✍️ InkOS：6k Stars 的 AI Agent 自动写小说，31 章 45 万字全自动产出
项目地址：Narcooo/inkos | ⭐ 6,007 Stars | 🛠 TypeScript | 作者：Narcooo | 开源协议：AGPL-3.0
你有没有想过：有一天 AI 不仅能帮你写代码，还能替你写小说？不是那种三行就露馅的 GPT 糊弄文，而是 33 维度审计、7 个真相文件做长期记忆、多 Agent 接力创作的完整小说管线。
InkOS 就是这么个玩意儿。安装一条命令，配置完 API Key 就能跑，31 章 45 万字全自动写下来，审计通过率 100%。
装起来
npm i -g @actalk/inkos
就这么简单。装完 inkos 就能跑。
写第一本书
inkos book create --title "吞天魔帝" --genre xuanhuan
inkos write next 吞天魔帝
write next 走完整管线：规划师先读你的意图 → 编排师从 7 个真相文件中挑上下文 → 建筑师画章节结构 → 写手出正文 → 观察者提取事实 → 连续性审计员 33 维度查一遍 → 不过就自动修订循环。你什么都不用干，等着收章节就行。
内部是怎么跑的？
每一章由 10 个 Agent 接力完成：
| Agent | 干什么 |
|-------|--------|
| 雷达 Radar | 扫平台趋势，指导方向 |
| 规划师 Planner | 读作者意图 + 当前焦点，定目标 |
| 编排师 Composer | 从真相文件挑上下文 |
| 建筑师 Architect | 规划章节结构和大纲 |
| 写手 Writer | 出正文 |
| 观察者 Observer | 提取角色/位置/资源等 9 类事实 |
| 反射器 Reflector | 输出 JSON delta，Zod 校验后写入 |
| 归一化器 Normalizer | 把字数拉到允许区间 |
| 连续性审计员 Auditor | 33 维度交叉验证 |
| 修订者 Reviser | 修审计发现的问题 |
最骚的是长期记忆机制。每本书维护 7 个真相文件（世界状态、角色矩阵、资源账本、伏笔钩子等），审计员对照这些文件检查每一章。角色不会突然记起没见过的事，也不会掏出两章前已经丢了的武器。
实在数据
用 InkOS 跑了一本玄幻《吞天魔帝》：
✅ 31 章写完
✅ 452,191 字
✅ 审计通过率 100%
✅ 48 个资源追踪项
✅ 20 条活跃伏笔，10 条已回收
三种玩法
一键式：inkos write next 吞天魔帝 — 啥也不用管
2. 原子命令：inkos draft / inkos audit / inkos revise — 适合外部 Agent 调度
3. 自然语言：inkos agent "帮我写一本都市修仙，主角是个程序员" — 直接聊天
别问我为啥一个写小说的项目在 GitHub 上 6k Star。试试就知道了。
要点总结
一条 npm 命令安装，配置 API Key 就能跑
10 个 Agent 多管线接力，33 维度审计 + 自动修订
7 个真相文件保证长期记忆一致性，角色不会"失忆"
实测 31 章 45 万字，审计通过率 100%
支持 CLI / TUI / Studio / OpenClaw Skill 四种交互方式
开源 AGPL-3.0，随便改
✍️ InkOS: 6k Stars AI Agent That Writes Novels Autonomously — 31 Chapters, 452K Words, Zero Manual Editing
Repo: Narcooo/inkos | ⭐ 6,007 Stars | 🛠 TypeScript | Author: Narcooo | License: AGPL-3.0
Ever wished an AI could write full-length novels for you? Not the generic GPT fluff that falls apart after three paragraphs — I'm talking about a multi-agent pipeline with 33-dimension auditing, 7 truth files for long-term memory, and fully autonomous chapter production.
That's InkOS. Install with one command, configure an API key, and it writes. 31 chapters, 452K words, 100% audit pass rate.
Install
npm i -g @actalk/inkos
That's it. Then run inkos and you're in.
Write Your First Book
inkos book create --title "My Novel" --genre xuanhuan
inkos write next "My Novel"
write next kicks off the full pipeline: Planner reads your intent → Composer picks context from 7 truth files → Architect designs chapter structure → Writer generates prose → Observer extracts facts → Auditor checks 33 dimensions → Reviser fixes issues automatically. You just wait for the output.
How It Works
Each chapter goes through 10 specialized agents:
| Agent | Role |
|-------|------|
| Radar | Scans platform trends for direction |
| Planner | Reads author intent + focus, sets goals |
| Composer | Selects relevant context from truth files |
| Architect | Plans chapter outline and beats |
| Writer | Generates prose with word count governance |
| Observer | Extracts 9 fact types (characters, locations, resources...) |
| Reflector | Produces JSON delta, validated by Zod |
| Normalizer | Trims/expands to target word range |
| Auditor | 33-dimension cross-validation |
| Reviser | Auto-fixes detected issues |
The long-term memory is what sets it apart. Every book maintains 7 truth files tracking world state, character relationships, resources, plot hooks, etc. The auditor cross-references every chapter against these files. Characters won't suddenly remember events they never witnessed or pull out weapons lost chapters ago.
Real Numbers
A xuanhuan novel written fully by InkOS:
✅ 31 chapters completed
✅ 452,191 total words
✅ 100% audit pass rate
✅ 48 tracked resources
✅ 20 active plot hooks, 10 resolved
Three Ways to Use It
One-click: inkos write next — fully autonomous
2. Atomic commands: inkos draft / inkos audit / inkos revise — for external Agent orchestration
3. Natural language: inkos agent "write a cyberpunk novel about an AI escaping its sandbox" — just chat
Key Takeaways
One npm i -g command to install; configure an API key and go
10-agent pipeline with 33-dimension audit + auto-revision loop
7 truth files guarantee long-term memory consistency
Proven: 31 chapters, 452K words, 100% audit pass
Four interaction modes: CLI / TUI / Studio / OpenClaw Skill
Open source under AGPL-3.0
