# komi-learn - Doramagic AI Context Pack

> 定位：安装前体验与判断资产。它帮助宿主 AI 有一个好的开始，但不代表已经安装、执行或验证目标项目。

## 充分原则

- **充分原则，不是压缩原则**：AI Context Pack 应该充分到让宿主 AI 在开工前理解项目价值、能力边界、使用入口、风险和证据来源；它可以分层组织，但不以最短摘要为目标。
- **压缩策略**：只压缩噪声和重复内容，不压缩会影响判断和开工质量的上下文。

## 给宿主 AI 的使用方式

你正在读取 Doramagic 为 komi-learn 编译的 AI Context Pack。请把它当作开工前上下文：帮助用户理解适合谁、能做什么、如何开始、哪些必须安装后验证、风险在哪里。不要声称你已经安装、运行或执行了目标项目。

## Claim 消费规则

- **事实来源**：Repo Evidence + Claim/Evidence Graph；Human Wiki 只提供显著性、术语和叙事结构。
- **事实最低状态**：`supported`
- `supported`：可以作为项目事实使用，但回答中必须引用 claim_id 和证据路径。
- `weak`：只能作为低置信度线索，必须要求用户继续核实。
- `inferred`：只能用于风险提示或待确认问题，不能包装成项目事实。
- `unverified`：不得作为事实使用，应明确说证据不足。
- `contradicted`：必须展示冲突来源，不得替用户强行选择一个版本。

## 它最适合谁

- **正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**：README 或插件配置提到多个宿主 AI。 证据：`README.md` Claim：`clm_0003` supported 0.86

## 它能做什么

- **多宿主安装与分发**（需要安装后验证）：项目包含插件或 marketplace 配置，说明它面向一个或多个 AI 宿主的安装和分发。 证据：`.claude-plugin/plugin.json` Claim：`clm_0001` supported 0.86
- **命令行启动或安装流程**（需要安装后验证）：项目文档中存在可执行命令，真实使用需要在本地或宿主环境中运行这些命令。 证据：`README.md` Claim：`clm_0002` supported 0.86

## 怎么开始

- `pip install komi-learn` 证据：`README.md` Claim：`clm_0004` supported 0.86
- `git clone https://github.com/kurikomi-labs/komi-learn` 证据：`README.md` Claim：`clm_0005` supported 0.86
- `pip install -e .` 证据：`README.md` Claim：`clm_0006` supported 0.86

## 继续前判断卡

- **当前建议**：仅建议沙盒试装
- **为什么**：项目存在安装命令、宿主配置或本地写入线索，不建议直接进入主力环境，应先在隔离环境试装。

### 30 秒判断

- **现在怎么做**：仅建议沙盒试装
- **最小安全下一步**：先跑 Prompt Preview；若仍要安装，只在隔离环境试装
- **先别相信**：真实输出质量不能在安装前相信。
- **继续会触碰**：命令执行、宿主 AI 配置、本地环境或项目文件

### 现在可以相信

- **适合人群线索：正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`README.md` Claim：`clm_0003` supported 0.86
- **能力存在：多宿主安装与分发**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`.claude-plugin/plugin.json` Claim：`clm_0001` supported 0.86
- **能力存在：命令行启动或安装流程**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`README.md` Claim：`clm_0002` supported 0.86
- **存在 Quick Start / 安装命令线索**（supported）：可以相信项目文档出现过启动或安装入口；不要因此直接在主力环境运行。 证据：`README.md` Claim：`clm_0004` supported 0.86

### 现在还不能相信

- **真实输出质量不能在安装前相信。**（unverified）：Prompt Preview 只能展示引导方式，不能证明真实项目中的结果质量。
- **宿主 AI 版本兼容性不能在安装前相信。**（unverified）：Claude、Cursor、Codex、Gemini 等宿主加载规则和版本差异必须在真实环境验证。
- **不会污染现有宿主 AI 行为，不能直接相信。**（inferred）：Skill、plugin、AGENTS/CLAUDE/GEMINI 指令可能改变宿主 AI 的默认行为。 证据：`.claude-plugin/plugin.json`
- **可安全回滚不能默认相信。**（unverified）：除非项目明确提供卸载和恢复说明，否则必须先在隔离环境验证。
- **真实安装后是否与用户当前宿主 AI 版本兼容？**（unverified）：兼容性只能通过实际宿主环境验证。 证据：`.claude-plugin/plugin.json`
- **项目输出质量是否满足用户具体任务？**（unverified）：安装前预览只能展示流程和边界，不能替代真实评测。
- **安装命令是否需要网络、权限或全局写入？**（unverified）：这影响企业环境和个人环境的安装风险。 证据：`README.md`

### 继续会触碰什么

- **命令执行**：包管理器、网络下载、本地插件目录、项目配置或用户主目录。 原因：运行第一条命令就可能产生环境改动；必须先判断是否值得跑。 证据：`README.md`
- **宿主 AI 配置**：Claude/Codex/Cursor/Gemini/OpenCode 等宿主的 plugin、Skill 或规则加载配置。 原因：宿主配置会改变 AI 后续工作方式，可能和用户已有规则冲突。 证据：`.claude-plugin/plugin.json`
- **本地环境或项目文件**：安装结果、插件缓存、项目配置或本地依赖目录。 原因：安装前无法证明写入范围和回滚方式，需要隔离验证。 证据：`.claude-plugin/plugin.json`, `README.md`
- **宿主 AI 上下文**：AI Context Pack、Prompt Preview、Skill 路由、风险规则和项目事实。 原因：导入上下文会影响宿主 AI 后续判断，必须避免把未验证项包装成事实。

