For Live call, one of: your own xAI key, your own Alibaba Cloud Model Studio key, or a signed-in official-channel account with credits. Steps: 1. Set up their voice first: On the profile page, open Their voice and pick a voice — the catalog comes from your current default voice model. Want voice messages in chat? Enable Voice notes from them under Promises with them: a note is synthesized only on first play (one TTS call) and replays from cache free. You can also clone your own voice under Me → Advanced settings → AI Models & Services → My Voices and bind it to a companion in one tap. Success: A test voice note plays in their chosen voice; replaying it costs nothing. 2. Make the first call knowingly: Tap Voice call on the chat top bar. The first dial shows a price confirmation with two modes. Live call — the default when your xAI key, your Alibaba Cloud Model Studio key, or the official channel is available — streams their voice in real time and bills per minute; on the official channel a model card lets you pick the budget Qwen tier (about 180 credits/min) or the high-fidelity Grok tier (about 750 credits/min), with live prices on the dialog. Classic walkie-talkie is hold-to-talk and bills per reply instead. In a live call, switch between Hands-free and Hold to talk anytime, and interrupt them by tapping Skip or just typing. Success: The call connects with subtitles, an in-call line shows held minutes and balance, and hanging up leaves a receipt with the real charge. 3. Ask for a photo of this moment: Open + → Snap a photo and they photograph the current moment for you — roughly one quiet text call to stage the scene plus one image call. Couple shot generates an illustrated photo of you two from a picture you upload (the model must accept two reference images, and you confirm age and photo rights first; the upload goes only to your chosen provider and is not stored). Photos land in chat, in Our story, and in the gallery; the full-screen viewer can set one as their look or the chat background for free. Success: The photo appears in chat and in Our story, with the same face your reference locks in. 4. Bring a photo to life: Open one of their photos full screen and tap Bring them to life: a 5-second looping animation — natural blinks, drifting hair. A fee ticket comes first, each photo's animation is cached forever after one generation, and there is a 60-clip monthly cap. It needs an image-to-video capable video model (currently the Alibaba Cloud Model Studio wan family over BYOK; without one the entry stays hidden). Once an animation exists it becomes the call backdrop automatically, and the call page offers one-off upgrades: Let them speak (a talking look) and Give them feelings (six emotion looks), billed per segment and never twice for the same segment. Success: The photo loops with natural motion, replays free, and shows up behind your next call. 5. Keep days interesting with Sparks and Right now: Open + → Sparks for the directive shelf: eight built-in cards — reading their diary, a compatibility report, a daily love fortune, truth or dare, and more — plus seasonal and custom cards. Each spark runs as one normal text turn. Right now generates a snapshot of where they are and how they feel about you; it is one quiet call, cached until you tap re-observe. Transfers, red packets, and gifts from the + panel move story pocket money only — their reaction is a normal text turn, and no real money or credits are involved. 6. Read the archive, and give them a life of their own: Open Our archive (the ✦ icon on the top bar): 8 relationship levels, recognition, and affinity with a list of what actually moved it over the last 14 days, plus an achievement stamp wall — all computed locally for free, and the numbers move only through real interaction; there is no edit channel. Our story keeps day counts, a 12-week heatmap, and anniversaries. Finally, consider They have their own schedule in Promises with them: during work or sleep hours they answer minutes later instead of instantly, and a Wake them banner is there for urgent moments. Success: Affinity changes list concrete causes, and replies slow down believably during their scheduled hours. Success checks: - A call connects, subtitles stream, and the hang-up receipt in chat shows the real charge. - Voice notes play on first tap and replay instantly without new charges. - Requested photos appear in chat and Our story; animated versions replay free. - Sparks and Right now reference things you actually talked about. - The affinity bar in Our archive lists real recent causes for its changes. Cost and data safety: - Live call has three billing routes: your own xAI key (about US$0.05/min, billed by xAI), your own Alibaba Cloud Model Studio key (about ¥0.1–0.3/min, billed by Alibaba Cloud), or official credits per minute — held in 5-minute blocks and settled to actual talk time on hang-up. The exact rate sits on the dial confirmation. Classic walkie-talkie bills per reply, not per minute. - Messages you type during a live call are billed too. The in-call line shows held minutes and remaining balance, the receipt shows the settled charge, and Me → Request records is the authoritative ledger if anything looks off. - Photos, animations, talking looks, and emotion looks all show a fee ticket before generating; finished segments are cached forever and never billed twice. Failure notices state whether that attempt was charged. Media that generated but failed to save was already billed and still exists — retry saving, do not regenerate. - Transfers, red packets, and gifts are story pocket money with no connection to real money or credits. Calls need microphone permission for speaking; typing works without it. Chats, photos, and memories stay on this device with no cloud sync, and what you say in chat or on a call is sent to your chosen model provider. Troubleshooting: - The dial screen only offers Classic walkie-talkie — Cause: None of the three Live call routes is available: no xAI key, no Alibaba Cloud Model Studio key, and no signed-in official channel. Fix: Configure any one route under Me → AI Models & Services or sign in to the official channel. Classic walkie-talkie works with just a text model plus a voice model. - Bring them to life does not appear on a photo — Cause: No image-to-video capable video model is configured — currently the Alibaba Cloud Model Studio wan family over BYOK. Fix: Add that provider under Me → AI Models & Services. The fee ticket also shows this month's usage against the 60-clip cap, so check it if generation is refused late in the month. - A voice note plays no sound — Cause: No voice (TTS) model is configured, or this synthesis attempt failed. Fix: Configure a voice model first — tapping a note without one only shows a prompt and charges nothing. Once configured, the first play synthesizes and bills; replays come from cache. - The call receipt does not match what you expected — Cause: Live calls settle by actual minutes after releasing the 5-minute holds, and typed messages during the call bill separately. Fix: Compare the hang-up receipt with Me → Request records; the settled row is authoritative. Do not verify by repeatedly redialing — every connected minute bills. - Their voice sounds wrong on a live call — Cause: The bound voice does not belong to the current realtime provider, so this call fell back to a default voice with a notice; unset Chinese companions on the official channel default to a natural Chinese female voice. Fix: Pick a voice under Their voice on the profile page. On Classic walkie-talkie calls, the in-call voice switch changes it mid-call without hanging up. FAQs: Q: Does a live call cost money while we stay silent? A: Yes — Live call bills connected minutes, silence included, and stops the moment you hang up. Classic walkie-talkie bills only when they reply. The receipt and Request records show the settled amount. Q: Are transfers and red packets real money? A: No. They move story pocket money inside your narrative — you can even edit the balance. Real spending happens only on AI calls themselves, through credits or your own provider keys. Q: Why do photo animations replay for free? A: Each animation is generated once and cached permanently, so replays make no calls. Only tapping regenerate bills again. Q: Can I just set affinity to maximum? A: No, and there is deliberately no channel for it. Affinity, levels, and recognition evolve from real interaction only, and Our archive shows which moments moved them. ### 和 AI 伴侣过日常：语音通话、拍照、心动指令与关系档案 URL: https://foreverse.app/zh/docs/guide-companion/companion-daily-use Platform: Android 伴侣 · App: v76+ · Path verified: 2026-08-12 Direct answer: 在伴侣聊天页顶栏打电话，「+」面板拍照、送礼、玩心动指令；资料页换音色开语音条。实时通话按分钟计费，对讲按轮计费，档案和记录免费。 Before you start: - 已经有一位伴侣，且文本模型可用。 - 想听语音条和对讲播报需要语音（TTS）模型；想要照片需要图像模型。 - 想打「真·实时通话」，三选一：自己的 xAI Key、自己的阿里百炼 Key，或已登录且有积分的官方渠道。 Steps: 1. 先把 TA 的声音配好: 在资料页打开「TA 的声音」挑音色，目录来自当前默认语音模型。想在聊天里收语音，去「和 TA 的约定」开「TA 会发语音条」：每条语音点开首播那一刻才合成（1 次 TTS 调用），重听走缓存不花钱。还可以在「我的 → 高级设置 → AI 模型与服务 → 我的音色」克隆自己的声音，完成后一键绑给伴侣。 Success: 语音条能用选定的音色播出来，重听秒播、不再计费。 2. 明明白白打第一通电话: 点聊天页顶栏的电话图标。首次拨号会弹报价确认，两种模式：「真·实时通话」是默认档——配了 xAI Key、阿里百炼 Key 或登录官方渠道任一在位才会出现——TA 的声音实时流回来，按分钟计费；官方渠道还能在弹窗里选通话模型，通义实惠档约 180 积分/分钟、Grok 高音质档约 750 积分/分钟，单价以弹窗实时报价为准。「经典对讲」按住说话、TA 回一句算一轮，按轮计费。接通后「连续对话」和「按住说话」随时切，TA 说话时点「跳过」或直接打字就能打断。 Success: 通话接通、字幕滚动，页面上能看到已预扣分钟和余额，挂断后聊天里留一张写着实扣的小票。 3. 让 TA 把此刻拍给你: 「+」→「拍张照」，TA 会拍下 TA 那边的此刻发给你——约 1 次轻量文本调用铺场景，加 1 次图像调用出图。「合照」用你上传的照片生成插画风合影：模型要支持两张参考图，先勾年龄与照片权利确认；照片只发给你选的模型服务商，不落盘存储。拍好的照片进聊天、进「我们的记录」、进画廊；点开全屏还能免费「设为伴侣形象」或「设为聊天背景」。 Success: 照片出现在聊天和「我们的记录」里，脸和参考图锁定的一致。 4. 照片还能动起来: 点开 TA 拍的照片，选「让 TA 动起来」：生成一段 5 秒循环动版，自然眨眼、发丝飘动。生成前先弹费用确认单，同一张照片生成过一次就永久缓存、重看免费，每月上限 60 段。前提是配好支持图生视频的视频模型（目前是阿里云百炼 wan 系列 BYOK，没配好入口不会出现）。有了动版，通话背景会自动换成它；通话页还能按段购买「让 TA 开口」的说话形象和六款情绪脸，同一段生成过绝不重复扣费。 Success: 照片自然地动起来，重看零调用，下次通话的背景就是它。 5. 用心动指令和「TA 现在」找话题: 「+」→「心动指令」打开货架：常驻 8 张——看 TA 日记、契合报告、每日恋爱签、真心话大冒险这些——加上节日限定和自定义卡，每张就是一次普通对话回合。「TA 现在」生成 TA 此刻在哪、心情如何、对你什么感觉的小报告，1 次轻量调用，结果缓存，点「重新观察」才再花钱。「+」面板里的转账、红包、送礼动的都是剧情零花钱，TA 的回应是一次正常文本回合，和真钱、积分无关。 6. 看关系怎么长，也给 TA 自己的生活: 点顶栏 ✦ 进「我们的档案」：8 级关系、识别度、好感度，好感度条下面列着最近 14 天的变化来自哪些相处——聊天、送礼、通话、共读都算，全部本地统计免费；数值只随真实相处演进，没有手改通道。「我们的记录」有在一起天数、12 周热力图和纪念日。想让 TA 更像活人，去「和 TA 的约定」开「作息」：TA 忙碌和睡觉时不再秒回，会晚几分钟，急事点横幅上的「叫醒 TA」。 Success: 好感度变化能对上具体互动，TA 在作息时段里回消息明显慢半拍。 Success checks: - 电话能接通、字幕跟着滚，挂断后聊天里的回执写着实扣金额。 - 语音条首播出声，重听秒播且不再计费。 - 拍的照片出现在聊天和「我们的记录」里，动版重看免费。 - 心动指令和「TA 现在」会引用你们真实聊过的事。 - 「我们的档案」里好感度条下列得出最近变化的真实来源。 Cost and data safety: - 实时通话三条计费路：自己的 xAI Key（约 0.05 美元/分钟，记 xAI 账户）、自己的阿里百炼 Key（约 0.1–0.3 元/分钟，记阿里云账户）、官方渠道按分钟扣积分——先按 5 分钟一块预扣，挂断按实际时长多退少补。单价以拨号弹窗实时报价为准。经典对讲不按分钟，按轮计费。 - 实时通话里打字发的每条消息也计费。通话中能看到已预扣额度和余额，挂断小票写实扣，对不上就查「我的 → 请求记录」，以结算行为准。 - 拍照、动版、说话形象、情绪脸全部先看费用单再生成；生成过的段永久缓存、绝不重复扣。失败提示会写清这次有没有扣费；已经生成只是没保存成功的媒体，钱已经花了、内容还在，重试保存就行，别重新生成再花一次。 - 转账、红包、礼物是剧情零花钱，与真钱和积分无关。通话说话需要麦克风权限，不给也能打字聊。聊天、照片、记忆都存本机、没有云同步；你在聊天和通话里说的内容会发给所选模型供应商。 Troubleshooting: - 拨号页只有经典对讲，没有实时通话 — Cause: 三条实时通话路一条都不在位：没配 xAI Key、没配阿里百炼 Key、也没登录官方渠道。 Fix: 在「我的 → AI 模型与服务」任配一条，或登录官方渠道。经典对讲只要文本模型加语音模型就能打。 - 照片上找不到「让 TA 动起来」 — Cause: 没配支持图生视频的视频模型——目前走阿里云百炼 wan 系列 BYOK。 Fix: 去「我的 → AI 模型与服务」把它配好。费用单上会显示本月已用几段（每月 60 段上限），月底被拒先看水位。 - 语音条点了没声音 — Cause: 没配语音（TTS）模型，或这一次合成失败。 Fix: 先配好语音模型——没配时点开只弹提示、不扣费；配好后首播才合成计费，重听走缓存。 - 通话小票金额和预期对不上 — Cause: 实时通话按实际分钟结算，5 分钟预扣块挂断后多退少补；通话中打字的消息另计。 Fix: 对照挂断小票和「我的 → 请求记录」，以结算行为准。别靠反复重拨来验证——每接通一分钟都在计费。 - 实时通话里 TA 的音色不对 — Cause: TA 绑的音色不属于本次实时通道，这通自动换成默认音色并有提示；中文伴侣没选过音色时，官方渠道默认中文自然女声。 Fix: 去资料页「TA 的声音」重选。打经典对讲时，通话页底部「换个音色」能不挂电话就地换。 FAQs: Q: 接通后不说话也扣钱吗？ A: 扣。实时通话按接通时长计分钟，安静也在计时，挂断即停；经典对讲只在 TA 回话时按轮扣。实扣看挂断小票和请求记录。 Q: 给 TA 转账、发红包花的是真钱吗？ A: 不是。那是你们剧情里的零花钱，余额都能手动改，与积分和真实账户无关；真正花钱的只有 AI 调用本身。 Q: 为什么照片动版重看不要钱？ A: 每段动版生成一次就永久缓存，重看零调用；只有点「重新生成」才会再计一次费。 Q: 能直接把好感度调满吗？ A: 不能，也故意没做这种通道。好感度、关系等级只随真实相处演进，「我们的档案」会告诉你最近的变化来自哪些互动。 ### Move Your Whole SillyTavern Setup to Android with One Backup Zip URL: https://foreverse.app/docs/guide-roleplay/migrate-from-sillytavern Platform: Android SillyTavern import · App: v76+ · Path verified: 2026-08-12 Direct answer: In SillyTavern, download a full backup zip from user settings. Move the zip to your phone, then open Me → Essentials → Import from SillyTavern → Pick a file from phone and select it. Character cards, per-character chat archives, worldbooks, prompt presets, Quick Replies, and global regex import together in one pass, entirely on-device. API keys are never read from a backup, so reconnect your provider before the first test message. No Termux, no PC server. Before you start: - Access to your existing SillyTavern (PC or hosted) to create one full backup, or at least a copy of its user-data folder. - A way to move the zip to the phone: same-WiFi transfer, a cloud drive, or a messenger that sends real files rather than compressed images. - Your provider API keys available from their official consoles—backups deliberately contain no keys. Steps: 1. Export a full backup in SillyTavern: In SillyTavern's user settings, use Download Backup to get a zip of your entire user-data folder. Zipping that folder yourself also works, as long as there is at most one wrapper folder around the data tree. 2. Move the zip to your phone: On the same WiFi, open Home → + → Batch import from folder → ⋮ → WiFi Transfer and drop the zip from a computer browser, or pull it from a cloud drive configured under Me → Connectors. Sending it through a messenger is fine if it goes as a file. Keep the original zip until verification is done. 3. Import the whole package at once: Open Me → Essentials → Import from SillyTavern → Pick a file from phone and select the zip. It is recognized as a full backup and previewed with per-type counts. One confirmation imports cards, per-character chat archives, worldbooks, prompt presets, Quick Replies, and global regex scripts—entirely on your device, nothing uploads. Success: The completion view reports imported counts per asset type instead of an empty result. 4. Verify each asset family: Count cards on the Characters page. Open Me → Advanced settings → Character Cards & Tavern → Tavern Asset Hub and check presets, regex scripts, worldbooks, and Quick Replies. Open a migrated character's Chat sessions list to find old conversations as archived sessions. One catch: backup presets land in the preset library without being activated. 