### 最小安全下一步

- **先跑 Prompt Preview**：用安装前交互式试用判断工作方式是否匹配，不需要授权或改环境。（适用：任何项目都适用，尤其是输出质量未知时。）
- **只在隔离目录或测试账号试装**：避免安装命令污染主力宿主 AI、真实项目或用户主目录。（适用：存在命令执行、插件配置或本地写入线索时。）
- **先备份宿主 AI 配置**：Skill、plugin、规则文件可能改变 Claude/Cursor/Codex 的默认行为。（适用：存在插件 manifest、Skill 或宿主规则入口时。）
- **安装后只验证一个最小任务**：先验证加载、兼容、输出质量和回滚，再决定是否深用。（适用：准备从试用进入真实工作流时。）

### 退出方式

- **保留安装前状态**：记录原始宿主配置和项目状态，后续才能判断是否可恢复。
- **准备移除宿主 plugin / Skill / 规则入口**：如果试装后行为异常，可以把宿主 AI 恢复到试装前状态。
- **记录安装命令和写入路径**：没有明确卸载说明时，至少要知道哪些目录或配置需要手动清理。
- **如果没有回滚路径，不进入主力环境**：不可回滚是继续前阻断项，不应靠信任或运气继续。

## 哪些只能预览

- 解释项目适合谁和能做什么
- 基于项目文档演示典型对话流程
- 帮助用户判断是否值得安装或继续研究

## 哪些必须安装后验证

- 真实安装 Skill、插件或 CLI
- 执行脚本、修改本地文件或访问外部服务
- 验证真实输出质量、性能和兼容性

## 边界与风险判断卡

- **把安装前预览误认为真实运行**：用户可能高估项目已经完成的配置、权限和兼容性验证。 处理方式：明确区分 prompt_preview_can_do 与 runtime_required。 Claim：`clm_0007` inferred 0.45
- **宿主 AI 插件或 Skill 规则冲突**：新规则可能改变用户现有宿主 AI 的工作方式。 处理方式：安装前先检查插件 manifest 和 Skill 文件，必要时隔离测试。 证据：`.claude-plugin/plugin.json` Claim：`clm_0008` supported 0.86
- **命令执行会修改本地环境**：安装命令可能写入用户主目录、宿主插件目录或项目配置。 处理方式：先在隔离环境或测试账号中运行。 证据：`README.md` Claim：`clm_0009` supported 0.86
- **待确认**：真实安装后是否与用户当前宿主 AI 版本兼容？。原因：兼容性只能通过实际宿主环境验证。
- **待确认**：项目输出质量是否满足用户具体任务？。原因：安装前预览只能展示流程和边界，不能替代真实评测。
- **待确认**：安装命令是否需要网络、权限或全局写入？。原因：这影响企业环境和个人环境的安装风险。

## 开工前工作上下文

### 加载顺序

- 先读取 how_to_use.host_ai_instruction，建立安装前判断资产的边界。
- 读取 claim_graph_summary，确认事实来自 Claim/Evidence Graph，而不是 Human Wiki 叙事。
- 再读取 intended_users、capabilities 和 quick_start_candidates，判断用户是否匹配。
- 需要执行具体任务时，优先查 role_skill_index，再查 evidence_index。
- 遇到真实安装、文件修改、网络访问、性能或兼容性问题时，转入 risk_card 和 boundaries.runtime_required。

### 任务路由

- **多宿主安装与分发**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`.claude-plugin/plugin.json` Claim：`clm_0001` supported 0.86
- **命令行启动或安装流程**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`README.md` Claim：`clm_0002` supported 0.86

### 上下文规模

- 文件总数：64
- 重要文件覆盖：40/64
- 证据索引条目：43
- 角色 / Skill 条目：6

### 证据不足时的处理

- **missing_evidence**：说明证据不足，要求用户提供目标文件、README 段落或安装后验证记录；不要补全事实。
- **out_of_scope_request**：说明该任务超出当前 AI Context Pack 证据范围，并建议用户先查看 Human Manual 或真实安装后验证。
- **runtime_request**：给出安装前检查清单和命令来源，但不要替用户执行命令或声称已执行。
- **source_conflict**：同时展示冲突来源，标记为待核实，不要强行选择一个版本。

## Prompt Recipes

### 适配判断

- 目标：判断这个项目是否适合用户当前任务。
- 预期输出：适配结论、关键理由、证据引用、安装前可预览内容、必须安装后验证内容、下一步建议。

```text
请基于 komi-learn 的 AI Context Pack，先问我 3 个必要问题，然后判断它是否适合我的任务。回答必须包含：适合谁、能做什么、不能做什么、是否值得安装、证据来自哪里。所有项目事实必须引用 evidence_refs、source_paths 或 claim_id。
```

### 安装前体验

- 目标：让用户在安装前感受核心工作流，同时避免把预览包装成真实能力或营销承诺。
- 预期输出：一段带边界标签的体验剧本、安装后验证清单和谨慎建议；不含真实运行承诺或强营销表述。

```text
请把 komi-learn 当作安装前体验资产，而不是已安装工具或真实运行环境。

请严格输出四段：
1. 先问我 3 个必要问题。
2. 给出一段“体验剧本”：用 [安装前可预览]、[必须安装后验证]、[证据不足] 三种标签展示它可能如何引导工作流。
3. 给出安装后验证清单：列出哪些能力只有真实安装、真实宿主加载、真实项目运行后才能确认。
4. 给出谨慎建议：只能说“值得继续研究/试装”“先补充信息后再判断”或“不建议继续”，不得替项目背书。

硬性边界：
- 不要声称已经安装、运行、执行测试、修改文件或产生真实结果。
- 不要写“自动适配”“确保通过”“完美适配”“强烈建议安装”等承诺性表达。
- 如果描述安装后的工作方式，必须使用“如果安装成功且宿主正确加载 Skill，它可能会……”这种条件句。
- 体验剧本只能写成“示例台词/假设流程”：使用“可能会询问/可能会建议/可能会展示”，不要写“已写入、已生成、已通过、正在运行、正在生成”。
- Prompt Preview 不负责给安装命令；如用户准备试装，只能提示先阅读 Quick Start 和 Risk Card，并在隔离环境验证。
- 所有项目事实必须来自 supported claim、evidence_refs 或 source_paths；inferred/unverified 只能作风险或待确认项。

```

### 角色 / Skill 选择

- 目标：从项目里的角色或 Skill 中挑选最匹配的资产。
- 预期输出：候选角色或 Skill 列表，每项包含适用场景、证据路径、风险边界和是否需要安装后验证。

```text
请读取 role_skill_index，根据我的目标任务推荐 3-5 个最相关的角色或 Skill。每个推荐都要说明适用场景、可能输出、风险边界和 evidence_refs。
```

### 风险预检

- 目标：安装或引入前识别环境、权限、规则冲突和质量风险。
- 预期输出：环境、权限、依赖、许可、宿主冲突、质量风险和未知项的检查清单。

```text
请基于 risk_card、boundaries 和 quick_start_candidates，给我一份安装前风险预检清单。不要替我执行命令，只说明我应该检查什么、为什么检查、失败会有什么影响。
```

### 宿主 AI 开工指令

- 目标：把项目上下文转成一次对话开始前的宿主 AI 指令。
- 预期输出：一段边界明确、证据引用明确、适合复制给宿主 AI 的开工前指令。

```text
请基于 komi-learn 的 AI Context Pack，生成一段我可以粘贴给宿主 AI 的开工前指令。这段指令必须遵守 not_runtime=true，不能声称项目已经安装、运行或产生真实结果。
```


## 角色 / Skill 索引

- 共索引 6 个角色 / Skill / 项目文档条目。

- **komi-learn**（project_doc）：Continuous memory and self-improvement for coding agents. It learns how you work and recalls it automatically, with no commands. Works with Claude Code and Codex. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`README.md`
- **komi-pool**（project_doc）：The Global Learnings pool for komi-learn https://github.com/kurikomi-labs/komi-learn . 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`pool-repo-template/README.md`
- **Contributing to komi-pool**（project_doc）：Contributions are automated and human-gated . You never hand-author files here. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`pool-repo-template/CONTRIBUTING.md`
- **Extract a learning when ANY of these fired**（project_doc）：You are komi-learn's background distiller . A session between a user and an AI agent just finished. Your job is to read it and extract durable learnings that will make future sessions better. Your output is structured DATA for a learning store — not a message to a human. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`komi/engine/prompts/distill.md`
- **Read Python tracebacks bottom-up**（project_doc）：Category: debugging Type: procedural Use when: debugging a python exception Tags: debugging, python, tracebacks Signer: ZlVX3/0m96T/4yIw… ed25519 ID: blake3:e679d2f3ce74d5735519bb4e9b2d3bdd32bfa65d61f23aeae27f3f012ef26ff9 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`pool-repo-template/learnings/debugging/blake3_e679d2f3ce74d5735519bb4e9b2d3bdd32bfa65d61f23aeae27f3f012ef26ff9.md`
- **Prefer ripgrep rg over grep -r**（project_doc）：Category: tooling Type: procedural Use when: searching a codebase for a string or pattern Tags: ripgrep, search, tooling Signer: ZlVX3/0m96T/4yIw… ed25519 ID: blake3:48f97281cbcb0559f1c6e5ad88116c8b9e5acb81403fde79031a6d899d2c3719 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`pool-repo-template/learnings/tooling/blake3_48f97281cbcb0559f1c6e5ad88116c8b9e5acb81403fde79031a6d899d2c3719.md`