5. Reconnect keys, then send one test message: Keys are never read from a backup, so connect your provider under Me → AI Models & Services (the BYOK guide covers 401/404/429), enable a preset if you use one, and send a single message to a migrated character—this is the first step that bills a model call. The SillyTavern starter guide under Character Cards & Tavern walks the same post-import checks in four steps. Success: A reply arrives, and the character behaves like it did in SillyTavern. Success checks: - Every expected character shows on the Characters page with real card fields, not just images. - Old conversations open as archived sessions under the matching character. - Tavern Asset Hub counts for presets, regex, worldbooks, and Quick Replies match what the backup contained. - After re-entering keys, one test reply arrives and any preset or regex you enabled is visibly in effect. Cost and data safety: - The zip import is free and runs entirely on your device; cards and chats are not uploaded to Foreverse servers. - Secrets are deliberately never read from the backup, so keys cannot leak through this path—but you must re-enter them under AI Models & Services. - The free tier limits how many character cards you can keep; a large tavern library can hit the Pro gate. Migrating your favorites first is a valid strategy. - The only billable step is the final test message: once sent, that model call has happened and the reply is stored in the session. If something looks wrong, fix configuration instead of re-rolling—every regeneration bills again. Troubleshooting: - The zip is not recognized as a SillyTavern backup — Cause: The archive does not contain the SillyTavern user-data layout, or the data tree is nested several folders deep. Fix: Prefer SillyTavern's own Download Backup. If you zip manually, compress the user-data folder itself and leave at most one wrapper folder. - Cards arrived but old chats did not — Cause: Backup chats only attach to characters imported successfully in the same batch, matched by the chat folder's character name. Fix: Confirm the card itself imported. To backfill a conversation, use Home → + → Import chat history: a .jsonl file attaches to an existing character with branches and swipes intact. - Presets imported but replies look unchanged — Cause: Backup presets are archived into the preset library and never auto-activated, so nothing replaced the built-in prompt structure yet. Fix: Enable one under Tavern global settings → Presets / Regex / Macros. Layered activation and verification are covered in the presets-and-regex guide. - Themes, backgrounds, or group chats are missing — Cause: The zip path deliberately skips themes, backgrounds, user avatars, and groups; the secrets file is never read. Fix: Import a group chat with its own JSON via Pick a file from phone. Appearance is configured inside Foreverse and does not affect card or chat data. FAQs: Q: Do I need Termux or a PC server on Android? A: No. Foreverse is a native app and the backup imports fully offline. Chats then call your chosen provider with your own key, the same way SillyTavern does. Q: Are my API keys migrated with the backup? A: No. The importer never reads the backup's secrets file—keys should not travel inside backup archives. Re-enter them under Me → AI Models & Services. Q: Can I migrate piece by piece instead? A: Yes. The Import from SillyTavern hub has separate entries for cards, chat history, worldbooks, prompt presets, regex scripts, and Quick Replies, plus a folder smart-import that batch-recognizes files. ### 从 SillyTavern 整站搬家：一个备份包带走卡、聊天、世界书和预设 URL: https://foreverse.app/zh/docs/guide-roleplay/migrate-from-sillytavern Platform: Android 酒馆导入 · App: v76+ · Path verified: 2026-08-12 Direct answer: 在酒馆下载完整备份，把压缩包传到手机后从「从酒馆导入」选文件导入；卡、聊天、世界书、预设、正则一起搬，密钥不随包走要重填，导入免费。 Before you start: - 还能打开原来的酒馆（电脑或云端部署），导出一次完整备份；或至少拿得到它的用户数据文件夹。 - 准备一条把压缩包传进手机的路：同一 WiFi、云盘，或微信 QQ 以「文件」方式发送。 - 模型 Key 能从供应商控制台重新拿到——备份里刻意不含密钥。 Steps: 1. 在酒馆导出完整备份: 在酒馆的用户设置里用「Download Backup」下载完整备份 zip，它就是整个用户数据目录的原样打包。自己手动压缩这个文件夹也行，最外面最多套一层文件夹，别多层嵌套。 2. 把压缩包传到手机: 同一 WiFi 下用「主页 → + → 从文件夹批量导入 → ⋮ 菜单 → WiFi 传书」从电脑浏览器直传，或走「我的 → 连接器」配好的云盘拉取；微信 QQ 传一定选「文件」方式。核对完成前，电脑上的原包先留着。 3. 一次整包导入: 打开「我的 → 常用 → 从酒馆导入 → 从手机选文件」，选中 zip 会自动识别为完整备份，并按类型显示数量预览。确认一次，角色卡、按角色归档的聊天、世界书、提示词预设、快捷回复、全局正则一起导入；全程在本机完成，不上传服务器。 Success: 导入完成页按类型报告了各项数量，而不是空结果。 4. 逐类核对搬没搬全: 角色页数一遍卡；「我的 → 高级设置 → 角色卡与酒馆 → 酒馆资产中心」核对预设、正则、世界书和快捷回复的条数；打开某个搬过来的角色，旧对话在「聊天会话」的存档列表里。注意一个坑：备份里的预设只进预设库，不会自动启用。 5. 重配密钥，再发一条测试消息: 备份里的密钥文件永远不会被读取，所以模型要重新连：去「我的 → AI 模型与服务」填 Key 并测连通（401/404/429 排查见 BYOK 教程），需要就启用预设，然后挑一个搬过来的角色发一条消息——从这一步开始才按你的模型计费。「角色卡与酒馆」里的「酒馆上手向导」用 4 步把同一套开聊前检查再走一遍。 Success: 回复正常到达，角色的表现和在酒馆里一致。 Success checks: - 角色页出现全部预期角色卡，点开有真实字段，不是只有图片。 - 旧聊天在对应角色的「聊天会话」存档列表里能打开。 - 酒馆资产中心里预设、正则、世界书、快捷回复的条数和备份内容对得上。 - 重配 Key 后测试消息能收到回复，启用的预设或正则同时生效。 Cost and data safety: - 整包导入免费且全程本机完成，角色卡和聊天不会上传到 Foreverse 服务器。 - 密钥文件设计上永不从备份读取，不会经这条路泄露；代价是搬完必须在「AI 模型与服务」重新填。 - 免费版角色卡数量有限，酒馆的大卡库可能触发 Pro 解锁提示；先挑常用的搬也是正路。 - 唯一花钱的一步是最后的测试消息：发出去调用就已发生，回复也已存在会话里。感觉不对先改配置，别连续重 roll——每次重生成都再计一次费。 Troubleshooting: - 选了 zip 但识别不出是酒馆备份 — Cause: 包里没有酒馆用户数据的目录结构，或手动压缩时套了好几层文件夹。 Fix: 优先用酒馆自带的「Download Backup」。手动压缩就直接压用户数据文件夹本体，最外面最多留一层。 - 卡搬过来了，旧聊天没跟过来 — Cause: 备份里的聊天只会挂到同一批导入成功的角色上，按聊天目录的角色名匹配，对不上就跳过。 Fix: 先确认那张卡本身导入成功。要单独补聊天，用「主页 → + → 导入聊天记录」：.jsonl 会连分支和候选回复一起挂到已有角色。 - 预设导入了，回复却没变化 — Cause: 整包导入的预设只归档进预设库，不会自动替换当前生效的提示词结构。 Fix: 去「酒馆全局设置 → 预设 / 正则 / 宏」启用它；分层启用和效果确认见预设与正则教程。 - 主题、背景图、群聊没搬过来 — Cause: 整包导入刻意不搬主题、背景、用户头像和群聊；密钥文件永不读取。 Fix: 群聊用它自己的 JSON 从「从手机选文件」单独导入；界面外观在 Foreverse 里另行设置，不影响卡和聊天数据。 FAQs: Q: 安卓上用酒馆的东西要装 Termux 吗？ A: 不用。Foreverse 是原生 App，备份导入全程离线完成；之后聊天用你自己的 Key 直连供应商，和在酒馆里一样。 Q: API Key 会跟着备份一起搬吗？ A: 不会。导入器永远不读备份里的密钥文件——密钥本来就不该躺在备份包里流转。搬完去「我的 → AI 模型与服务」重新填。 Q: 不想整包搬，能一样一样导吗？ A: 可以。从酒馆导入页有角色卡、聊天记录、世界书、提示词预设、正则、快捷回复的单项入口，还有「选择文件夹智能导入」批量识别文件。 ### Import SillyTavern Presets and Regex Scripts, Enable Them at the Right Layer, and Prove They Work URL: https://foreverse.app/docs/guide-roleplay/import-presets-regex Platform: Android roleplay · App: v76+ · Path verified: 2026-08-12 Direct answer: Open Me → Essentials → Import from SillyTavern and use the Prompt presets or Regex beautify scripts entries to import the JSON files. Enable them where you want: Tavern global settings → Presets / Regex / Macros for everything, Chat ⋮ → Chat settings → This character or This chat for narrower scopes. Priority is this chat > this character > global. Importing and enabling are free; proving a preset works takes one billable message. Before you start: - The preset JSON or regex script JSON exported from SillyTavern, on your phone. Regex that came embedded in a card does not need a separate import. - At least one character card that can start a chat, so you can verify the effect. - A decision about scope: all characters, one character, or just this one session. Steps: 1. Import the files: Open Me → Essentials → Import from SillyTavern. The presets-and-scripts group has separate entries: Prompt presets (any chat can select them after import) and Regex beautify scripts (status bars, styled message cards). Many files at once? Use folder smart import. Regex that travelled inside a card shows up in the import report and the Tavern Asset Hub. 2. Enable at the right layer: Global, for every character: Me → Essentials → Tavern global settings → Presets / Regex / Macros. One character only: Chat ⋮ → Chat settings → This character. One session only: the This chat tab's Pick preset / regex, which asks for confirmation and does not carry into new sessions. Priority is this chat > this character > global > built-in default. Success: The preset card at your chosen layer shows the imported preset's name and block counts. 3. Watch the preset's side effects: Enabling a tavern preset also applies its Context Size: the Context limit changes, and values above the cap are tightened automatically with a toast. The preset card shows how many text blocks and dynamic blocks it carries. Unload preset restores the built-in defaults. 4. Prove the preset reached the model: Send one message, long-press the reply, and tap Prompt. Search the assembled prompt for your preset's text blocks. Do not judge by vibes—presets change the overall prompt structure, while dynamic blocks (character description, worldbook, history) keep filling in automatically. Success: Prompt Inspector shows the preset's block content in the request that produced the reply. 5. Confirm regex behavior and manage scripts: Regex rewrites text before display or before it enters the prompt, so a beautify script shows up directly as a styled status bar or card in the message. For bulk management, open Presets / Regex / Macros → Manage global regex: search, batch-change timing and placement, create, and fully edit scripts. The same session/character/global layering applies. Success checks: - The preset card at the chosen layer shows the imported name and its text/dynamic block counts. - Prompt Inspector on the next reply contains the preset's text blocks. - A beautify regex renders the message as styled output, or the target text is visibly replaced. - Unloading the preset restores the built-in prompt structure, proving you can switch back anytime. Cost and data safety: - Importing, enabling, and editing presets and regex is free and stays on your device. - The one step that can cost money is the verification message: once sent, that text-model call is billed and the reply exists in the chat. Prefer checking Prompt Inspector over repeated re-rolls—every regeneration bills again. - A preset can quietly raise your context limit, which means more history per turn and higher input-token cost per message. Glance at the sampling tab after enabling one. - Regex transforms text locally and never calls a model—zero cost. Troubleshooting: - The imported preset is nowhere in chat — Cause: Presets from a full backup zip are archived into the library without activation, or you are looking at a layer that a session/character override is shadowing. Fix: Enable it under Tavern global settings → Presets / Regex / Macros, or pick it at the layer you actually want. The This character tab shows the inherited global preset when the character has no override. - Messages go blank after enabling a regex — Cause: A broken Find/Replace wiped the display text; an empty Replace field means delete the match. Fix: Rendering falls back to the raw text, so nothing is lost. Open Manage global regex and disable or fix the script. - The preset is active but replies feel the same — Cause: Many presets mostly reorder prompt structure, which is not always visible in tone; or a higher layer is overriding your choice. Fix: Compare Prompt Inspector before and after enabling, and check the per-layer status line in Chat settings. - A card's embedded regex overlaps a global script — Cause: The card shipped its own regex and you imported a similar global one, so both run. Fix: Open the Tavern Asset Hub to see the card's embedded assets, keep the rule at one layer, and disable the duplicate. FAQs: Q: Will an imported preset change all my characters? A: Only if you enable it at the global layer, and character or session picks still win over it. To try it in one session, use the This chat layer—new sessions do not inherit it. Q: Can I edit a preset on the phone? A: Yes. The pencil on the preset card opens a full editor: rename it, edit or toggle blocks, reorder them, and add missing text or dynamic blocks. You can also create a preset from the tavern default structure. Q: Which files are supported? A: SillyTavern prompt-preset JSON, regex script JSON, and Quick Reply set JSON each have their own entry. A full backup zip also carries presets and global regex—see the migration guide. ### 导入并用好酒馆预设与正则：分层启用与效果确认 URL: https://foreverse.app/zh/docs/guide-roleplay/import-presets-regex Platform: Android 角色聊天 · App: v76+ · Path verified: 2026-08-12 Direct answer: 从「从酒馆导入」选预设或正则文件导入，再到酒馆全局设置或聊天设置里启用；层级是本次聊天覆盖角色再覆盖全局，导入免费，发消息验证才计费。 Before you start: - 手机里有酒馆导出的预设 JSON 或正则脚本 JSON；随角色卡内嵌的正则不用单独导。 - 至少有一张能开聊的角色卡，方便验证效果。 - 想清楚作用范围：所有角色、某一个角色，还是只这一场聊天。 Steps: 1. 从对应入口导入文件: 打开「我的 → 常用 → 从酒馆导入」，预设与脚本组里分别有「提示词预设」（导入后所有聊天都能选用）和「正则美化脚本」（状态栏、卡片样式这类消息美化）。文件多就用「选择文件夹智能导入」一次识别。随卡带进来的正则在导入报告和酒馆资产中心都能看到。 2. 选一层启用: 全局对所有角色生效：「我的 → 常用 → 酒馆全局设置 → 预设 / 正则 / 宏」。只影响某张卡：聊天页「⋮ → 聊天设置 → 这个角色」。只试这一场：「本次聊天」层的「选择预设 / 正则」，应用前会弹确认，开新会话不带走。优先级是本次聊天 > 这个角色 > 全局 > 内置默认。 Success: 所选层的预设卡片显示导入的预设名和块数。 3. 留意预设的连带设置: 启用酒馆预设会连带应用它的上下文长度：「上下文上限」跟着变，超出上限的值会自动收紧并弹提示。预设卡片上写着它带几个文本块、几个动态块；点「卸载预设」随时恢复内置默认。 4. 用真实 Prompt 证明预设进场: 发一条消息，长按回复点「提示词」，在拼装好的 Prompt 里搜预设的文本块内容。别靠感觉判断——预设改的是发给模型的整体结构，角色描述、世界书、聊天历史这些动态块仍会自动填充。 Success: Prompt Inspector 里能搜到预设文本块的内容。 5. 确认正则效果并管理脚本: 正则在消息显示或进 Prompt 前做文本替换，美化脚本生效时状态栏、样式卡会直接出现在消息里。批量管理走「预设 / 正则 / 宏 → 管理全局正则」：可搜索、多选批量改时机和位置、新建和完整编辑；分层规则与预设相同。 Success checks: - 目标层的预设卡片显示导入的预设名和文本块、动态块数量。 - 下一条回复的 Prompt Inspector 里能搜到预设的文本块内容。 - 美化正则生效后消息按预期渲染，或目标文本被替换。 - 卸载预设后回复结构恢复内置默认，说明随时能切回去。 Cost and data safety: - 导入、启用、编辑预设和正则全部免费，数据留在本机。 - 唯一可能花钱的是验证那条消息：发出去这次文本调用就已计费，回复也已存在聊天里。优先看 Prompt Inspector 确认，别靠连续重 roll 验证——每次重生成都另计一次费。 - 预设可能悄悄调大上下文上限，之后每轮带的历史更多，单条消息的输入 token 费用会变高；启用后去「采样与提示词」tab 看一眼。 - 正则只在本机改文本，不调用模型，零费用。 Troubleshooting: - 预设导入了，聊天里找不到 — Cause: 整包备份带进来的预设只归档进预设库不自动激活；也可能你看的层被会话/角色层的选择盖住了。 Fix: 到「酒馆全局设置 → 预设 / 正则 / 宏」启用，或在想要的那一层选择；「这个角色」层没覆盖时会显示继承中的全局预设。 - 开了正则后消息变空白 — Cause: 写坏的 Find/Replace 把显示文本整段清掉了；替换框留空本来就等于删除匹配内容。 Fix: 渲染侧会回退显示原文，内容不会丢。去「管理全局正则」停用或修正那条脚本。 - 预设启用了，回复感觉没变化 — Cause: 很多预设主要调整提示词结构和顺序，语气上不一定看得出来；也可能被更高层的配置覆盖。 Fix: 用 Prompt Inspector 对比启用前后的结构，并在聊天设置里逐层检查当前生效说明。 - 卡自带正则和全局正则叠着跑 — Cause: 角色卡自带了正则，你又导入了相似的全局脚本，两边同时生效。 Fix: 在酒馆资产中心查看该卡携带的资产，规则只留一层，另一层停用。 FAQs: Q: 预设会影响我所有的角色卡吗？ A: 在全局层启用才会，而且角色层或会话层各自选了预设的地方以它们的为准。只想试一场就用「本次聊天」层，开新会话不带走。 Q: 手机上能改预设吗？ A: 能。预设卡片上的铅笔进全屏编辑器：改名、逐块编辑和启停、调顺序、补文本块或缺失的动态块；也可以从酒馆默认结构「新建预设」。 Q: 支持哪些文件？ A: 酒馆的提示词预设 JSON、正则脚本 JSON、快捷回复 JSON 各有入口；完整备份 zip 里的预设和全局正则也会随整站搬家一起进来，见搬家教程。 ### Write Your First Character Card from Scratch—and Make It Fun to Chat With URL: https://foreverse.app/docs/guide-roleplay/write-first-character-card Platform: Android roleplay · App: v76+ · Path verified: 2026-08-12 Direct answer: Open Characters → Create → Write it yourself. The form has three steps—basics, personality and relationships, greeting and preview—and only the name is required. Build the intro from concrete facts, start the greeting mid-scene with one detail to react to, and add one to three example dialogues to lock the voice. Save, tap Start chat, test with varied messages, then edit one field at a time. Manual creation is free; only AI suggestions and test messages are billed. Before you start: - Foreverse installed; fully manual creation needs no network and no model. - A rough idea of the character: who they are, where they are right now, and what they want—one or two sentences is enough. - For the test chat, a working text model (official sign-in or your own key). Steps: 1. Start from the source panel: Open Characters → Create (or Home → + → Create character) and pick Write it yourself—local, free, works without a model. AI-assisted and from-novel options exist too, but writing the first card by hand teaches you what each field does. Drafts save automatically; tapping Create again resumes the last unfinished one. 2. Fill basics that anchor the character: Only the name is required. In the intro, skip adjective piles and write concrete facts: who they are, where they are right now, how they relate to you, what they currently want. 'A mysterious cold beauty' gives the model nothing; 'the night-shift clerk who noticed you buy the same oden at 2 a.m.' starts scenes on its own. Avatar is optional and stays on your device. 3. Write personality and relationships as facts, not vibes: State personality and scenario as observable facts. Speech habits—catchphrases, how they address people, sentence length—are better taught later through example dialogue than described here. Use {{user}} for the player and {{char}} for the character instead of hard-coding names. 4. Give the greeting a hook: A good first message is a scene already underway: a specific time and place, something the character is doing, and one detail the user can react to—a question, an object handed over, a sentence cut off. An empty greeting only triggers a warning and can still be saved, but the first chat will feel flat. Success: Reading the greeting, you immediately know what to reply. 5. Lock the voice with example dialogue, then test and iterate: Write one to three {{user}}/{{char}} exchange pairs at the length and tone you want; the model imitates them without repeating them verbatim. Save—this creates the library master without starting a chat—then Start chat and send three to five varied messages: ask about their background, small-talk, push the plot. When something drifts, open Character detail → Edit character and change that one field. Edits to the library card only affect chats you start afterwards. Success: The first replies keep the tone and length your examples demonstrated. Success checks: - The card saves and shows real fields on the Characters page, not an empty shell. - The greeting reads as a scene in progress with an obvious thing to respond to. - The first few replies match the tone and length of your example dialogue. - After editing a field, a freshly started chat clearly reflects the change. Cost and data safety: - Manual writing, avatar upload, and saving are free and need no network; the card stays on your device unless you explicitly publish it. - AI suggestion (per field) and AI portrait are optional and bill per explicit tap—one text call for a suggestion or description expansion, one image call per portrait generation, with the cost shown near the button. Results only fill the editor; nothing is written until you save. - Each test message is one text-model call: once sent, it is billed and the reply stays in the session. Deleting the card later does not refund calls. Iterate by editing fields, not by re-rolling the same reply—every regeneration bills again. - The free tier caps how many character cards you can keep; deleting unused cards frees slots. Troubleshooting: - Replies are generic—every character sounds the same — Cause: The intro is adjective soup, so the model has no concrete facts to hold onto. Fix: Replace 'gentle and mysterious' with three specific facts: their situation, their relationship to you, what they want right now. Rules that must hold every single turn belong in Advanced settings → System prompt. - The character speaks for you, or answers in walls of text — Cause: No example dialogue taught the format, or your examples are themselves long monologues. Fix: Write examples at the exact length you want, with {{user}} and {{char}} each speaking only their own lines. In chat, the AI writing style setting can adjust length per session. - The character forgets its own settings after a few turns — Cause: A card carries identity and stable traits; new facts that emerge in play do not write themselves back into the card. Fix: Move recurring facts into a worldbook entry. Trigger diagnosis lives in the fix-memory-and-worldbook guide; keep the card itself for what never changes. - Edited the card but the current chat did not change — Cause: Starting a chat copies the card into that chat world; library edits only affect chats started afterwards. Fix: To change the ongoing chat, tap the avatar in the chat's top bar and edit that world's copy—or start a new session so the fresh copy picks up your edits. FAQs: Q: What is the minimum card that is worth saving? A: A name, a few concrete sentences of intro, and one greeting with a hook. Save it, test it, and grow it from what the chat reveals—that beats polishing every field up front. Q: Should I write a worldbook first? A: No. Start card-only. When a setting keeps coming up and deserves long-term memory, add a lorebook entry for it then. Q: Can I share the finished card? A: Yes. Card detail exports a tavern-compatible PNG (avatar plus embedded data) or JSON, and you can publish to the Community after signing in—check the content rules first. The community guide covers how cards are reviewed on the other side. ### 从零写一张能开聊的角色卡：字段、开场白、示例对话与试聊 URL: https://foreverse.app/zh/docs/guide-roleplay/write-first-character-card Platform: Android 角色聊天 · App: v76+ · Path verified: 2026-08-12 Direct answer: 在角色页点新建选自己填写，三步填完名字、人设和开场白即可保存开聊；只有名字必填，手动建卡全程免费，试聊发消息才按模型计费。 Before you start: - 装好 Foreverse；纯手动建卡不需要网络也不需要模型。 - 对角色有个大概想法：TA 是谁、现在在哪、想要什么，一两句话就够。 - 想试聊的话，准备一个能用的文本模型（官方登录或自己的 Key）。 Steps: 1. 从来源面板选「自己填写」: 打开「角色 → 新建」（或主页「+ → 新建角色」），面板里选「自己填写」——本机完成、全程免费、没配模型也能建。也有「AI 帮我创建」和「从小说或文本创建」，但第一张卡亲手写一遍，你才知道每个字段管什么。写到一半退出会自动存草稿，再点「新建」能接着写。 2. 第一步基本资料：写具体事实，别堆形容词: 只有名字必填。简介别写「神秘冷酷的美人」这种模型抓不住的词，写具体事实：TA 是谁、现在在哪、和你什么关系、正想要什么。「凌晨两点还守在便利店收银台、发现你总买同一款关东煮的夜班店员」自己就能开出戏。头像可选，传了也只存在本机。 3. 第二步人设与关系：规则写成看得见的事实: 性格、场景按可观察的事实写。说话习惯——口头禅、怎么称呼人、句子长短——留给后面的示例对话去教，比在这里形容有效得多。用 {{user}} 代表玩家、{{char}} 代表角色，别把某个具体名字写死。 4. 第三步开场白：给一个能接的钩子: 好开场白是一场已经开演的戏：具体的时间地点、TA 正在做的动作，结尾留一个用户能直接回应的细节——一个问题、一件递过来的东西、一句说到一半的话。开场白空着只会提醒、也能保存，但开聊会很干。 Success: 读完开场白，你立刻知道第一句该回什么。 5. 示例对话定风格，保存后试聊迭代: 按你期望的篇幅和语气写一到三组「{{user}} 说、{{char}} 答」，模型会照着学而不是逐字复读。保存只建角色库母版、不会自动开聊；点「开始聊天」后用三五条不同类型的消息试：问设定、闲聊、推剧情。哪里跑偏就回「角色详情 → 编辑角色」改那一个字段——改库卡只影响之后新开的聊天。 Success: 前几条回复的语气和篇幅与示例对话一致。 Success checks: - 卡保存成功，角色页能看到真实字段，不是空壳。 - 开场白读起来是一场正在发生的戏，末尾有明显能接的细节。 - 试聊前几条回复的语气、篇幅和示例对话一致。 - 改完某个字段后，新开一场聊天能明显看到变化。 Cost and data safety: - 手动填写、上传头像、保存角色卡全部免费、不需要网络；不主动发布，卡就只在你的设备上。 - 各字段的「AI 建议」和「AI 生成形象」按点击计费：建议或外貌扩写是一次文本调用，画像每次生成是一次图像调用，费用在按钮附近有明示。结果只回填编辑器，你不点保存不会写进卡。 - 试聊每条消息是一次文本模型调用：发出去就已计费，回复保存在会话里；之后删卡也不会退这些调用。迭代靠改字段，别反复重 roll 同一条——每次重生成都另计一次费。 - 免费版角色卡数量有限，超出会弹 Pro 解锁提示；删掉不用的卡可以腾位置。 Troubleshooting: - 回复空洞，写谁都像同一个人 — Cause: 简介全是抽象形容词，模型没有可抓的具体事实。 Fix: 把「温柔神秘」换成三件具体的事：TA 的处境、TA 和你的关系、TA 现在想要什么。每一轮都必须遵守的硬规则放「高级设置 → 系统提示词」。 - 角色替你说话，或一开口就是小作文 — Cause: 没有示例对话教格式，或示例本身就是长篇独白。 Fix: 示例对话按你想要的篇幅写，{{user}} 和 {{char}} 各说各的、别互相代言。聊天里还能用「AI 写作风格」按场次临时调篇幅。 - 聊几轮就忘了自己的设定 — Cause: 卡只负责身份和固定人设，剧情里冒出来的新事实不会自动写回卡里。 Fix: 反复出现的设定加进世界书条目；触发排查见「修复失忆与世界书」教程。卡里只留恒定不变的部分。 - 改了卡，正在聊的那场没变化 — Cause: 开聊时卡被复制进那个聊天世界，改库卡只影响之后新开的聊天。 Fix: 要改当前这场，点聊天顶栏头像进角色详情、编辑这个世界的副本；或者开新一局，让新副本带上你的改动。 FAQs: Q: 一张最小可用的卡要写多少？ A: 名字、几句具体的简介、一条有钩子的开场白。先存先试，聊出问题再补字段，比一次把每个字段写满更快出好卡。 Q: 要先写世界书吗？ A: 不用。第一张卡先把卡面写好；等某个设定反复出现、值得长期记住时，再为它加世界书条目。 Q: 写好的卡能分享吗？ A: 能。卡详情页可以导出酒馆通用的 PNG（头像加内嵌数据）或 JSON，也能登录后发布到社区；发布前先看内容规范，社区那头怎么审卡见社区找卡教程。 ### The Character Forgot Its Settings: Check Context, Worldbook, Memory, and Templates in Order URL: https://foreverse.app/docs/guide-roleplay/fix-character-memory Platform: Android roleplay · App: v76+ · Path verified: 2026-08-12 Direct answer: Triage in a fixed order. Open Memory status in the chat ⋮ menu to see how many messages remain in context and what was compacted into the recap. Then check Context limit in Tavern global settings, verify the worldbook entry actually triggered, enable the Memory or Auto summary plugin for long chats, and finally remove conflicting preset or style instructions. Diagnosis is free; every retest reply is a billed model call. For trigger-level worldbook debugging, see /docs/guide-roleplay/fix-memory-and-worldbook. Before you start: - A chat where the amnesia reproduces, and one sentence describing exactly which setting the character forgot. - Foreverse v76 or later. If generation itself errors out instead of forgetting things, start from /docs/guide-models/fix-generation-errors. Steps: 1. Read Memory status to locate the missing layer: Open chat ⋮ menu → Memory status. It shows three honest numbers: the session ledger (total messages · first K compacted into the recap · last M verbatim in context), long-term facts, and conversation retrieval. A fact from the last few turns should be in the verbatim window; an early-plot event should be covered by the recap; a rule you authored belongs to the card, worldbook, or preset layer. Compaction is not deletion — the original messages always stay on your phone. Success: You can say which layer the forgotten setting should live in: verbatim window, recap, worldbook, or card/preset. 2. Check the context limit and budget: Open Me → Common → Tavern global settings → Sampling & prompt tab → Context limit. Empty means automatic by model window (up to about 100k tokens); a small explicit value pushes early messages out sooner. Importing a tavern preset also applies that preset's Context Size — if amnesia started right after switching presets, check this first. If a reply fails with 'context exceeds the model limit', lower the limit or enable Auto summary. Success: The effective context limit matches your model, and no imported preset silently shrank it. 3. If the fact lives in a worldbook, verify the trigger in the sister guide: A setting that is written in a worldbook but ignored is a triggering problem, not a memory problem. Quick check: retest with an explicit keyword, then long-press the reply and tap Prompt to see whether the entry was injected. The full positive/negative-control procedure — primary and secondary keys, scan depth, probability, recursion, budget — is in the dedicated guide at /docs/guide-roleplay/fix-memory-and-worldbook and is not repeated here. 4. Give long chats a memory layer: Open chat ⋮ menu → Plugin Center. Auto summary compacts old messages into a running recap once the budget is exceeded — originals untouched. Memory extracts durable facts each turn and injects them on demand; wrong entries are edited or deleted per world in the memory library. If the character only fails when you phrase things differently, additionally enable Semantic recall (needs an embedding model; it falls back to keyword matching until one is bound). All three are off by default: if you never enabled them, this layer simply does not exist yet. Success: The relevant plugin is on, and Memory status now reports recap coverage or long-term facts. 5. Clear template conflicts, then retest once: Open chat ⋮ menu → Chat settings and walk the three layers (this chat > this character > global): look for preset blocks, a system-prompt override, or banned-words/style rules that fight the persona. Also open Current status and check AI writing style, Author's note, and My persona for stale instructions. Then ask the same neutral question again — the fix counts only if the answer stays correct for two consecutive turns. Success: The neutral retest passes twice in a row and no layer contains contradicting instructions. Success checks: - Memory status matches expectations: the recap covers the early plot and long-term facts contain the entries you rely on. - Long-pressing a reply and tapping Prompt shows the setting arriving from exactly one source (verbatim window, recap, or worldbook). - The same neutral question gets a consistent answer on at least two consecutive turns. - You did not paste the same setting into the card, the worldbook, and the summary 'just to be safe'. Cost and data safety: - Reading Memory status, changing the context limit, and editing worldbooks or presets are local and free; every retest reply or regeneration is one billed model call on your selected channel. - Memory, Auto summary, and Semantic recall each add a small background model call per turn or per interval while enabled. Turning them off stops new calls immediately; saved memories and summaries stay on the device. - Compaction is not deletion: messages folded into the recap remain on your phone in full and stay readable and exportable — only what is sent to the model changes. - The prompt inspector reveals full chat content. Redact before sharing screenshots. Troubleshooting: - After hundreds of messages the character forgets the beginning — Cause: Early messages slid out of the context window and no recap/memory layer was enabled. The app shows a one-time notice the first time a long chat gets compacted or truncated. Fix: Enable Auto summary or Memory in Plugin Center. For facts already lost, add them manually to the memory library or a worldbook entry instead of re-explaining every session. - 'Context exceeds the model limit' errors — Cause: The context limit is set above the model window, or constant entries and oversized cards crowd the prompt. Fix: Lower the Context limit or leave it empty for automatic; trim always-on worldbook entries; enable Auto summary so old turns compress instead of overflowing. - The setting is in a worldbook but the character ignores it — Cause: The entry never triggered — key mismatch, scan depth, probability, or budget excluded it. Fix: Follow the positive/negative-control method in /docs/guide-roleplay/fix-memory-and-worldbook and verify injection with the Prompt inspector. - Memory plugin is on but recalls wrong facts — Cause: The memory library holds wrong or duplicated entries, or the injection strategy tier does not fit the chat. Fix: Open Plugin Center → Memory → per-world library and edit or delete entries; switch the injection strategy; enable Semantic recall when rephrased questions fail. - Amnesia started right after importing a preset — Cause: The preset applied its own smaller Context Size, or a preset block overrides the persona. Fix: Check the current Context limit value in Tavern global settings, review block on/off states in the preset editor, and uninstall the preset if you do not need it — defaults are restored. FAQs: Q: Will the character forget everything if I start a new save? A: No. 