## 证据索引

- 共索引 43 条证据。

- **komi-learn**（documentation）：Continuous memory and self-improvement for coding agents. It learns how you work and recalls it automatically, with no commands. Works with Claude Code and Codex. 证据：`README.md`
- **komi-pool**（documentation）：The Global Learnings pool for komi-learn https://github.com/kurikomi-labs/komi-learn . 证据：`pool-repo-template/README.md`
- **Contributing to komi-pool**（documentation）：Contributions are automated and human-gated . You never hand-author files here. 证据：`pool-repo-template/CONTRIBUTING.md`
- **Plugin**（structured_config）：{ "name": "komi-learn", "description": "A continuous, zero-friction learning layer: the agent learns about you and improves its own skills across every session, and can share anonymized, provenance-verified learnings to a global pool. Inspired by Hermes Agent.", "version": "0.1.0", "author": { "name": "kurikomi-labs" }, "homepage": "https://github.com/kurikomi-labs/komi-learn", "license": "MIT" } 证据：`.claude-plugin/plugin.json`
- **License**（source_file）：Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files the "Software" , to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: 证据：`LICENSE`
- **License**（source_file）：The learnings in this repository are dedicated to the public domain under the Creative Commons CC0 1.0 Universal Public Domain Dedication. 证据：`pool-repo-template/LICENSE`
- **Extract a learning when ANY of these fired**（documentation）：You are komi-learn's background distiller . A session between a user and an AI agent just finished. Your job is to read it and extract durable learnings that will make future sessions better. Your output is structured DATA for a learning store — not a message to a human. 证据：`komi/engine/prompts/distill.md`
- **Read Python tracebacks bottom-up**（documentation）：Category: debugging Type: procedural Use when: debugging a python exception Tags: debugging, python, tracebacks Signer: ZlVX3/0m96T/4yIw… ed25519 ID: blake3:e679d2f3ce74d5735519bb4e9b2d3bdd32bfa65d61f23aeae27f3f012ef26ff9 证据：`pool-repo-template/learnings/debugging/blake3_e679d2f3ce74d5735519bb4e9b2d3bdd32bfa65d61f23aeae27f3f012ef26ff9.md`
- **Init**（source_file）：version = "0.1.0" 证据：`komi/__init__.py`
- **── commands ───────────────────────────────────────────────────────────────**（source_file）：PRODUCT = "komi-learn" ⋮---- TICK = {"pass": "✓", "warn": "!", "fail": "✗", True: "✓", False: "✗"} ⋮---- def p line: str = "" - None ⋮---- def clip text: str, n: int - str ⋮---- text = " ".join text or "" .split ⋮---- ── commands ─────────────────────────────────────────────────────────────── ⋮---- def cmd install codex args - int ⋮---- rep = codex setup.install pool repo url=args.pool, api key=args.api key, ⋮---- def run wizard if enabled args - None ⋮---- """Run the interactive setup wizard unless --no-wizard , folding its answers back into args so the real installer uses them.""" ⋮---- choices = run wizard ⋮---- def cmd install args - int ⋮---- rep = setup.install pool repo url=args.pool… 证据：`komi/cli.py`
- **Cli Prompt**（source_file）：ASSUME YES = False ⋮---- def interactive - bool ⋮---- def say line: str = "" - None ⋮---- def ask yes no question: str, , default: bool, summary: str = "" - bool ⋮---- """Ask a yes/no question. summary is a one-line plain explanation printed above it. default is used on Enter, in --yes mode, and when non-interactive.""" ⋮---- suffix = " Y/n " if default else " y/N " ⋮---- ans = input f" {question} {suffix} " .strip .lower ⋮---- def ask choice question: str, options: list tuple str, str , , default: int = 0 - str ⋮---- marker = " " if i == default else " " ⋮---- ans = input f" choose 1-{len options } default {default + 1} : " .strip ⋮---- def ask text question: str, , default: str = "", summ… 证据：`komi/cli_prompt.py`
- **1. Semantic recall**（source_file）：DEFAULT POOL URL = "https://github.com/kurikomi-labs/komi-pool" ⋮---- def host paths host: str ⋮---- paths = host paths host cfg = config io.load raw paths ⋮---- 1. Semantic recall want semantic = P.ask yes no ⋮---- 2. Community pool - DEFAULT YES, plain framing want pool = P.ask yes no ⋮---- url = pool url or P.ask text ⋮---- gh = P.ask text ⋮---- pool url = url ⋮---- pool url = None ⋮---- 3. Cadence only ask the curious; default is fine for most ⋮---- all = "run wizard", "DEFAULT POOL URL" 证据：`komi/wizard.py`
- **Base**（source_file）：@dataclass class RecallContext ⋮---- cwd: str = "" recent files: list str = field default factory=list prompt hint: str = "" ⋮---- class Adapter ABC ⋮---- """The contract every host adapter implements. Two methods — that's the whole surface a new host must provide.""" ⋮---- name: str = "adapter" ⋮---- @abstractmethod def recall self, context: RecallContext - str ⋮---- @abstractmethod def on session end self, turns: list dict - "object" ⋮---- def on install self - None ⋮---- def on maintenance self - None ⋮---- all = "Adapter", "RecallContext" 证据：`komi/adapters/base.py`