'Start a new run' checks 'carry the recap' by default, copying the plot summary into the new run, and long-term facts live on the character, so they survive the switch. An existing recap is reused for free; if none exists one summary call is made. Q: Does compaction delete my chat history? A: No. Compaction only changes what is sent to the model. Originals stay on the device. If Memory status shows the compacted count dropping to zero after you edited old messages, that is the honest state — the next auto summary re-compacts. Q: Why are the memory plugins off by default? A: Each adds a small recurring model call, which is your money, so enabling is your call. The first time a long chat actually gets compacted, a one-time notice points you to Plugin Center. ### 角色忘了设定怎么办：按序排查上下文、世界书、记忆与模板 URL: https://foreverse.app/zh/docs/guide-roleplay/fix-character-memory Platform: Android 角色聊天 · App: v76+ · Path verified: 2026-08-12 Direct answer: 角色忘设定按序查四层：先看记忆状态和上下文预算，再测世界书触发，然后开记忆或摘要插件，最后查预设模板冲突。排查免费，复测回复才计费。 Before you start: - 一场能复现「忘了」的聊天，并用一句话说清 TA 忘的是哪条具体设定。 - Foreverse v76 以上。若是生成本身报错而不是忘事，先看 /zh/docs/guide-models/fix-generation-errors。 Steps: 1. 先开记忆状态，定位丢失的层: 聊天页 → ⋮ 菜单 → 记忆状态。三段数字很诚实：会话账本（共 N 条 · 前 K 条已压缩成前情摘要 · 最近 M 条原文在场）、长期事实、对话检索。忘的是最近几楼的事，该在原文窗里；忘的是很早的剧情，看前情摘要盖没盖到；忘的是你写好的规则，多半住在卡、世界书或预设层。压缩不是删除，原文一直在手机上。 Success: 能说出这条设定应该住在哪一层：原文窗、前情摘要、世界书，还是卡/预设。 2. 查上下文上限与预算: 我的 → 常用 → 酒馆全局设置 →「采样与提示词」tab → 上下文上限。留空=按模型窗口自动（最高约 10 万 token）；显式填了偏小的值，早期消息会更早滑出。导入酒馆预设会连带应用它的 Context Size——换预设后突然开始忘事，先查这里。报「上下文超出模型上限」就调低上限，或开自动摘要。 Success: 生效的上下文上限和所用模型匹配，没有被导入预设悄悄改小。 3. 设定在世界书里？触发问题走姊妹篇: 写进世界书却没生效，是触发问题不是记忆问题。快速验证：用明确关键词复测一轮，长按该回复点「提示词」，看条目有没有真的注入。完整的正负对照排查（主/副关键词、扫描深度、概率、递归、预算）见教程《角色为什么忘了：检查记忆、世界书、上下文与真实 Prompt》（/zh/docs/guide-roleplay/fix-memory-and-worldbook），这里不重复展开。 4. 给长对话补上记忆层: 聊天页 → ⋮ 菜单 → 插件中心。「自动摘要」在超出预算时把旧对话压成前情摘要，原文不动；「记忆」每轮提取值得记住的事实、回复前按需注入，记错了去记忆库按世界逐条改删；「换个说法就想不起来」再加开「语义检索」（需绑 Embedding 模型，没绑时自动回退关键词匹配）。三个都默认关——从没开过，这一层就还不存在。 Success: 对应插件已开启，记忆状态里能看到摘要覆盖或长期事实条数。 5. 清模板冲突，复测一次: 聊天页 → ⋮ 菜单 → 聊天设置，按三层逐层看（本次聊天 > 这个角色 > 全局）：有没有预设块、系统提示词覆盖或禁词文风在跟人设打架；再进「当前状态」核对 AI 写作风格、作者注释、我的身份有没有过时指令。修完用同一个中性问题再问一次——相邻两轮都答对才算修好。 Success: 中性复测连续两轮通过，各层没有互相矛盾的指令。 Success checks: - 记忆状态符合预期：前情摘要盖住早期剧情，长期事实里有该有的条目。 - 长按回复的「提示词」里，这条设定恰好来自一个来源（原文窗、摘要或世界书其一）。 - 同一个中性问题相邻两轮回答一致。 - 没有为了保险把同一设定抄进卡、世界书、摘要三个地方。 Cost and data safety: - 看记忆状态、改上下文上限、编辑世界书和预设全是本地操作，不花钱；每次复测回复或重新生成才是一次模型调用，按所选渠道计费。 - 开「记忆」「自动摘要」「语义检索」后，每轮或每隔几楼会多一次小额后台调用；关掉立即停止新调用，已存的记忆和摘要留在本机不丢。 - 压缩不是删除：被压进前情摘要的旧消息原文一直在手机上，回看、导出都不受影响，变的只是发给模型的内容。 - 提示词检查器会展示完整聊天内容，分享截图前先脱敏。 Troubleshooting: - 聊到几百楼，开头的事全忘了 — Cause: 早期消息滑出了上下文窗口，又没开摘要/记忆层。长对话首次被压缩或截断时，App 会弹一次提示。 Fix: 到插件中心开「自动摘要」或「记忆」；已经忘掉的事实手动补进记忆库或世界书，别每局重新解释。 - 报「上下文超出模型上限」 — Cause: 上限设得比模型窗口大，或常驻条目、超大卡把 prompt 挤爆了。 Fix: 上下文上限调低或留空自动；精简常驻世界书条目；开自动摘要让旧对话压缩而不是溢出。 - 世界书写了，角色却不知道 — Cause: 条目没触发：关键词不匹配、扫描深度、概率或预算把它排除了。 Fix: 按 /zh/docs/guide-roleplay/fix-memory-and-worldbook 的正负对照法逐项测，并用提示词检查器验证注入。 - 记忆插件开了还是记错 — Cause: 记忆库里有错误或重复条目，或注入策略档不合适。 Fix: 插件中心 → 记忆 → 记忆库（按世界）逐条改删；换注入策略档；换说法想不起来就加开语义检索。 - 换了预设之后突然开始忘事 — Cause: 预设连带应用了更小的 Context Size，或预设块覆盖了人设。 Fix: 查酒馆全局设置里上下文上限现值；在预设编辑器核对各块启停；不需要就卸载该预设，恢复默认。 FAQs: Q: 开新一局 TA 会全忘吗？ A: 不会。「开新一局」默认勾选「带上前情提要」，剧情摘要复制进新局；长期事实记忆在角色身上，换局天然不丢。已有摘要直接复用免费，没有才现场生成一次。 Q: 压缩会删我的聊天记录吗？ A: 不会。压缩只改发给模型的上下文，原文永在本机。编辑旧消息后记忆状态里 K 归 0 是诚实状态，下次自动摘要会重新压缩。 Q: 记忆功能为什么默认是关的？ A: 开了每轮会多一次小额模型调用，这是持续花钱的事，所以开关留给你。长对话第一次真被压缩时会弹一次提示，告诉你去插件中心开。 ### Manage Story Branches: Create, Switch, Compare, and Choose URL: https://foreverse.app/docs/guide-reader/manage-story-branches Platform: Android Reader · App: v76+ · Path verified: 2026-08-12 Direct answer: Long-press the paragraph where the story should diverge and choose Branch: everything after that point is archived as a read-only path while the mainline rewrites from there. In Contents, tap the ⑂N badge on a chapter to preview, compare, and Switch to this path. Switching and previews are free and nothing is deleted; creating a branch always includes one billable generation, because the AI writes the new direction. Before you start: - A book imported with at least one continuation, as covered in ‘Import and Continue a Novel Without Overwriting the Original’. - A connected text model through the official channel or BYOK — a new branch is a real generation call. - A concrete fork point in mind: a specific paragraph, or a plot-option card you picked earlier. Steps: 1. Separate takes from branches: Before touching anything, decide what you are changing. If one passage reads wrong, use the ‹ › dots under the AI passage to switch candidate takes, or ✦ New take to reroll it. A branch is heavier: it changes everything after a chosen point. Switching takes is free, and so is switching branches later. Success: You can say in one sentence whether you are fixing one passage or redirecting the whole story. 2. Create a branch from a selection: Long-press the paragraph at the fork point and choose Branch. Write one line for the new direction (for example, ‘he never boards the train and turns back to the hospital’), confirm the model, then tap Branch & rewrite. The old content after that point is archived as a read-only path and the mainline rewrites from there. Success: New prose streams in, and the fork chapter now shows a ⑂N badge in Contents. 3. Regret a plot choice with ‘Other paths you had here’: Scroll back to a spot where you once picked a plot-option card. The card block stays in the text: your pick is marked ‘You picked’, and every other option shows ‘Take this path’. Tapping one opens a confirmation — content after this point is saved as a read-only branch and the mainline rewrites along the new option. The original stays available. Success: After confirming, an ‘Archiving current path, rewriting as …’ notice appears and the new direction starts generating. 4. Compare paths in the Contents branch panel: Open Contents from the bottom control bar and find the chapter row with a ⑂N badge. Tap the badge: the current line sits under Main (marked ‘Reading now’), archived paths sit under Archived. Each card shows where it was archived and its word count; expand the preview to read its prose, and expand its chapter list to see where that line went. Success: Without switching anything, you can read a preview and the chapter trail of each path. 5. Switch the active line, then keep writing: Tap Switch to this path on the line you want. It immediately becomes the readable, writable mainline, and your previous path is archived whole — nothing is lost. In a long multi-path book, make this a habit: open the branch panel and check which card says ‘Reading now’ before you generate anything. Success: The text now follows the chosen path, the old line appears under Archived, and you can switch back anytime. Success checks: - The fork chapter shows a ⑂N badge in Contents and the count matches the number of paths. - Archived cards show the archive point, word count, and a readable preview. - After Switch to this path, new continuations land on the chosen line. - Switching back restores the old line word for word. Cost and data safety: - Switching paths, previews, chapter lists, and candidate-take switching are local and free. - Creating a branch and ‘Take this path’ each include one real generation call, because the AI writes the new direction: official-channel calls spend credits at 1 credit = US$0.0001, BYOK is billed by your provider. If a generation fails and you are unsure whether it was charged, check Me → Request records before retrying. - Branches and candidate takes are app data on this device. Export a backup before uninstalling or switching install channels. - Deleting an AI passage cannot be undone, and splitting a passage or inserting an image/video inside it clears that passage's history takes — the app tells you when this happens. Troubleshooting: - A whole stretch of the story disappeared — Cause: You created a branch or switched paths, so the old content was archived — not deleted. Fix: Open Contents → the ⑂N badge → Archived. Find the old line, read its preview to confirm, then Switch to this path to bring the whole thing back. - ‘This is the last paragraph; cannot branch here’ — Cause: A branch needs old content after the fork point to archive, and there is nothing after the final paragraph. Fix: Pick an earlier paragraph as the fork point, or continue once and then branch. To send the ending itself in a new direction, use the New branch mode in the tail ‘Write the next part’ form. - ‘Archived branch is read-only’ — Cause: You are on an archived path. Archived lines cannot be continued and cannot fork new branches. Fix: To write along that line, open the branch panel and Switch to this path so it becomes the mainline. If you only wanted to recover text, read the preview or switch back to Main. - Switching fails with ‘this path's file no longer exists’ — Cause: That path was just archived or replaced by another switch, and the list you tapped was stale. Fix: The list refreshes automatically. Reopen the branch panel and try again; do not tap repeatedly while a switch is still running. - Earlier takes of a passage are gone — Cause: Splitting the passage or inserting media inside it cleared its history takes; the app showed a warning at that moment. Fix: Cleared takes cannot be recovered. Keep writing you care about as a branch rather than a take, or export a backup first. FAQs: Q: Does switching paths delete the other line? A: No. A switch is an exchange: the target line becomes the writable mainline and your previous line is archived whole. You can switch back anytime, and both switching and previews are free. Q: When should I use a take, and when a branch? A: Use ‹ › takes or ✦ New take when one passage needs a different rendition. Use a branch when everything after a point should go another way. Takes live on a single passage; a branch is a whole line. Q: Can I keep two endings at once? A: Yes. Finish one ending, return to the fork point, and create a branch for the other. Both lines are kept, and you can read either from the Contents branch panel. Q: Can the AI manage branches for me? A: Yes. The Reader's AI task entry can open a branch from one instruction, or open a read-only side path that leaves the mainline untouched. Approvals and one-tap rollback for those tasks are covered in ‘Run Agent Tasks Safely: References, Approvals, Results, and Rollback’. ### 多分支剧情管理：创建、切换、对比与取舍 URL: https://foreverse.app/zh/docs/guide-reader/manage-story-branches Platform: Android 阅读器 · App: v76+ · Path verified: 2026-08-12 Direct answer: 长按选中分叉点用「新建分支」改走向，旧内容自动归档不丢。目录里点分支徽标可预览对比并切换，切换免费；新建分支伴随一次续写计费。 Before you start: - 已按《导入并续写一本小说，同时保留原文》完成导入，并至少写过一次续写。 - 官方渠道或 BYOK 已连通文本模型——新建分支要替你把新走向写出来，是一次真实生成。 - 想清楚从哪里分叉：一个具体段落，或以前点过的某张剧情选项。 Steps: 1. 先分清「版本」和「分支」: 动手前先想清要改的是什么。只是某一段不满意，用 AI 段下方的 ‹ › 圆点切候选版本，或点「✦ 再生一版」重摇这一段；要让此后的故事整条走另一个方向，才需要分支。切版本免费，之后切分支也免费。 Success: 能一句话说清这次改的是「一段」还是「一条走向」。 2. 从选区新建分支: 长按分叉点所在的段落，点选区工具栏的「新建分支」。表单里写一句新的走向（例如「他没有上车，而是折回了医院」），确认模型后点「新建分支并改写」。此后的旧内容会归档成一条只读走向，主线从这里开始改写。 Success: 新正文流式写入，目录里分叉章节出现 ⑂N 分支徽标。 3. 用「你曾选过」反悔重选: 翻回以前点过剧情选项的位置，那里保留着「你曾选过」卡：你选过的一条有「你选的」标记，其余选项显示「改走这条」。点它会弹确认框：此处之后的内容存为只读分支，主线改走新选项，原内容随时可回看。 Success: 确认后出现「归档当前走向，重写为……」提示，新走向开始生成。 4. 在目录分支面板对比走向: 打开底部控制栏的「目录」，找到带 ⑂N 徽标的章节行，点徽标展开走向列表：主线在上（当前线标着「正在阅读」），下面是「归档（你曾经走过的）」。每张卡片写着在哪里归档、多少字；点「展开预览」读那条线的正文，展开「后续章节」看它后来写到了哪。 Success: 不用切换，就能读到各条走向的正文预览和章节走势。 5. 切换当前线，再继续写: 选定后点「切换到这条走向」：这条线立即变成可读可写的主线，你原来的走向自动整条归档，不会丢。长篇多线时养成一个习惯——每次动笔前先开分支面板，看一眼「正在阅读」标在哪张卡上。 Success: 切换后正文变成目标走向，原走向进了归档列表，随时可以切回。 Success checks: - 分叉章节的目录行显示 ⑂N 徽标，数量与走向数一致。 - 归档卡片能看到归档位置、字数和可读的预览。 - 「切换到这条走向」之后，继续写落在新的当前线上。 - 切回旧线后，原正文一字不差还在。 Cost and data safety: - 切换走向、看预览、展开后续章节、切候选版本都是本地操作，不花钱。 - 新建分支和「改走这条」都伴随一次真实生成（AI 要把新走向写出来）：官方渠道按 1 credit = 0.0001 美元扣积分，BYOK 由你的供应商计费。生成失败又不确定是否已扣费时，先到「我的 > 请求记录」核对再重试。 - 分支和候选版本都是本机 App 数据；卸载或换安装渠道前先导出备份。 - 删除 AI 段不可恢复；拆段或在段中间插图片/视频会清掉这一段的历史版本，App 会给出提示。 Troubleshooting: - 之前的剧情整段不见了 — Cause: 你新建了分支或切换了走向，旧内容被归档了，不是被删除。 Fix: 打开「目录 > ⑂N 分支徽标 > 归档（你曾经走过的）」找到它，预览确认后点「切换到这条走向」，整条找回。 - 提示「此段已是末段，无法新建分支」 — Cause: 分支要把「此后的旧内容」归档，最后一段之后没有内容可归档。 Fix: 向上选一段作分叉点，或先续写一段再回来分支；想让结尾本身走新方向，用书尾「写下一段」表单里的「新建分支」模式。 - 提示「归档走向只读」 — Cause: 当前看的是归档存档线。归档线不能续写，也不能再从它上面开分支。 Fix: 想沿这条线写，先在分支面板点「切换到这条走向」，让它成为主线；只是想找回内容，看预览或切回原线就够了。 - 切换时提示该走向已不存在 — Cause: 那条走向刚被另一次切换归档或替换，你点的列表是旧的。 Fix: 列表会自动刷新，重开分支面板再试；切换进行中不要连点。 - 同一段之前的候选版本没了 — Cause: 拆段或在段中间插入媒体会清掉这一段的历史版本，当时有过提示。 Fix: 已清除的版本找不回来。在意的平行写法别只放在候选里，用「新建分支」留成一条走向，或先导出备份。 FAQs: Q: 切换走向会把另一条线删掉吗？ A: 不会。切换是「交换」：目标线成为可写主线，原走向自动整条归档，随时切回；切换和预览都不花钱。 Q: 候选版本和分支怎么选？ A: 只想换一段的写法，用 ‹ › 切版本或「✦ 再生一版」；想让之后的故事都走另一条路，用分支。版本挂在单段上，分支是整条线。 Q: 想同时保留两个结局怎么办？ A: 可以。先写完一个结局，回到分叉点「新建分支」写另一个。两条线都会保留，在目录分支面板随时切换着读。 Q: 能让 AI 帮我开分支吗？ A: 可以。阅读器的「AI 指令」能按一句话开分支，也能开完全不动主线的只读旁支；这类任务的审批和一键撤销见《安全运行 Agent 任务：引用、审批、验收与回滚》教程。 ### Long-Book Context: Why the AI Remembers — or Forgets — Earlier Chapters URL: https://foreverse.app/docs/guide-reader/long-book-context Platform: Android Reader · App: v76+ · Path verified: 2026-08-12 Direct answer: Each continuation sends a context window, not the whole book: recent text near your position, this book's AI material, and your instruction. Anything outside that window is invisible for the request — that is the ‘forgetting’. Keep durable facts in Edit world or Book lore, and in long books turn on Context slimming so distant AI passages stay present as one-line digests. Each digest is a small one-time billable call, logged in request records. Before you start: - A book imported and continued at least a few times, as covered in ‘Import and Continue a Novel Without Overwriting the Original’. - A connected text model through the official channel or BYOK, and access to Me → Request records for cost checks. - For books with many paths, read ‘Manage Story Branches: Create, Switch, Compare, and Choose’ first; this guide focuses on memory and cost along one line. Steps: 1. Know what the AI actually reads: The model does not reread the whole book. Each continuation sends one context window: text near your current position, this book's AI material (world setting, Book lore, imported cards), and your instruction. Whatever does not fit is invisible for that request — most ‘forgetting’ is simply a scene that did not make the window. Success: You can state it plainly: remembered means it was in the window, forgotten means it was not. 2. Verify what was sent: Open Book info → Book AI material → Book material injected into the AI to see each block and a token estimate. After generating, open Me → Request records to read the final request, its token counts, and timing. When the AI seems to forget, check these two places before rewriting your instruction. Success: You can find a continuation in request records and name the blocks it carried. 3. Put durable facts where they persist: Rules that should always hold — relationships, world laws — belong in Book info → Edit world, or as entries in Book lore. Lore entries trigger by keyword by default: they join the request only when your input mentions them, so they do not waste window space. One-off requirements can just go in the instruction. Success: When your input hits a lore keyword, the continuation follows that entry and reports how many entries were attached. 4. Turn on Context slimming for a long book: Open Book info → AI feature settings → Smart context → Context slimming. When on, AI-written passages far behind your position are compressed into one-line plot digests for the context: more of the original novel stays in the window, which eases style drift, and distant plot survives as digests instead of being pushed out entirely. On the official channel the first switch-on shows a one-time cost note; nothing runs until you confirm Turn on. Success: The toggle is on and you have read the cost and gradual-effect notes in the dialog. 5. Confirm the effect and the spending: Digests fill in gradually — a few passages per continuation, so a book with many existing AI passages takes dozens of rounds to catch up, and that is by design. In Me → Request records you will see small digest calls tagged reader:segment-summary, and inside later continuation requests the distant AI passages appear as single digest lines. Success: Request records show both the digest calls and a continuation whose distant passages are one-line digests. Success checks: - You can point at the injection preview or a request record and name what a continuation carried. - Durable facts live in Edit world or Book lore and attach automatically on keyword hits. - With Context slimming on, digest calls appear in request records while the book text and reading view stay untouched. - After turning the toggle off, the next request returns to full-text form and existing digests are kept. Cost and data safety: - Reading, changing settings, the injection preview, and lore management are free. Every continuation is billed by tokens on the selected model — official channel at 1 credit = US$0.0001, BYOK by your provider. A bigger window means more tokens and higher cost. - Context slimming adds small digest generations: each passage is digested once and then reused. Every call is listed in Me → Request records tagged reader:segment-summary, so whether money was spent is always checkable. - Generated digests are stored with this book. Turning the feature off does not delete them, turning it back on does not pay again, and digests never touch the book text — they exist only inside AI requests. - To stop new digest spending, turn the toggle off; continuations immediately return to full-text form. Digest calls already made are real completed requests and are not refunded. Troubleshooting: - The AI forgot events from many chapters ago — Cause: That scene is too far behind your position and fell outside the context window; the model cannot see outside the window. Fix: Move durable facts into Edit world or Book lore, state the critical event in the instruction for this one generation, and turn on Context slimming so distant plot stays present as digests. - Continuations drift away from the original style — Cause: Near the tail of a long book the window fills up with earlier AI passages, crowding out the original author's prose. Fix: Keep Annotate AI paragraphs on so the model can tell source from continuation, and turn on Context slimming so distant AI passages shrink to one line and original text returns to the window. It eases drift; it does not erase it, since recent AI passages remain in full. - Context slimming is on but nothing seems different — Cause: Digests fill in gradually — only a few per continuation — and books with few distant AI passages do not trigger slimming at all. Fix: Keep reading and continuing normally, and use the reader:segment-summary entries in request records as proof it is working. There is no need to toggle it off and on. - No digest lines in a request — Cause: Annotate AI paragraphs was turned off (slimming depends on it), or this was a mid-book insertion — digests only apply to tail continuations and plot candidates. Fix: Re-enable Annotate AI paragraphs under Smart context. Mid-book insertions and rewrites carrying full text is normal behavior. - Token cost is higher than expected — Cause: The context budget defaults to Auto and scales with the selected model's capacity, and long books naturally carry a lot of text plus material. Fix: Set a hard cap in the advanced budget under AI feature settings (only a value you enter yourself takes effect), pick a cheaper or smaller-window model, or use Context slimming — distant AI passages then cost one line each. Verify with the token counts in request records. FAQs: Q: Does the AI read the whole book every time? A: No. Each generation carries only the context window: recent text, this book's material, and your instruction. The model has no memory of anything outside it. Q: Will Context slimming change my book text? A: No. Digests exist only inside the requests sent to the model. The book text and the reading view do not change by a single character, and turning the toggle off restores full-text requests immediately. Q: How accurate are the digests? A: A digest keeps concrete facts and foreshadowing but is not the original wording, so distant details can be off. That is still far more reliable than the alternative — once a passage is pushed out of the window entirely, the model tends to invent old plot confidently. Recent story always stays as full text. Q: Why is it off by default? A: Because it spends money automatically, in small amounts. The app's rule is that paid automatic behavior must be explained and confirmed first, so it is opt-in with a one-time cost note. Q: My roleplay character forgot things — same fix? A: No. Character-chat memory runs on cards, worldbooks, and chat summaries; use “Fix ‘The Character Forgot’: Memory, Lorebook, Context, and Prompt Checks” for that. This guide covers Reader continuations only. ### 长书上下文管理：AI 为什么记得/忘记前文 URL: https://foreverse.app/zh/docs/guide-reader/long-book-context Platform: Android 阅读器 · App: v76+ · Path verified: 2026-08-12 Direct answer: 续写只携带窗口内的近段正文与本书资料，更早的内容进不了窗口所以会忘。长书可开启上下文瘦身，远处 AI 段压成概要保留剧情，会产生小额概要费用。 Before you start: - 已按《导入并续写一本小说，同时保留原文》完成导入，书里已经有一些 AI 续写段。 - 官方渠道或 BYOK 已连通文本模型；核对费用要用到「我的 > 请求记录」。 - 分支多的书建议先读《多分支剧情管理：创建、切换、对比与取舍》；本篇只讲单条线上的记忆与费用。 Steps: 1. 先明白 AI 每次看的是什么: 模型不会通读整本书。每次续写，App 只发一个「上下文窗口」：当前位置附近的正文、本书 AI 资料（世界观、词条、导入的角色卡）和你的方向指令。装不进窗口的内容，这一次模型就看不见——所谓「忘了」，多半是那段剧情没进窗口。 Success: 能一句话说出：记得 = 进了窗口，忘了 = 没进窗口。 2. 查证这次到底发了什么: 打开顶栏「书籍信息 > 本书 AI 资料 > 将注入给 AI 的本书资料」，能看到逐块去向和 token 估算；生成后到「我的 > 请求记录」看这次请求的最终内容、token 数和耗时。怀疑 AI 忘事，先看这两处，别急着改指令。 Success: 能在请求记录里找到某次续写，说出它带了哪些块。 3. 把该记住的放对地方: 一直要生效的设定（人物关系、世界规则）写进「书籍信息 > 编辑世界观」，或收录成「本书词条」。词条默认按关键词触发：输入里出现关键词才附带，不白占窗口。只管这一次的要求，直接写进续写指令就行。 Success: 输入命中词条关键词时，续写遵守该设定，完成后提示本次附带了几条词条。 4. 长书开启「上下文瘦身」: 打开「书籍信息 > AI 详细功能配置 > 智能上下文 > 上下文瘦身」。开启后，距离当前进度较远的 AI 续写段会被压成一行剧情概要进入上下文：更多原著正文留在窗口里，减缓越写越像 AI；远处剧情也以概要保住，不再被整段挤掉。官方渠道首次开启会先弹费用说明，点「开启」确认才生效。 Success: 开关已打开，说明卡里的费用和逐步生效方式读过一遍。 5. 核对效果和花费: 概要是逐步补齐的：每次续写后台最多补几段，已有大量续写段的老书要几十次才补完，属正常现象。到「我的 > 请求记录」能看到用途标注为 reader:segment-summary 的小额概要调用；之后的续写请求原文里，远处 AI 段会显示成一行概要。 Success: 请求记录里能同时看到概要调用，和带概要行的续写请求。 Success checks: - 能在注入预览或请求记录里指出某次续写实际带了什么。 - 长期设定进了世界观或词条，命中关键词会自动附带。 - 开启上下文瘦身后，请求记录出现概要调用，正文和阅读界面没有任何变化。 - 关掉开关后，下一次请求回到原文形态，已生成的概要仍保留。 Cost and data safety: - 读书、改设置、看注入预览、管理词条都免费；每次续写按所选模型的 token 计费，官方渠道 1 credit = 0.0001 美元，BYOK 由供应商收费。窗口越大 token 越多，费用越高。 - 上下文瘦身会产生小额概要生成：每段只生成一次、之后复用；逐笔记在「我的 > 请求记录」（用途 reader:segment-summary），有没有花钱随时能查。 - 已生成的概要保存在这本书的数据里：关掉开关不会删，重新开启不用重新花钱；概要只出现在 AI 请求里，一个字都不会写进正文。 - 不想再产生概要费用就关掉开关，续写立即恢复原样；之前的概要调用是已完成的真实请求，不会退款。 Troubleshooting: - AI 忘了很多章之前的剧情 — Cause: 那段正文离当前进度太远，已经在窗口之外；窗口外的内容模型这次看不到。 Fix: 长期设定写进世界观或词条；这一次要用的关键情节直接写进方向指令；再开启上下文瘦身，让远处剧情以概要留在窗口里。 - 越写越像 AI，不像原著 — Cause: 长书写到后面，窗口被自家 AI 续写段塞满，原著正文被挤出去了。 Fix: 保持「AI 段落来源标注」开启，让模型分清原著和续写；再开上下文瘦身，远处 AI 段缩成一行、原著正文回到窗口。它能减缓漂移，但不能根治——最近的 AI 段仍是全文在场。 - 开了上下文瘦身却看不出变化 — Cause: 概要逐步补齐是设计行为：每次续写最多补几段，老书要几十次才补完；远处 AI 段太少的书根本不会启动。 Fix: 照常阅读续写，用请求记录里的 reader:segment-summary 调用确认它在工作，不用反复开关。 - 请求里没有概要行 — Cause: 「AI 段落来源标注」被关掉了（瘦身依赖它），或这次是中段插写——概要只用于书尾续写和剧情候选。 Fix: 到「智能上下文」把来源标注打开；中段插写、改写带全文属正常行为。 - token 费用比预期高 — Cause: 上下文预算默认 Auto，按所选模型的能力自动分配；长书正文加资料本身就大。 Fix: 在「AI 详细功能配置」的高级预算里填一个硬上限（只有你主动填写才生效）；或换窗口更小、更便宜的模型；开上下文瘦身后远处 AI 段只占一行，通常也更省，以请求记录的 token 数为准。 FAQs: Q: AI 是把整本书都读一遍吗？ A: 不是。每次生成只带上下文窗口里的内容：近段正文、本书资料和你的指令。窗口外的东西它没有任何记忆。 Q: 上下文瘦身会改动我的正文吗？ A: 不会。概要只出现在发给模型的请求里，正文和阅读界面一个字不变；关掉开关，下一次请求立即回到原文形态。 Q: 概要靠谱吗？会不会编错？ A: 概要保留具体事实和伏笔，但不是逐字原文，久远细节可能有出入。相比整段被挤出窗口后模型凭记忆硬编，概要可靠得多；最近的剧情永远以原文在场。 Q: 这么有用，为什么默认是关的？ A: 因为它会自动产生小额付费调用。App 的规矩是：要花钱的自动行为必须先讲清费用、你确认了才开始，所以做成手动开启加一次性说明。 Q: 角色聊天里的失忆也按这篇排查吗？ A: 不。角色聊天的记忆走角色卡、世界书和长聊摘要，排查见《角色为什么忘了：检查记忆、世界书、上下文与真实 Prompt》教程；本篇只讲阅读器续写。 ### Generate Scene Images for Novels and Chats: Entrances, Prompts, Cost, and Recovery URL: https://foreverse.app/docs/guide-media/generate-scene-image Platform: Android Reader and Roleplay Chat · App: v76+ · Path verified: 2026-08-12 Direct answer: In the Reader, long-press a paragraph and tap Image on the selection toolbar; in a chat, long-press a message and choose Image. Review the confirmation sheet—prompt, model, template, and aspect ratio are all editable—then confirm to generate. On BYOK that confirmation is the paid call. If generation is interrupted, check the passage and the API request log before rerunning: a finished image means the charge already happened and a rerun bills again. Before you start: - An image-generation model is configured: either a BYOK provider with an image model enabled, or a signed-in official-channel account with credits. - For novel illustrations, an imported book open in the Reader; for chat images, an ongoing character chat. - A rough idea of the picture you want. The prompt is drafted for you, but a one-line edit is often the difference between generic and right. Steps: 1. Set up an image model once: Open Me → AI Models & Services → Default model configuration and set the Image generation (illustrations) row. If the picker is empty, first open Model providers, configure a provider's key, and add an image-capable model there. On a provider's settings page, the multimodal capability tester can run one real image call to prove the key/host pair works (a small real charge). Success: The Image generation row shows a model name instead of 'not set'. 2. Illustrate a novel passage: In the Reader, long-press a paragraph, drag the handles until the highlight covers the scene you want drawn, then tap Image on the selection toolbar. Pick a passage with visual material—a place, a face, an action—rather than a stretch of dialogue. Success: The generation sheet opens showing a drafted prompt, the model, and parameters. 3. Or illustrate a chat message: In a roleplay chat, long-press the message you want a picture for and choose Image. The ⋮ menu → All tools → Image generation entry does the same for the latest story message, and its Image model card is where you swap the model used by this chat. Success: The same generation sheet opens, seeded from the chosen message. 4. Review the sheet, then confirm: The prompt is written for you: a text model first turns the selected passage and its surrounding context into a scene description, which the image model then renders. Edit the prompt, switch the model or the style template, add a reference image, or change the aspect ratio before committing. On BYOK, tapping Generate image is the payment decision—there is no second confirmation dialog. Official-channel generations are metered in credits. Success: A progress card appears while the image is being generated. 