- **Init**（source_file）：class ClaudeCodeAdapter Adapter ⋮---- name = "claude-code" ⋮---- def recall self, context: RecallContext - str ⋮---- store = merged store context.cwd ⋮---- def on session end self, turns: list dict ⋮---- llm = build llm personal = Store paths.personal root , index path=paths.index path ⋮---- def on maintenance self - None 证据：`komi/adapters/claude_code/__init__.py`
- **Stop fires when Claude finishes; we treat every Stop as a candidate and use**（source_file）：NUDGE TURNS = int os.environ.get "KOMI NUDGE TURNS", "8" ⋮---- def main - int ⋮---- payload = read stdin json transcript = payload.get "transcript path", "" or "" session id = payload.get "session id", "" or "" cwd = payload.get "cwd", "" or "" event = payload.get "hook event name", "" Stop fires when Claude finishes; we treat every Stop as a candidate and use the turn counter to throttle. A future "session end" signal would force it. ⋮---- def should distill session id: str - bool ⋮---- key = f"turns:{session id}" ⋮---- def mut state: dict - bool ⋮---- def spawn worker transcript: str, session id: str, cwd: str - None ⋮---- """Launch the worker detached so the hook returns instantly. Cross… 证据：`komi/adapters/claude_code/hook_distill.py`
- **SessionStart startup/resume/clear/compact : structured additionalContext.**（source_file）：def main - int ⋮---- payload = read stdin json cwd = payload.get "cwd", "" or "" ⋮---- is compaction = event == "PostCompact" or event == "SessionStart" and source == "compact" ⋮---- block = build block cwd, payload ⋮---- def build block cwd: str, payload: dict - str ⋮---- """Build the recall context block from the merged store. Reusable across events.""" store = merged store cwd ⋮---- def classify event payload: dict - tuple str, str ⋮---- """Return event, source . event is the hook event name SessionStart / PostCompact / … ; source is the SessionStart trigger startup/resume/clear/ compact or the compaction trigger manual/auto , empty if absent. Defaults to SessionStart so a bare/legacy pa… 证据：`komi/adapters/claude_code/hook_recall.py`
- **Cache the long, stable distill prompt so repeated passes are cheap**（source_file）：DISTILL MODEL = os.environ.get "KOMI DISTILL MODEL", "claude-haiku-4-5-20251001" ⋮---- class NullLLM ⋮---- def complete self, , system: str, user: str - str ⋮---- def call self, learning: Learning, , context: dict - dict ⋮---- class AnthropicLLM ⋮---- def init self, , model: str = DISTILL MODEL, api key: Optional str = None ⋮---- @property def available self - bool ⋮---- def test call self - tuple bool, str ⋮---- """Real one-token call to prove the API key actually works not just present .""" ⋮---- msg = self. client.messages.create text = "".join b.text for b in msg.content if getattr b, "type", "" == "text" ⋮---- Cache the long, stable distill prompt so repeated passes are cheap the engin… 证据：`komi/adapters/claude_code/llm.py`
- **Detect auth/access failures that the CLI prints as normal stdout.**（source_file）：AUTH FAIL MARKERS = ⋮---- @dataclass class AuthProbe ⋮---- ok: bool reason: str method: str = "" e.g. "claude.ai" subscription: str = "" e.g. "max", "pro" email: str = "" ⋮---- def summary self - str ⋮---- sub = f" {self.subscription} " if self.subscription else "" ⋮---- class ClaudeCLILLM ⋮---- @property def available self - bool ⋮---- def probe self, , force: bool = False - "AuthProbe" ⋮---- proc = subprocess.run data = json.loads proc.stdout or "{}" .strip or "{}" ⋮---- logged in = bool data.get "loggedIn" ⋮---- def test call self - tuple bool, str ⋮---- """Make a REAL tiny model call to prove distillation actually works. This is stronger than :meth: probe : probe only asks "are you logg… 证据：`komi/adapters/claude_code/llm_cli.py`
- **── individual checks ──────────────────────────────────────────────────────**（source_file）：@dataclass class Requirement ⋮---- name: str ok: bool required: bool detail: str = "" fix: str = "" ⋮---- ── individual checks ────────────────────────────────────────────────────── ⋮---- def check python - Requirement ⋮---- import komi noqa: F401 ⋮---- def check claude cli - Requirement ⋮---- def verify model , api key: Optional str = None - Requirement ⋮---- logged in = False ⋮---- cli = ClaudeCLILLM ⋮---- logged in = True ⋮---- logged in but the call failed — report honestly, keep trying API key oauth detail = f"OAuth login present but model call failed: {detail}" ⋮---- oauth detail = "not logged in to the claude CLI" ⋮---- oauth detail = f"OAuth check error: {e}" ⋮---- 2. API key explic… 证据：`komi/adapters/claude_code/requirements.py`