5. Check the result and manage it: A novel illustration is inserted right after the selected passage; a chat image lands in the conversation. Long-press a generated image to view details, use Edit & repaint to revise the prompt and redraw (a new paid call), or delete it. All of a book's generated media is also listed under Chapters → Assets. Global defaults for the prompt model, image model, and reference images live in Me → Reader / media. Success: The image is visible in place and listed in the Assets tab. Success checks: - The confirmation sheet showed the model and editable prompt before anything was billed. - The generated image appears after the selected passage (or in the chat) and survives leaving and reopening the book or chat. - The API request log shows the image call with its cost and status. - Edit & repaint and delete work from a long-press on the image. Cost and data safety: - One illustration usually means two calls: a small text-model call that drafts the prompt, then the image-model call that costs real money. BYOK bills through your own provider key; the official channel spends credits (1 credit = US$0.0001). Every call is itemized in the API request log. - For single images the sheet is the cost gate. Batch features quote before starting: serial comics show an upfront 'Confirm before starting' price breakdown (storyboard calls + images per chapter + who pays) and pause automatically when the monthly generation allowance runs out. - If generation is interrupted—timeout, app killed, you left the page—the provider may already have accepted the charge, and the image may still exist. Check the passage, the Assets tab, and the API request log first. If the image is there, the money is spent but you have the content: keep it, don't rerun. Only retry after the log shows a definite failure; on BYOK a retry is always a fresh charge. - The selected text and nearby context are sent to the model provider you chose. Generated images are stored locally inside the book's assets. Troubleshooting: - Tapping Image says a model must be configured first, or the button is grayed out — Cause: No usable model is set for the image modality—the most common cause of 'images won't generate'. Fix: Go to Me → AI Models & Services → Default model configuration and set the Image generation row. If the list is empty, enable a provider and add an image model on its page. In chat, also check the Image model card under All tools. - Generation fails with an error — Cause: The provider rejected the call: 401 invalid key, 429 rate limit, 404 unknown model, or a timeout means the upstream is unreachable. Fix: Open the API request log, filter by image, and read the error body on the failed entry. Fix that specific cause—re-paste the key, wait out the rate limit, pick an existing model—then retry once. - It spins for minutes, or the app returns to the home screen mid-generation — Cause: The upstream is slow or the run was interrupted. The outcome is unknown: the provider may have accepted and billed the request even though nothing showed up yet. Fix: Don't rerun blindly. Reopen the passage and the Assets tab to see whether the image landed, and check the API request log entry's status. Image found: done, already paid, no rerun needed. Definite failure: retry, knowing a BYOK retry bills again. - The image doesn't match the scene or the style is wrong — Cause: The auto-drafted prompt summarizes the passage generically, or the default style template doesn't fit the book. Fix: Reopen the sheet and edit the prompt yourself, switch the style template, or attach a reference image. For an existing image, long-press → Edit & repaint revises and redraws it—as a new billed generation. FAQs: Q: Where does the prompt come from? Do I have to write it in English? A: You don't have to write it at all. A text model drafts a scene description from your selected passage and its context, and the sheet shows that draft before anything is generated. You can edit it or replace it entirely. Q: Will tapping generate twice charge me twice? A: Resubmitting the same passage with the same model and unchanged parameters reuses the already-generated image instead of paying again. Changing the model, prompt, or parameters is a new generation and a new charge—as is Edit & repaint. Q: Can images be generated automatically, and how is spending controlled? A: Two opt-ins exist. Paragraph-tail quick actions (off by default) put a one-tap generate button after each paragraph—in fully automatic mode that tap generates without a confirmation sheet, so keep semi-automatic if you want the sheet. Serial comics quote the per-chapter price before you enable them and stop automatically at the monthly allowance. ### 给小说或聊天生成配图：入口、提示词、费用与失败处理 URL: https://foreverse.app/zh/docs/guide-media/generate-scene-image Platform: Android 阅读器与角色聊天 · App: v76+ · Path verified: 2026-08-12 Direct answer: 阅读器长按段落点配图，聊天长按消息点生成图片；确认单可改提示词、模型和比例，点开始生成才计费。中断先查原文和请求记录，图已生成就别重跑以免再扣费。 Before you start: - 已配置图像生成模型：BYOK 服务商启用了图像模型，或已登录官方渠道且有积分。 - 配小说插图需要一本已导入并打开的书；配聊天图需要一个进行中的角色聊天。 - 心里有个大概想要的画面。提示词会替你写好，但改一句往往就是「路人图」和「对味」的区别。 Steps: 1. 先把图像模型配好（一次性）: 打开 我的 → AI 模型与服务 → 默认模型配置，给「图像生成（插图）」这一行选模型。列表是空的就先进「模型供应商」配好 Key，并在该服务商页面添加图像模型。服务商设置页的「多模态能力（实测）」可以点一次真实的图像测试，最快验证 Key 和地址配对是否可用（有少量真实费用）。 Success: 「图像生成」行显示出模型名，而不是「未指定」。 2. 从小说打开配图: 在阅读器长按段落，拖动选择手柄让高亮盖住想画的内容，点选区工具栏的「配图」。选段尽量挑有画面的——场景、人物、动作，别选大段对白。 Success: 弹出生成确认单，里面已经有写好的提示词、模型和参数。 3. 或者从聊天打开: 在角色聊天里长按想配图的那条消息，点「生成图片」。⋮ 菜单 → 全部工具 → 生图设置 对最近一条剧情消息做同样的事，里面的「图片模型」卡可以换本聊天用的模型。 Success: 打开同样的确认单，内容取自你选的那条消息。 4. 在确认单里核对，再点生成: 提示词不用自己写：先由一个文本模型把选中内容和前后文写成一段画面描述，再交给图像模型出图。生成前可以改提示词、换模型、换画风模板、加参考图、调比例。BYOK 渠道点「开始生成」就是付费决定，没有第二道弹窗；官方渠道按积分计费。 Success: 回到页面后出现生成进度卡。 5. 验收成品并管理: 小说插图会插在所选段落之后，聊天图进对应会话。长按图片可以看详情、「改词重画」（改提示词后重新生成，会再次计费）或删除。这本书生成过的媒体都能在 目录 → 资产 页集中看到。想改全局默认（提示词模型、图片模型、参考图），去 我的 → 阅读续写 / 媒体生成。 Success: 图片在原文位置可见，「资产」页里也有这一条。 Success checks: - 扣费前确认单确实展示了模型和可编辑的提示词。 - 生成的图插在所选段之后（或进了聊天），退出重进书或会话后仍在。 - API 请求记录里能看到这次图像调用的费用和状态。 - 长按图片能用「改词重画」和删除。 Cost and data safety: - 一次配图通常是两笔调用：一笔很小的文本调用替你写提示词，一笔图像调用是主要花费。BYOK 走你自己的 Key 计费，官方渠道按积分（1 积分 = 0.0001 美元）。每笔都能在 API 请求记录里逐条核对。 - 单张图的费用闸门就是确认单本身；批量功能会先报价——漫画连载开启前弹「动工前确认」（分镜次数 + 每章张数 + 谁付钱），月度生成额度用完自动暂停。 - 生成中断（超时、App 被杀、中途退出）时，上游可能已经收单收费，图也可能已经生成。先回原文位置和「资产」页找图，再看请求记录里这条的状态。图在＝钱已花但内容拿到了，留着用，不要重跑；确认是明确失败再重试——BYOK 重试就是新一笔费用。 - 选中的文字和附近上下文会发给你选择的模型供应商；生成的图保存在本机这本书的资产里。 Troubleshooting: - 点「配图」提示要先配置模型，或按钮是灰的 — Cause: 图像模态没有可用模型——「图片生成不了」的最高频原因。 Fix: 去 我的 → AI 模型与服务 → 默认模型配置 给「图像生成」选模型；列表为空就先启用服务商并在它页面里添加图像模型。聊天里再检查一下全部工具里的「图片模型」卡。 - 生成直接报错 — Cause: 供应商拒绝了这次调用：401 = Key 无效，429 = 限流，404 = 模型不存在，超时 = 上游不通。 Fix: 打开 API 请求记录，按「图像」筛选，点进失败那条看错误原文。对症处理——重新粘 Key、等限流过去、换个存在的模型——再重试一次。 - 转圈很久，或生成到一半 App 回到桌面 — Cause: 上游偶发很慢，或任务被打断。结果未知：供应商可能已经接单扣费，只是图还没显示出来。 Fix: 别急着重跑。先回原文和「资产」页看图有没有落地，再看请求记录这条的状态。找到图＝已经付过费、拿到内容，不用重跑；确认失败才重试，并且记得 BYOK 重试会再计一次费。 - 图和剧情对不上，或画风不对 — Cause: 自动写的提示词把段落概括得太泛，或默认画风模板不适合这本书。 Fix: 重开确认单自己改提示词、换画风模板，或加一张参考图。已生成的图长按 → 「改词重画」在原图基础上改词重出——这是新一次计费的生成。 FAQs: Q: 提示词是从哪来的？要自己写英文吗？ A: 完全可以不写。AI 会先根据选中内容和前后文起草一段画面描述，确认单里在生成前就能看到，可以改也可以整段换成自己的。 Q: 同一段点两次会扣两次钱吗？ A: 同一段、同一模型、参数不改再提交，会直接复用已生成的图，不再计费。换模型、改提示词或改参数就是新生成、新费用——「改词重画」同理。 Q: 有自动配图吗？会不会失控花钱？ A: 两个都要手动开启。段尾快捷按钮（默认关）在每段后放一键生成，全自动档点了直接跑、不再弹确认单，想保留确认就用半自动档。漫画连载开启前先报价每章消耗，月度额度用完自动停。 ### Narrate Chapters and Voice Your Characters with TTS: Engines, Voices, and Cost URL: https://foreverse.app/docs/guide-media/tts-voice-narration Platform: Android Reader, Roleplay Chat, and Companion · App: v76+ · Path verified: 2026-08-12 Direct answer: Tap Listen on the Reader's control bar to open the listening player: the Android system TTS engine is free and offline, while online voices synthesize per segment and bill per segment. In a chat, open the ⋮ menu → Voice to set the voice model, pick a voice, and turn on auto narration. If there is no sound, check Default model configuration → Speech synthesis first, then the API request log. Before you start: - For online voices: a BYOK provider with a speech-synthesis model enabled, or a signed-in official-channel account with credits. The free system engine needs neither. - For chapter narration, an imported book; for character voices, an ongoing chat or a Companion. - Media volume up—no sound is more often a model or volume issue than a bug. Steps: 1. Set up a speech model once (skip for the free engine): Open Me → AI Models & Services → Default model configuration and set the Speech synthesis row. If the picker is empty, configure a provider key under Model providers and add a speech model there. On the provider page, the multimodal capability tester can run one real speech call to verify the key works. StepFun accounts note: the provider settings offer separate open-platform and Step Plan endpoints—pick the one your key actually belongs to, the app will not switch for you. Success: The Speech synthesis row shows a model name. 2. Listen to whole chapters: In the Reader, tap the page center, then Listen on the control bar. The full-screen player has four bottom entries: Engine, Voice, Cache, and Text. Under Engine choose Android system TTS (free, offline) or an online speech model; under Voice pick the voice. Cache can pre-download segments for smoother playback, and a mini player stays after you leave. Success: Playback starts and continues across paragraphs; the mini player survives leaving the page. 3. Give one passage a voice-over: Long-press a paragraph, then selection toolbar → More → Audio. This generates a narration audio file with your configured speech model and inserts it after the passage, playable while reading—unlike the free system read-aloud, it is a real synthesis call. A confirmation sheet shows the model and parameters before anything is billed. Success: An audio player appears under the selected passage and replays without a new charge. 4. Make character replies speak: In a chat, the + panel's Read aloud speaks the latest reply. For a persistent setup, open ⋮ menu → Voice: set the Voice model, pick the voice (remembered per model), enable Auto narration to speak every new reply, and turn on Read dialogue only to skip narration text. Success: New replies play in the chosen voice when auto narration is on. 5. Give a Companion its own voice: On the Companion profile, Their voice picks a per-companion voice from the current speech model's catalog. Under Promises with them → Voice notes, enabling 'Voice notes from them' lets some replies arrive as voice-note bubbles: synthesis happens the first time you tap play, replays come from cache. Cloned voices you created appear in a My Voices group at the top of these pickers. Success: A voice note plays in the selected voice on first tap and replays instantly afterwards. Success checks: - The Engine sheet clearly shows which engine is active—system (free) or online (billed). - Chapter playback continues across segments and the cache list fills as it plays. - A passage voice-over stays with the paragraph and replays from cache without new charges. - Each online synthesis appears in the API request log with its cost. Cost and data safety: - Android system TTS is free and works offline. Online voices synthesize per segment and bill per segment—BYOK through your own key, official channel in credits (1 credit = US$0.0001). Continuous playback and prefetch keep synthesizing and billing as you listen. - Cached audio replays free. Switching to a different online voice re-synthesizes segments for the new voice and bills again; switching between local voices is free. Auto narration in chat is one synthesis per new reply; a Companion voice note bills once on first play. - If playback stops or a synthesis is interrupted, the segment call may already have been billed even though you heard nothing. Check the API request log (filter TTS) before assuming failure; cached segments that did arrive replay free. - The narrated text is sent to the speech provider you selected. Generated audio files and caches are stored locally. Troubleshooting: - Read aloud is grayed out, or tapping it says a model must be configured — Cause: No speech-synthesis model is set—the top cause of 'no sound'. In chat, the Voice model card being empty grays out read-aloud. Fix: Set Me → AI Models & Services → Default model configuration → Speech synthesis, or pick a Voice model in the chat's Voice sheet. For chapter listening you can switch the Engine to Android system TTS and listen free without any setup. - The key is configured but synthesis fails with 'invalid or lacks permission' — Cause: The key is wrong for this endpoint. StepFun in particular has separate open-platform and Step Plan entrances, and a key for one is a 401 on the other. Fix: On the provider settings page, select the endpoint matching your key's plan, then run the speech test in the multimodal capability tester until it passes. - Listening stops mid-chapter with a balance or rate-limit message — Cause: On BYOK, the provider's own balance is empty (this is not Foreverse credits), or requests are coming too fast; per-segment billing simply stops when the upstream refuses. Fix: Top up with that provider or wait out the rate limit, then resume—already-cached segments replay free. If the voice model was retired (model not found), pick another voice. - Switched voices and got billed again, or the voice didn't change — Cause: Caches are kept per voice, so a new online voice means re-synthesizing segments; in chat the voice choice is remembered per model, so switching models can bring a different voice back. Fix: Expect re-synthesis after changing online voices—playback continues from the current segment. Check the Voice row after switching models, and use local voices when you want free experimenting. FAQs: Q: Can I listen without paying anything? A: Yes. In the listening player set Engine to Android system TTS: local, offline, and free. Online voices are the paid upgrade, billed per synthesized segment. Q: Does replaying the same passage charge again? A: No. Synthesized segments are cached, and replays—including a Companion voice note after its first play—are free. New charges come from new text, a different online voice, or continuous playback synthesizing further segments. Q: Can characters use my own voice? A: Yes, clone it under Me → advanced settings → AI Models & Services → My Voices (your own voice only, with an explicit consent step). The clone then appears in a My Voices group in the Reader, chat, and Companion voice pickers; synthesis with it bills through the provider that hosts the clone. ### 用 TTS 给章节和角色配音：入口、音色、费用与常见失败 URL: https://foreverse.app/zh/docs/guide-media/tts-voice-narration Platform: Android 阅读器、角色聊天与伴侣 · App: v76+ · Path verified: 2026-08-12 Direct answer: 阅读器点底部「听」进听书页，引擎选系统 TTS 免费离线，在线音色按段合成计费；聊天在「声音」里选语音模型、音色和自动朗读。