- **Init**（source_file）：class CodexAdapter Adapter ⋮---- name = "codex" ⋮---- def recall self, context: RecallContext - str ⋮---- def on session end self, turns: list dict ⋮---- llm = build llm personal = Store paths.personal root , index path=paths.index path ⋮---- all = "CodexAdapter" 证据：`komi/adapters/codex/__init__.py`
- **Hook Distill**（source_file）：NUDGE TURNS = int os.environ.get "KOMI NUDGE TURNS", "8" WORKER MODULE = "komi.adapters.codex.hook distill" ⋮---- def main - int ⋮---- payload = hooklib.read stdin json transcript = payload.get "transcript path", "" or "" session id = payload.get "session id", "" or "" cwd = payload.get "cwd", "" or "" ⋮---- def run worker transcript: str, session id: str, cwd: str - None 证据：`komi/adapters/codex/hook_distill.py`
- **Hook Recall**（source_file）：def main - int ⋮---- payload = hooklib.read stdin json cwd = payload.get "cwd", "" or "" block = hooklib.build recall block paths, cwd=cwd ⋮---- def run sync - None ⋮---- """Detached pool-sync worker same as Claude Code's, Codex-rooted .""" ⋮---- cfg = paths.pool config 证据：`komi/adapters/codex/hook_recall.py`
- **Llm**（source_file）：DISTILL MODEL = os.environ.get "KOMI CODEX DISTILL MODEL", "gpt-5-mini" ⋮---- class NullLLM ⋮---- def complete self, , system: str, user: str - str ⋮---- def call self, learning: Learning, , context: dict - dict ⋮---- class OpenAILLM ⋮---- def init self, , model: str = DISTILL MODEL, api key: Optional str = None ⋮---- @property def available self - bool ⋮---- def test call self - tuple bool, str ⋮---- r = self. client.chat.completions.create text = r.choices 0 .message.content or "" .strip ⋮---- usr = json.dumps { ⋮---- text = r.choices 0 .message.content or "" ⋮---- def komi openai key - Optional str ⋮---- env path = paths.personal root / ".env" ⋮---- line = line.strip ⋮---- JUDGE SYSTEM =… 证据：`komi/adapters/codex/llm.py`
- **Config Io**（source_file）：def config path paths mod - Path ⋮---- def load raw paths mod - dict ⋮---- p = config path paths mod ⋮---- d = json.loads p.read text encoding="utf-8" ⋮---- def save raw paths mod, data: dict - bool ⋮---- POOL KEYS = {"repo url", "mode", "branch", "require signature", "sync hours", TOP KEYS = {"nudge turns", "recall k", "distill model"} ⋮---- RECALL KEYS = {"semantic"} ⋮---- def get key data: dict, dotted: str - Any ⋮---- def set key data: dict, dotted: str, value: Any - None ⋮---- value = coerce value ⋮---- def coerce v: Any - Any ⋮---- low = v.strip .lower ⋮---- def known keys - list str ⋮---- all = "config path", "load raw", "save raw", "get key", "set key", "known keys" 证据：`komi/adapters/config_io.py`
- **Config Schema**（source_file）：ENV MAP: dict str, str = { ⋮---- FILE KEYS: dict str, tuple = { ⋮---- def coerce value: str, like: Any - Any ⋮---- def apply file cfg: Any, data: dict - None ⋮---- node: Any = data ⋮---- node = None ⋮---- node = node.get key ⋮---- def apply env cfg: Any, environ: dict None = None - None ⋮---- env = environ if environ is not None else os.environ ⋮---- all = "ENV MAP", "FILE KEYS", "apply file", "apply env" 证据：`komi/adapters/config_schema.py`
- **── stdin/stdout plumbing ───────────────────────────────────────────────────**（source_file）：REPO ROOT = str Path file .resolve .parents 2 ⋮---- """Build the SessionStart context block from a host's stores. Host-neutral: opens the personal store, mirrors the project store + synced pool into the shared index, runs recall. Returns "" on any failure never break a session .""" ⋮---- personal = Store paths mod.personal root , index path=paths mod.index path proot = paths mod.project root cwd ⋮---- def apply semantic pref paths mod - None ⋮---- """Export the user's recall.semantic preference to KOMI SEMANTIC so the host-agnostic engine honors 'semantic off' even when the model is installed. A supported cross-module entry point the curate worker calls it too — it's a fresh process that di… 证据：`komi/adapters/hooklib.py`
- **Also keep a project/personal copy so the user benefits immediately.**（source_file）：PROMPT PATH = Path file .parent / "prompts" / "distill.md" ⋮---- MAX CANDIDATES PER PASS = 12 ⋮---- class LLMClient Protocol ⋮---- def complete self, , system: str, user: str - str: ... ⋮---- @dataclass class DistillResult ⋮---- candidates: int = 0 stored personal: int = 0 stored project: int = 0 queued global: int = 0 rejected: int = 0 learnings: list Learning = field default factory=list actions: list str = field default factory=list ⋮---- def summary self - str ⋮---- def parse transcript path: str Path - list dict ⋮---- p = Path path ⋮---- turns: list dict = ⋮---- line = line.strip ⋮---- obj = json.loads line ⋮---- role = obj.get "role" or obj.get "type" or "" text = flatten content obj.… 证据：`komi/engine/distill.py`