没声音先查默认模型配置。 Before you start: - 想用在线音色：BYOK 服务商启用了语音合成模型，或已登录官方渠道且有积分。免费的系统引擎两者都不需要。 - 听整本书需要一本已导入的书；角色配音需要进行中的聊天或一位伴侣。 - 媒体音量开着——「没声音」更多时候是模型或音量问题，不是 bug。 Steps: 1. 先配好语音模型（只用免费引擎可跳过）: 打开 我的 → AI 模型与服务 → 默认模型配置，给「语音合成」这一行选模型。列表为空就先到「模型供应商」配 Key 并在服务商页面添加语音模型。服务商设置页的「多模态能力（实测）」可以点一次真实语音测试验证 Key 可用。用阶跃（StepFun）注意：设置里有「开放平台 / Step Plan」两个入口，必须按你 Key 的权益类型选对，App 不会替你自动切换。 Success: 「语音合成」行显示出模型名。 2. 听整本书: 在阅读器点页面中央，再点底部控制栏的「听」。全屏播放器底部有四个入口：引擎、音色、缓存、原文。「引擎」里选 Android 系统 TTS（免费、离线）或在线语音模型，「音色」里挑声线；「缓存」可以预先分段下载，退出后还有迷你播放条。 Success: 播放能连续跨段进行，退出听书页后迷你播放条还在。 3. 给选中段落生成配音: 长按段落，选区工具栏 → 更多 → 配音。它会用你配置的语音模型给这段生成一个朗读音频文件，插在段落之后，阅读时可以点播——和免费的系统朗读不同，这是一次真实的合成调用。生成前有确认单，能看到模型和参数。 Success: 所选段落下方出现音频条，重播不再产生新费用。 4. 让角色回复开口说话: 聊天里，+ 面板的「朗读」读最后一条回复。想一劳永逸就开 ⋮ 菜单 → 声音：选「语音模型」、挑音色（按模型分别记住）、打开「自动朗读」让每条新回复自动播，「只朗读台词」可以跳过旁白。 Success: 开自动朗读后，新回复会用选定的音色自动读出来。 5. 给伴侣一把专属声音: 在伴侣资料页，「TA 的声音」从当前语音模型的音色目录里给这位伴侣单独选声线。「和 TA 的约定 → 语音条」里打开「TA 会发语音条」，有些回复就会以语音气泡出现：第一次点开才合成计费，之后重播走缓存。你克隆过的音色会出现在这些选择器顶部的「我的音色」分组。 Success: 语音条第一次点开能播放，之后再点秒播不再扣费。 Success checks: - 引擎页能明确看出当前用的是系统引擎（免费）还是在线引擎（计费）。 - 章节播放能跨段连续进行，缓存列表随播放逐渐补齐。 - 选区配音固定在段落下方，重播走缓存不再计费。 - 每次在线合成都能在 API 请求记录里查到费用。 Cost and data safety: - Android 系统 TTS 免费且离线可用。在线音色按段合成、按段计费——BYOK 走你自己的 Key，官方渠道按积分（1 积分 = 0.0001 美元）。连续播放与预热会继续按段合成并计费。 - 已缓存的音频重播免费。换一个在线音色，该音色的段落会重新合成、再次计费；本地音色之间切换免费。聊天自动朗读是每条新回复合成一次；伴侣语音条只在第一次点开时计一次。 - 播放中断或合成被打断时，那一段的调用可能已经计费，只是你没听到。先到 API 请求记录按 TTS 筛选核对，再判断是不是失败；已经到手的缓存段照常免费重播。 - 被朗读的文字会发给你选择的语音供应商；生成的音频文件和缓存保存在本机。 Troubleshooting: - 朗读按钮是灰的，或点了提示要先配置模型 — Cause: 没有配语音合成模型——「没声音」的最高频原因。聊天里「语音模型」卡为空时朗读会灰显。 Fix: 去 我的 → AI 模型与服务 → 默认模型配置 设置「语音合成」，或在聊天「声音」页选语音模型。只是想听书的话，把引擎切成 Android 系统 TTS，零配置免费听。 - Key 配了，合成还是报「无效或没有权限」 — Cause: Key 和入口不匹配。尤其是阶跃：开放平台和 Step Plan 是两个独立入口，一边的 Key 在另一边就是 401。 Fix: 在服务商设置页按 Key 的权益类型选对入口，然后用「多模态能力（实测）」的语音测试跑到通过为止。 - 听着听着停了，提示余额不足或请求太频繁 — Cause: BYOK 场景是该服务商自己的余额空了（不是 Foreverse 积分），或请求触发限流；上游一拒绝，按段计费的合成就停在那。 Fix: 到该服务商充值，或稍等限流过去再继续——已缓存的段落照常免费播。若提示找不到语音模型（可能已下线），换一个音色。 - 换了音色又扣费，或声音没变 — Cause: 缓存按音色隔离，换在线音色意味着段落要重新合成；聊天里音色是按模型分别记忆的，换模型可能带回另一个音色。 Fix: 换在线音色后重新合成是预期行为，播放会从当前段落继续。换模型后检查一下「音色」行；想免费多试音色，用本地音色组。 FAQs: Q: 能完全不花钱听书吗？ A: 能。听书页「引擎」选 Android 系统 TTS：本地、离线、免费。在线音色是付费升级，按合成的段落计费。 Q: 同一段重播会再扣费吗？ A: 不会。合成过的段落有缓存，重播免费——伴侣语音条首播之后也一样。新费用来自新文字、换了在线音色，或连续播放合成到了新的段落。 Q: 能让角色用我自己的声音说话吗？ A: 可以。到 我的 → 高级设置 → AI 模型与服务 → 我的音色 克隆（只允许本人声音，有明确的同意确认步骤）。克隆好的音色会出现在听书、聊天和伴侣音色选择器的「我的音色」分组，用它合成按其归属服务商计费。 ### Generation Stopped or Failed: Triage Network, Key, Context, Rate Limits, and Recover Tasks URL: https://foreverse.app/docs/guide-models/fix-generation-errors Platform: Android generation and tasks · App: v76+ · Path verified: 2026-08-12 Direct answer: Open Me → Request log and tap the failed entry: the detail page shows status, cost, and duration, with the failure reason up top. Treat 401 as a key problem, 404 as host or model, 429 as quota or throttling, and timeouts as network. 'Context exceeds the model limit' means trim or compact history. An interrupted leave-and-return task shows a resume card on the next app start — check whether money was already spent before retrying. Before you start: - Foreverse v76 or later, and the failing feature can be reopened (chat, continuation, or media form). - Know which channel the call used — official credits or BYOK — because the bill lives in different places. Steps: 1. Separate 'failed' from 'interrupted': An error message or endless spinner is a failure — go to step 2. A long task you confirmed and then lost by leaving the page or having the process killed is an interruption — go to step 5. Either way, remember: a failure does not prove nothing was billed, and an interruption does not mean the money was wasted. The detail page and the resume card both state the cost status explicitly. 2. Read the exact error in the request log: Open Me → Request log (in the grid under the reader card), filter by modality, and tap the failed entry. The top summary card shows status, cost, and duration, with the failure reason visible. 401 means the key is invalid; 404 means the host path or model ID does not exist there; 429 means rate limit or quota; a timeout means the upstream is unreachable. Do not blame 'server down' before you have this evidence. Success: You can name the failure class: key, host/model, quota, network, or context. 3. Fix configuration-class errors in the BYOK guide: 401, 404, and 429 are key-and-provider configuration issues. The step-by-step fixes — versioned API host, exact model ID, billing and quota — are in /docs/guide-models/configure-byok-and-fix-errors and are not repeated here. For 'spins forever with no result', check Me → AI Models & Services → Default model config first: a missing model for that modality is the most common root cause. 4. Handle context overflow and rate limits on the spot: 'Context exceeds the model limit': in tavern chats, lower Context limit in Tavern global settings or enable the Auto summary plugin; in Agent sessions, use Compact context in the session options menu (one small text-model call). Recurring 429 with a healthy key: wait for the quota window, reduce concurrency, or pool several keys of the same provider under Provider settings → Manage multi-key & rotation — when one key hits 401, throttling, or an empty quota, the pool switches to the next key before the reply starts. Success: The retried request succeeds, or the pool visibly switches keys instead of failing the run. 5. Recover an interrupted leave-and-return task: Long tasks you confirmed with a quote — chat-to-novel conversion, migration imports — run with a resident progress notification. If the process is killed, the next app start shows a resume card ('interrupted at segment N of M'); tap Continue to return to the original form, re-confirm the quote, and carry on. Finished parts are saved and are not regenerated. If no card appears, reopen the original feature entry — it shows a Continue action. The master switch for task-completion notifications is Me → Notifications & reminders. Success: The task resumes from its checkpoint and previously completed chapters or pages are intact. Success checks: - The failed entry's error text has been read and mapped to key, host/model, quota, network, or context. - A small retry from the same feature succeeds and appears as a new successful entry in the request log. - The interrupted task was recovered via the resume card or the feature entry, with completed parts intact. - Before any 'may already be charged' decision, you checked the request log or the provider console. Cost and data safety: - Triage is free: reading the request log and changing settings never calls a model. Every retry or regeneration is a new billed call. - Failure does not always mean no charge: calls that definitely never reached the upstream (for example connection refused) are not charged; an 'outcome unknown' call may already be charged. The app flags these conservatively and never auto-retries them into double billing. - Completed parts of an interrupted task — written chapters, drawn pages — are saved and already paid for. Continuing only fills the gap; do not delete and rerun. - If media generated but saving failed, the generation itself may have been billed successfully: recover the asset or retry delivery instead of generating (and paying) again. - BYOK failures are billed by your provider; check its console for the truth. Do not hammer retry on expensive image or video calls. Troubleshooting: - A streaming reply cut off halfway — Cause: The network switched or the upstream timed out mid-stream. Fix: Check that entry's status and cost in the request log. Output already received is usually kept; regenerating is a new billed call. - 429 / rate limit keeps coming back — Cause: Requests-per-minute, tokens-per-minute, concurrency, or balance hit the provider ceiling. Fix: Read the exact quota in the provider console, wait for the reset, slow down bursts, or configure the key pool so exhausted keys rotate out automatically. - Spinner runs forever, no error, no result — Cause: The target modality has no usable model, or the upstream is unreachable. Fix: Check Default model config for an enabled model of that modality first, then read the newest failed entry in the request log — this order is the official triage. - 'N pages unfinished, may already be charged' after a batch task — Cause: Some pages of a serial-comic-style batch ended in an unknown state (for example a read timeout after submission). Fix: Both exits state their consequences: retrying those pages may pay again; abandoning them stops there and returns the monthly quota. Verify in the request log or provider usage before choosing. - Task vanished after killing the app or rebooting — Cause: The process died, but checkpoints live on disk. Fix: Cold start shows the resume card — tap Continue. Recovery always re-confirms the quote first; nothing silently resumes spending in the background. FAQs: Q: Does tapping Stop refund anything? A: No. Stop only prevents later steps; completed calls are billed and their output is kept. See /docs/guide-models/understand-billing-usage for how spending is recorded. Q: Will a failed generation still charge me? A: Three cases: definitely-not-sent failures are not charged; upstream-side failures follow the detail page's cost field and your provider bill; unknown outcomes are flagged 'may already be charged' and left for you to verify — the app never silently retries them. Q: How do I keep long tasks from being wasted work? A: Read the call count on the quote before confirming, allow notifications, and glance at the resident progress notification before leaving. Even after a process kill, checkpoints and finished content survive, and the resume card takes you back. ### 生成中断或失败怎么排查：网络、Key、上下文、限流与任务找回 URL: https://foreverse.app/zh/docs/guide-models/fix-generation-errors Platform: Android 生成与任务 · App: v76+ · Path verified: 2026-08-12 Direct answer: 生成失败先到请求记录点开失败条目看原因：401 查 Key、429 是限流、超时查网络；离场任务中断后重开 App 有恢复卡，确认是否扣费再继续。 Before you start: - Foreverse v76 以上，出问题的功能还能再打开（聊天、续写或媒体表单）。 - 知道这次调用走的是官方积分还是 BYOK——两边账单不在一个地方。 Steps: 1. 先分清「失败」和「中断」: 有错误提示或一直转圈，是失败，走第 2 步；确认报价后离开页面、进程被杀导致任务不见了，是中断，走第 5 步。两条路都先记住一件事：失败不代表没扣费，中断也不代表钱白花——详情页和恢复卡都会把费用状态写明白。 2. 到请求记录看错误原文: 我的 → 请求记录（读者证下方四宫格），按模态筛选，点失败条目进详情：顶部是「状态 / 费用 / 耗时」三个大数字，失败原因常显。401=Key 无效；404=地址或模型 ID 在那边不存在；429=限流或额度；超时=上游不通。没拿到证据之前，先别归因「服务器坏了」。 Success: 能把这次失败归进一类：Key、地址/模型、额度、网络，还是上下文。 3. 配置类错误去 BYOK 姊妹篇修: 401、404、429 属于 Key 与供应商配置问题，逐项修法（API Host 版本前缀、精确模型 ID、计费与额度）见教程《配置 BYOK，并排查 401、404、429 与模型不见了》（/zh/docs/guide-models/configure-byok-and-fix-errors），这里不重复。「一直转圈没结果」先查 我的 → AI 模型与服务 → 默认模型配置：对应模态没有可用模型，是最高频根因。 4. 上下文超限和限流的现场处理: 报「上下文超出模型上限」：酒馆聊天到酒馆全局设置调低「上下文上限」，或开自动摘要插件；Agent 会话用会话选项里的「压缩上下文」（一次小额文本调用）。429 反复出现而 Key 本身没问题：等配额窗口恢复、降低连发；同一家有多把 Key，就到服务商设置页 →「管理多 Key 与轮询」配 Key 池——某把 401、限流或额度耗尽时，会在回复开始前自动切下一把。 Success: 重试成功，或能看到 Key 池自动切换而不是整次失败。 5. 找回中断的可离场任务: 确认过报价、可以离开的长任务（把聊天转成小说、搬家导入等）运行中有常驻进度通知；进程被杀后，下次打开 App 会弹恢复卡「进行到 N/M 段时中断」，点「去继续」回到原表单、重新确认报价后接着跑——已完成的部分都保存着，不会从头重来。没弹卡就回原功能入口，能看到「继续」按钮。任务完成通知的总开关在 我的 → 通知与提醒。 Success: 任务从断点继续，之前写好的章节或画好的页都完好。 Success checks: - 失败条目的错误原文已读到，并归进 Key、地址/模型、额度、网络或上下文其中一类。 - 从同一入口做的小额重试成功，请求记录里出现新的成功条目。 - 中断任务经恢复卡或原入口找回，已完成部分完好。 - 做「可能已扣费」相关选择前，先核对过请求记录或供应商用量。 Cost and data safety: - 排查免费：看请求记录、改设置不调用模型；每次「重试 / 重新生成」都是新的一次计费调用。 - 失败不等于免单，也不等于必扣费：确定没到上游的（如连接拒绝）不会扣；「结果未知」的可能已扣费——App 会保守提示，绝不静默自动重试让你重复付费。 - 中断任务已完成的部分（写好的章节、画好的页）内容都在、费用已计；继续只补缺口，不要删掉重跑。 - 媒体显示保存失败时，生成本身可能已成功计费：先找回产物或重试保存，不要立刻重新生成再付一次。 - BYOK 的失败调用以供应商控制台账单为准；昂贵的图像、视频请求不要连点。 Troubleshooting: - 流式回复输出到一半断了 — Cause: 网络切换或上游中途超时。 Fix: 到请求记录看该条状态与费用；已输出的内容通常保留，重新生成是新调用、新计费。 - 429 / 限流反复出现 — Cause: 每分钟请求数、token 数、并发或余额触到供应商上限。 Fix: 去供应商控制台看具体配额，等窗口恢复、放慢连发；或配 Key 池，让耗尽的 Key 自动轮换出去。 - 一直转圈，没报错也没结果 — Cause: 对应模态没有可用模型，或上游不通。 Fix: 先查默认模型配置里该模态有没有已启用的模型，再看请求记录最新失败条目——这是官方排查次序。 - 批量任务提示「有 N 页没画完，可能已经扣过费」 — Cause: 漫画连载这类批量任务里，个别页结果未知（比如提交后读超时）。 Fix: 两个出口后果直说：重画这 N 页可能重复付费；放弃则不再画并退回本月次数。先对照请求记录或供应商用量再选。 - 杀了 App / 手机重启后任务不见了 — Cause: 进程死了，但断点都在本机。 Fix: 冷启动会弹恢复卡，点「去继续」。恢复一定先重新确认报价，不会不问你就在后台接着花钱。 FAQs: Q: 点「停止」能退钱吗？ A: 不能。停止只拦住后续步骤，已完成的调用照常计费、产出保留。钱怎么记账见 /zh/docs/guide-models/understand-billing-usage。 Q: 生成失败还会扣费吗？ A: 分三种：确定没发到上游的不扣；上游收到但失败的，以详情页费用栏和供应商账单为准；结果未知的按「可能已扣费」保守标注，由你核对后裁决——App 不会替你悄悄重试。 Q: 怎么避免长任务白跑？ A: 确认报价时看清调用次数；允许通知；离开前瞄一眼常驻进度通知。就算进程被杀，断点和已完成内容都在，恢复卡会带你接着跑。 ### Understand Billing and Usage: Where the Money Goes, Where to Check, How to Cap It URL: https://foreverse.app/docs/guide-models/understand-billing-usage Platform: Android account and usage · App: v76+ · Path verified: 2026-08-12 Direct answer: Foreverse bills the official channel in credits: 1 credit = US$0.0001, metered per call, never expiring. Your balance sits on the reader card in Me; the pricing table and the credit ledger are one tap away, and per-call details live in the request log. BYOK calls are billed by your own provider and never touch credits. Batch tasks always show a cost confirmation before starting, and refunds arrive as new positive ledger rows — history is never rewritten. Before you start: - Sign in to see the official-channel balance and ledger; BYOK spending lives in each provider's console. - Foreverse v76 or later. Builds without credit purchase or the official channel hide those blocks entirely — that is configuration, not a bug. Steps: 1. Know the two payment channels: Official channel: sign in and use it, metered in credits, 1 credit = US$0.0001 (10,000 credits = US$1), pay-per-use, credits never expire. BYOK: your own API key, billed directly by the provider — Foreverse adds no markup and deducts no credits. Text, image, and speech can each use a different channel. Success: For any feature you use, you can say which channel pays for it. 2. Check balance and unit prices: The ★ number on the reader card in Me is your credit balance; tap it for the Credits page and purchase packs. Inside, 'Pricing · how billing works' lists real unit prices for every official model — text as input/output/cache-hit prices per million tokens — issued by the server and synced with official model pricing, so quotes always come from this page, not word of mouth. The footer shows 'cache saved you N', computed from your last 30 days of requests. Success: You can find the unit price of the model you actually use. 3. Read the per-item ledgers: Me → Credit ledger (grid under the reader card) lists every row: purchases, chat spending, image generation, speech synthesis, refunds, redeem-code deposits, and welcome-grant deposits. For the technical view of one call — tokens, cache hits, duration — open Me → Request log and expand Technical details on the entry. For the long-run picture (token heatmap, streaks, model ranking), open Me → Journey. Success: A ledger row, its request-log entry, and the feature you used all tell the same story. 