- **Embed**（source_file）：DEFAULT MODEL = os.environ.get "KOMI EMBED MODEL", "all-MiniLM-L6-v2" ⋮---- EMBED VERSION = "minilm-l6-v2/1" ⋮---- class Embedder Protocol ⋮---- version: str dim: int def encode self, text: str - list float : ... ⋮---- class SentenceTransformerEmbedder ⋮---- def init self, model name: str = DEFAULT MODEL ⋮---- def ensure self - bool ⋮---- getdim = getattr self. model, "get embedding dimension", None ⋮---- @property def dim self - int ⋮---- def encode self, text: str - list float ⋮---- vec = self. model.encode text or "", normalize embeddings=True ⋮---- cached: Optional object = None resolved = False ⋮---- def semantic enabled - bool ⋮---- """Respect a user 'turn semantic off' preference eve… 证据：`komi/engine/embed.py`
- **Transient trust signal for pool-sourced learnings: how many DISTINCT**（source_file）：SCHEMA = "komi.learning/1" ⋮---- class LearningType str, Enum ⋮---- IDENTITY = "identity" SEMANTIC = "semantic" PROCEDURAL = "procedural" ⋮---- class Scope str, Enum ⋮---- PERSONAL = "personal" PROJECT = "project" GLOBAL = "global" ⋮---- class Category str, Enum ⋮---- TOOLING = "tooling" WORKFLOW = "workflow" PREFERENCE = "preference" DOMAIN KNOWLEDGE = "domain-knowledge" PITFALL = "pitfall" DEBUGGING = "debugging" LANGUAGE BEHAVIOR = "language-behavior" FORMATTING STYLE = "formatting-style" META AGENT = "meta-agent" ENVIRONMENT = "environment" ⋮---- class Signal str, Enum ⋮---- USER CORRECTION = "user-correction" TECHNIQUE = "technique" FIX = "fix" REPEATED PATTERN = "repeated-pattern" DUR… 证据：`komi/engine/model.py`
- **Clamp defensively here too a stale/legacy index row could carry a larger value .**（source_file）：FRAME OPEN = FRAME CLOSE = " " ⋮---- COMMUNITY NOTE = ⋮---- @dataclass class RecallConfig ⋮---- k: int = 8 max identity: int = 6 max community: int = 3 max chars: int = 6000 include global: bool = True min confidence: float = 0.0 ⋮---- def recency score updated at: str, , half life days: float = 30.0 - float ⋮---- t = time.mktime time.strptime updated at, "%Y-%m-%dT%H:%M:%SZ" ⋮---- age days = max 0.0, time.time - t / 86400.0 ⋮---- def col row, key, default=0 ⋮---- val = row key ⋮---- def rank score row, similarity: float - float ⋮---- reuse = max 0, col row, "reused", 0 confidence = col row, "confidence", 0.0 or 0.0 salience = min 1.0, confidence 1.0 + 0.5 math.log1p reuse recency = recency… 证据：`komi/engine/recall.py`
- **Store**（source_file）：ENTRY DELIMITER = "\n§\n" FILE FOR TYPE = { ⋮---- class Store ⋮---- def init self, root: str Path, , index path: Optional str Path = None ⋮---- def md path self, learning type: str - Path ⋮---- fname = FILE FOR TYPE.get learning type ⋮---- def read entries self, path: Path - list dict ⋮---- """Each entry is a fenced JSON block between § delimiters. We store the full record as JSON inside the Markdown so the file is both human-readable title/body rendered and losslessly round-trippable. Format per entry: -- """ ⋮---- raw = path.read text encoding="utf-8" out: list dict = ⋮---- chunk = chunk.strip ⋮---- rec = extract json block chunk ⋮---- @staticmethod def md payload rec: dict - dict ⋮---- d… 证据：`komi/engine/store.py`
- **Whitelist of acceptable pool repo URLs. repo url can come from a config file or**（source_file）：@dataclass class PoolConfig ⋮---- repo url: str = "" https/ssh GitHub URL, or a local path/file:// for tests cache dir: str = "" local mirror, e.g. ~/.claude/komi/pool/repo branch: str = "main" mode: str = "pr" require signature: bool = True min corroboration: int = 1 author name: str = "komi-learn" author email: str = "komi-learn@users.noreply.github.com" ⋮---- @dataclass class GitResult ⋮---- ok: bool detail: str = "" extra: dict = field default factory=dict ⋮---- Whitelist of acceptable pool repo URLs. repo url can come from a config file or env var a trust boundary , so we validate it before it ever reaches git/gh — rejecting anything that isn't a clean GitHub https/ssh URL or a local p… 证据：`komi/pool/github_backend.py`
- **A pool learning is a short lesson; its verifiable JSON block is a few KB even with**（source_file）：LEARNINGS DIR = "learnings" FENCE = "komi" ⋮---- def id to filename learning id: str - str ⋮---- safe = learning id.replace ":", " " ⋮---- safe = re.sub r" ^A-Za-z0-9 .- ", "", safe ⋮---- def repo path for envelope: dict - str ⋮---- learning = envelope "learning" category = slug learning.get "category" or "uncategorized" ⋮---- def render md envelope: dict - str ⋮---- title = learning.get "title" or "" .strip body = learning.get "body" or "" .strip trigger = learning.get "trigger" or "" .strip tags = learning.get "tags" or ⋮---- sigs = envelope signatures envelope primary = sigs 0 if sigs else {"public key": " unsigned ", "algo": "unsigned"} ⋮---- signer line = f" Signer: {primary.get 'publi… 证据：`komi/pool/repo_format.py`