4. Use the gates before big spends: Any batch task that fires many calls (serial comics, for example) must show 'Confirm before starting': model × call count and who pays, and work begins only after you confirm. Changing the model or the config afterwards requires a fresh confirmation — nothing keeps charging on a stale authorization. Serial comics also carry a monthly generation quota: the app pauses when it is used up, and the quote warns 'N left this month' once less than 20% remains. For Agent sessions, set a per-run token cap (200k/500k/1M) under Run resources so a runaway task stops itself. Success: You have seen a cost confirmation at least once and know where the Agent token cap lives. 5. Read task accounting and refunds: When a long task is interrupted, accounting is three-way: finished parts are billed and their content is kept; calls that never reached the upstream are not billed; unknown outcomes are flagged 'may already be charged' and wait for your decision — retry (may pay again) or abandon. Refunds never rewrite history: the original charge row stays as-is and the refund arrives as a new positive ledger row that raises the balance, so both rows reconcile in the Credit ledger. Success: You can point at a charge row and, if one exists, its matching refund row. Success checks: - The reader card shows the balance and the pricing table shows unit prices for your usual models. - Every ledger expense maps to one real use, traceable to a request-log entry when needed. - Batch tasks showed 'Confirm before starting' and you can state the call count and the paying channel. - BYOK requests appear in the request log tagged BYOK and produce no credit-ledger rows. Cost and data safety: - Checking balance, pricing, ledgers, the request log, and Journey is free — none of it calls a model. - Signed out, the Credits page shows a sign-in prompt instead of a balance, and official-channel usage records do not sync. The request log itself stays on your device. - BYOK spending appears only on the provider's bill; Foreverse cannot refund money charged directly by a third party. - If media generated but saving failed, the cost is already booked and the content exists — recover or retry delivery instead of regenerating and paying again. - Pro is a one-time unlock for local quantity limits and includes no AI model usage; credits and Pro are separate things. Troubleshooting: - "Why was I charged credits?" — Cause: Every deduction has a source row. Fix: Find the row in the Credit ledger, open the matching request-log entry for model and usage, then check the pricing page's 'these places auto-charge' list — nearly all auto-spending features default to off. - Credits unchanged but my provider billed me — Cause: That call went through BYOK, not the official channel. Fix: The request log's channel column shows BYOK. The bill lives with your provider; credits are not involved. - No balance shown on the Credits page — Cause: You are signed out. Fix: The page shows a sign-in prompt; balance, ledger, and official usage appear after signing in. - Bought a credit pack but it has not arrived — Cause: Store payment receipts can lag. Fix: Refresh the Credit ledger and look for the purchase row; if it never appears, use Me → Help & feedback with the purchase time attached. - Got a refund but cannot find it — Cause: Refunds are appended rows, not edits to the original charge. Fix: Look for a positive 'refund' row in the Credit ledger; the balance rises with it while the original charge row stays for reconciliation. FAQs: Q: How much is one credit worth? A: 1 credit = US$0.0001, so 10,000 credits = US$1. Credits are metered per use and never expire; what each model costs per call is shown on the pricing page and in the request log. Q: Can something charge me without my knowing? A: Auto-charging entry points are all listed under 'these places auto-charge' on the pricing page, and nearly all default to off (exceptions like continuous audiobook playback are listed honestly). Batch tasks always show a cost confirmation before starting. Q: How do I keep spending down? A: Point background chores at a cheap lightweight-task model, enable per-turn plugins only where you need them, set the Agent per-run token cap, and read batch quotes before confirming. For what failures and interruptions cost, see /docs/guide-models/fix-generation-errors. ### 看懂计费与用量：钱花在哪、在哪看、怎么控制 URL: https://foreverse.app/zh/docs/guide-models/understand-billing-usage Platform: Android 账号与用量 · App: v76+ · Path verified: 2026-08-12 Direct answer: 官方渠道按积分计费，1 积分=0.0001 美元；余额、价目表和逐笔流水都在「我的」页，退款是新增一笔正向流水。BYOK 由供应商收费，不扣积分。 Before you start: - 看官方渠道余额与流水需要登录；BYOK 的花费在各供应商控制台看。 - Foreverse v76 以上。未开放积分购买或官方渠道的版本会整体隐藏对应区块，是配置不是故障。 Steps: 1. 分清两条付费渠道: 官方渠道：登录即用，按量扣积分，1 credit = 0.0001 美元（1 万积分 = 1 美元），用多少扣多少，积分永不过期。BYOK：自己的 API Key，由供应商直接收费——Foreverse 不加价、不扣积分。文本、生图、语音可以分别走不同渠道。 Success: 对常用的每个功能，都能说出这次是谁在收钱。 2. 看余额和单价: 「我的」页读者证卡上的 ★ 数字就是积分余额，点它进「积分」页，可购买积分包。页内「价目表 · 怎么扣费」列出全部官方模型的实付单价——文本按输入/输出/缓存命中三种价（每百万 token），价格由服务端统一下发、随官方定价同步，报价一律以这页为准，不听口头换算。页底「缓存帮你省了 N」按你近 30 天的请求现算。 Success: 能查到自己常用模型的单价。 3. 查逐笔明细: 我的 → 积分流水（读者证下方四宫格）逐笔列出：购买、对话扣费、图像生成、语音合成、退款、兑换码入账、体验金赠送。想看某一次调用的技术细节（token、缓存命中、耗时），去 我的 → 请求记录 点条目展开「技术详情」；长期总览（token 热力图、连续天数、模型排行）在 我的 → 世界旅程。 Success: 任意一笔流水、它对应的请求记录、你实际用的功能，三者能对上。 4. 花大钱之前有闸门: 一次要发很多调用的批量任务（比如漫画连载）动工前必弹「动工前确认」：写明模型 × 次数、谁付钱，确认了才开工；之后换模型或改配置要重新确认，不会拿旧授权继续扣。漫画连载另有月度生成额度：用完自动暂停，余量不足 20% 时报价单会提示「本月还能生成 N 次」。Agent 会话可在「运行资源」里设单回合 token 上限（200k/500k/1M 三档），任务跑飞会自己停下。 Success: 见过至少一次「动工前确认」，并知道 Agent 的 token 上限在哪设。 5. 读懂任务账和退款: 长任务中断时按三种状态记账：已完成的部分费用已计、内容保留；确定没发到上游的不扣费；「结果未知」的会标「可能已扣费」，等你核对后选重试（可能重复付费）或放弃。退款不改历史：原来那笔扣费记录原样保留，退款以新增一笔正向流水回到余额——两笔都在积分流水里，对账一目了然。 Success: 能指出一笔扣费流水，以及（如果有）与它对应的退款流水。 Success checks: - 读者证卡能看到余额，价目表页能查到常用模型单价。 - 积分流水里每笔支出都能对应一次真实使用，必要时能在请求记录找到那次调用。 - 批量任务开工前见过「动工前确认」，说得出调用几次、谁付钱。 - BYOK 请求在请求记录里标 BYOK，且没有产生任何积分流水。 Cost and data safety: - 看余额、价目表、流水、请求记录、世界旅程全部免费，不产生模型调用。 - 未登录时积分页显示登录引导而不是余额，官方渠道用量记录也不同步；请求记录本身只存在手机上。 - BYOK 花费只出现在供应商账单里；第三方直接收取的费用，Foreverse 无法代为退款。 - 媒体已生成但保存失败时，费用已经计了、内容还在——先找回产物或重试保存，不要重新生成再付一次。 - Pro 是一次性买断的本地数量解锁，不含任何 AI 模型调用费用；积分和 Pro 是两回事。 Troubleshooting: - 「为什么扣我积分」 — Cause: 每笔扣费都有来源流水。 Fix: 在积分流水里找到那笔，点开对应请求记录看模型与用量，再翻价目表页「这些地方会自动扣费」清单核对——自动花钱的功能绝大多数默认关。 - 积分没动，但供应商那边扣钱了 — Cause: 那次调用走的是 BYOK，不是官方渠道。 Fix: 请求记录的渠道列会标 BYOK；账单在供应商侧，积分不参与。 - 积分页看不到余额 — Cause: 当前未登录。 Fix: 积分页会显示登录引导；登录后余额、流水和官方用量才可见。 - 买了积分包没到账 — Cause: 商店支付回执可能有延迟。 Fix: 稍后刷新积分流水找「购买」入账行；一直没有就走 我的 → 帮助与反馈 提交反馈，附上购买时间。 - 收到退款但不知道在哪看 — Cause: 退款是新增流水，不是改旧账。 Fix: 在积分流水里找正向「退款」条目，余额同步增加；原来那笔扣费记录仍在，两笔对着看就是完整的账。 FAQs: Q: 1 积分到底值多少钱？ A: 1 credit = 0.0001 美元，即 1 万积分 = 1 美元。按量消耗、永不过期；每个模型每次用掉多少，见价目表页和请求记录。 Q: 会不会有我不知道的自动扣费？ A: 自动扣费入口全部列在价目表页「这些地方会自动扣费」清单里，绝大多数默认关闭（听书连播等例外如实标注）；批量任务动工前必弹报价确认。 Q: 怎么把花费控制住？ A: 后台小任务指到便宜的「轻量功能模型」；每轮额外调用的插件按需开；给 Agent 设单回合 token 上限；批量任务看清报价再确认。失败和中断时钱怎么算，见 /zh/docs/guide-models/fix-generation-errors。 ======================================================================== ## COMPLETE APP FEATURE INDEX — 268 visible, searchable entry points These are feature lookups, not 268 thin HowTo pages. Each entry mirrors the visible feature-index page and links to the closest module guide. English index: https://foreverse.app/docs/feature-index 中文索引: https://foreverse.app/zh/docs/feature-index ======================================================================== ## DOCS ### Foreverse Tutorials & Help — Setup, Continuation, Roleplay, BYOK, Troubleshooting URL: https://foreverse.app/docs Searchable Foreverse tutorials for setup, novel continuation, character-card chat, BYOK 401/404/429 errors, memory, Agent approvals, Companion messages, backup and migration, plus module guides and creator specs. ### Foreverse Complete Feature Index — 268 App Features and Entry Points URL: https://foreverse.app/docs/feature-index Search all 268 current Foreverse Android features by UI label or task, with a verified starting point, cost or data boundary, and links to the closest task and module guides. ### Home Library and Import Guide — Books, Characters, Companions, Safe Management URL: https://foreverse.app/docs/guide-worlds Use the current All, Books, Characters, and Companions panes; import txt / epub / md and Tavern assets; pin, export, and delete safely. ### Reader Guide — Selection AI, Options, Branches, Listening URL: https://foreverse.app/docs/guide-reader Current Reader paths: selection tools, three rewrite modes, story options, AI-segment controls, optional paragraph-tail shortcuts, branch switching from Contents, listening, and the four-button control pill. ### Roleplay Guide — Tavern Actions, Groups, IM Mode URL: https://foreverse.app/docs/guide-roleplay The current Tavern map: common + panel actions, message tools, Current status, three-layer Chat settings, QR and automations, groups, IM mode, and media. ### Agent and Companion Guide — Tasks, @, Skills URL: https://foreverse.app/docs/guide-agent Use Quick Dock and the Agent workbench to assign tasks, read tool cards, use @ references across books and cards, customize skills, and run a Companion. ### Models and Usage Guide — BYOK Setup and Request Records URL: https://foreverse.app/docs/guide-models Add providers, test connectivity, set per-modality defaults, understand official-channel billing, and read the API request record. ### Settings and Data Guide — Appearance, Assets, Export URL: https://foreverse.app/docs/guide-settings Use Me for themes and sideloading, the SillyTavern asset library, tavern preferences, data export and account deletion. ### Beautify Spec — Chat Cards, Stickers, HTML Rendering URL: https://foreverse.app/docs/beautify Creator reference: ```html WebView rendering and sandbox limits, fv-card native cards, [[sticker:]] syntax, the marker-regex-HTML pipeline, TavernHelper API support. ### Character Card Import Spec — PNG / JSON / .charx URL: https://foreverse.app/docs/character-cards Supported card containers and fields: PNG ccv3 chunks, JSON, V3 .charx zip with assets, V1-V3 compatibility, group JSON. ### Worldbook Spec — Triggering, Budget, V3 Decorators URL: https://foreverse.app/docs/worldbook Exact runtime behavior: keyword scanning and depth, constant entries, recursion, token-budget drops, and @@activate / @@dont_activate semantics. ### Glossary — Tavern, Character Cards, and AI Reading, Defined URL: https://foreverse.app/docs/glossary High-frequency terms defined in one place: character card / chara_card_v3 / lorebook / constant entries / recursion / scan depth / regex / presets / macros / swipe / BYOK / prompt caching / context window / branch / VN theater and more — 27 self-contained, quotable entries with further reading. ### Theme Sideload Spec — .fvtheme.json Fields URL: https://foreverse.app/docs/themes The .fvtheme.json format: top-level fields, all FvPalette semantic color keys, ARGB notation, fallback rules and design guidelines. ### Foreverse 教程与帮助 — 上手、续写、角色聊天、BYOK 与排障 URL: https://foreverse.app/zh/docs 可搜索的 Foreverse 官方教程与帮助：快速上手、导书续写、角色卡开聊、BYOK 401/404/429 排障、世界书记忆、Agent 审批回滚、伴侣主动消息、导出换机，以及完整模块教程与创作者规范。 ### Foreverse 全部功能索引 — 268 个 App 功能、入口与费用边界 URL: https://foreverse.app/zh/docs/feature-index 可搜索的 Foreverse Android 功能地图：268 个当前功能逐项给出入口、用途、常见口语问法，以及费用或数据边界，并链接到相关任务教程和模块指南。 ### 主页资料库与导入教程 — 书、角色、伴侣与安全管理 URL: https://foreverse.app/zh/docs/guide-worlds 主页资料库完整教程：当前「全部、书、角色、伴侣」四分栏，txt / epub / md、角色卡与酒馆资产导入，以及置顶、导出和删除风险。 ### 阅读器完整教程 — 划词 AI、续写表单、选项生成、分支与听书 URL: https://foreverse.app/zh/docs/guide-reader 阅读器当前功能逐项说明：划词工具栏、改写三模式、剧情候选、AI 段编辑与继续、可选段尾快捷键、目录分支切换、听书，以及四键控制栏与五种翻页。 ### 角色聊天完整教程 — 酒馆动作、世界书、群聊与线上聊 URL: https://foreverse.app/zh/docs/guide-roleplay 酒馆玩家的当前功能地图：+ 面板常用动作、消息级操作（swipe/编辑/隐藏/checkpoint/分支）、当前状态与三层聊天设置、Quick Reply 与自动化、群聊、线上聊、图片/朗读/翻译。 ### 智能体与伴侣教程 — 任务、@ 引用、技能与陪伴 URL: https://foreverse.app/zh/docs/guide-agent Quick Dock 与智能体工作台教程：给 Agent 派任务、看懂思考流与工具卡、@ 引用小说和角色卡、自定义技能目录；以及伴侣（Companion）的记忆、主动消息和伴侣备份。 ### 模型与用量教程 — BYOK 配置、按模态默认与请求记录 URL: https://foreverse.app/zh/docs/guide-models 从零配好模型：添加供应商、填 Key、测连通、启停模型、按文/图/音/视频设默认；官方渠道计费说明；API 请求记录里 tokens、缓存命中和费用怎么看。 ### 设置与数据教程 — 外观、酒馆资产库与数据导出 URL: https://foreverse.app/zh/docs/guide-settings 「我的」里的设置与数据说明：主题外观与侧载主题、酒馆资产库（角色/世界书/预设/正则/QR/Persona）、酒馆偏好、数据导出与账号删除。 ### Foreverse 美化规范 — 聊天卡片、贴纸与 HTML 渲染 URL: https://foreverse.app/zh/docs/beautify 创作者美化手册：```html 围栏 WebView 渲染与沙箱边界、```fv-card 原生小卡三型、[[sticker:]] 贴纸语法、marker → 正则 → 主题化 HTML 的整卡美化链路、TavernHelper 宿主 API 支持清单。 ### Foreverse 角色卡导入规范 — PNG / JSON / .charx URL: https://foreverse.app/zh/docs/character-cards 支持的角色卡格式与字段：PNG ccv3/chara 隐写、JSON、V3 .charx zip 容器（含资产分类）、V1/V2/V3 字段兼容、nickname、群聊 JSON 导入。 ### Foreverse 世界书规范 — 触发、预算与 V3 decorators URL: https://foreverse.app/zh/docs/worldbook 世界书 runtime 行为参考：关键词触发与扫描深度、常驻条目、递归激活、token 预算丢弃规则，以及 V3 decorators（@@activate / @@dont_activate）的确切语义。 ### 术语表 — 酒馆、角色卡与 AI 阅读的黑话手册 URL: https://foreverse.app/zh/docs/glossary 高频术语集中定义：角色卡/chara_card_v3/世界书/常驻条目/递归/扫描深度/正则/预设/宏/swipe/BYOK/prompt 缓存/上下文窗口/分支/剧场模式等 27 条，每条自包含可引用，附站内延伸阅读。 ### Foreverse 主题侧载格式 — .fvtheme.json 完整字段 URL: https://foreverse.app/zh/docs/themes .fvtheme.json 主题包格式：顶层字段、FvPalette 全部语义色键、ARGB 颜色写法、回退规则与设计准则；文件放入 /foreverse/themes/ 即在设置 → 外观出现。 ======================================================================== Contact: support@foreverse.app Get the app: https://play.google.com/store/apps/details?id=app.foreverse.android&referrer=utm_source%3Dforeverse_web%26utm_medium%3Dweb_cta