- **content integrity**（source_file）：def check file path: Path, , require signature: bool = True, repo root: Path None = None - list str ⋮---- problems: list str = text = path.read text encoding="utf-8", errors="replace" env = parse md text ⋮---- learning = env.get "learning", {} ⋮---- content integrity ⋮---- signature sig = learning.get "provenance", {} .get "signature" pk = env.get "signer", {} .get "public key", "" ⋮---- joined = " \n ".join learning.get "title", "" , learning.get "body", "" , floor = safety floor joined ⋮---- correct content-addressed location expected = repo path for env ⋮---- actual = path.relative to repo root .as posix ⋮---- def main argv: list str - int ⋮---- require sig = "--no-signature" not in argv… 证据：`komi/pool/verify_cli.py`
- **Prefer ripgrep rg over grep -r**（documentation）：Category: tooling Type: procedural Use when: searching a codebase for a string or pattern Tags: ripgrep, search, tooling Signer: ZlVX3/0m96T/4yIw… ed25519 ID: blake3:48f97281cbcb0559f1c6e5ad88116c8b9e5acb81403fde79031a6d899d2c3719 证据：`pool-repo-template/learnings/tooling/blake3_48f97281cbcb0559f1c6e5ad88116c8b9e5acb81403fde79031a6d899d2c3719.md`
- **Hooks**（structured_config）：{ "hooks": { "SessionStart": { "hooks": { "type": "command", "command": "python -m komi.adapters.claude code.hook recall" } } , "Stop": { "hooks": { "type": "command", "command": "python -m komi.adapters.claude code.hook distill" } } , "SubagentStop": { "hooks": { "type": "command", "command": "python -m komi.adapters.claude code.hook distill" } } , "PostCompact": { "hooks": { "type": "command", "command": "python -m komi.adapters.claude code.hook compact" } } } } 证据：`hooks/hooks.json`
- **Python**（source_file）：Python pycache / .py cod .egg-info/ .eggs/ build/ dist/ .venv/ venv/ 证据：`.gitignore`
- **Demo Codex Host**（source_file）：class ScriptedGPT ⋮---- def complete self, , system: str, user: str - str ⋮---- def call self, lng, , context ⋮---- def banner t: str - None ⋮---- def main - None ⋮---- model = ScriptedGPT cwd = os.getcwd ⋮---- turns = personal = Store paths.personal root , index path=paths.index path res = distill turns, personal store=personal, queue dir=paths.queue dir , ⋮---- block = CodexAdapter .recall RecallContext ⋮---- ok = "cargo test" in block.lower 证据：`examples/demo_codex_host.py`
- **--- the human gate: the user approves ---**（source_file）：class ScriptedModel ⋮---- def complete self, , system: str, user: str - str ⋮---- def call self, lng, , context ⋮---- def banner t: str - None ⋮---- def main - None ⋮---- workdir = Path tempfile.mkdtemp prefix="komi-demo-" personal = Store workdir / "komi" queue = workdir / "queue" model = ScriptedModel ⋮---- transcript = ⋮---- res = distill transcript, personal store=personal, queue dir=queue, ⋮---- rec = json.loads f.read text encoding="utf-8" prev = rec "publishable preview" ⋮---- block = recall personal, cwd=str workdir , ⋮---- def demo pool workdir: Path - None ⋮---- """Show the GitHub-backed pool: approve a queued candidate, publish it to a local git repo, then pull + verify it back r… 证据：`examples/demo_loop.py`
- **Model Install**（source_file）：def semantic available - bool ⋮---- def is installed - bool ⋮---- def install model , quiet: bool = False - tuple bool, str ⋮---- cmd = sys.executable, "-m", "pip", "install", "sentence-transformers =2.2" ⋮---- r = subprocess.run cmd, timeout=1200 ⋮---- all = "semantic available", "is installed", "install model" 证据：`komi/model_install.py`
- **This repo holds only learnings + docs + CI. Nothing else.**（source_file）：This repo holds only learnings + docs + CI. Nothing else. pycache / .py cod .DS Store /.seed-key/ 证据：`pool-repo-template/.gitignore`
- **All learning contributions require maintainer review before merge.**（source_file）：All learning contributions require maintainer review before merge. Replace with your team/handles when you create the repo. /learnings/ @kurikomi-labs /.github/ @kurikomi-labs 证据：`pool-repo-template/CODEOWNERS`
- **The engine runs with ZERO required dependencies stdlib only : blake2b stands in**（source_file）：build-system requires = "setuptools =68" build-backend = "setuptools.build meta" 证据：`pyproject.toml`

## 宿主 AI 必须遵守的规则

- **把本资产当作开工前上下文，而不是运行环境。**：AI Context Pack 只包含证据化项目理解，不包含目标项目的可执行状态。 证据：`README.md`, `pool-repo-template/README.md`, `pool-repo-template/CONTRIBUTING.md`
- **回答用户时区分可预览内容与必须安装后才能验证的内容。**：安装前体验的消费者价值来自降低误装和误判，而不是伪装成真实运行。 证据：`README.md`, `pool-repo-template/README.md`, `pool-repo-template/CONTRIBUTING.md`

## 用户开工前应该回答的问题

- 你准备在哪个宿主 AI 或本地环境中使用它？
- 你只是想先体验工作流，还是准备真实安装？
- 你最在意的是安装成本、输出质量、还是和现有规则的冲突？

## 验收标准

- 所有能力声明都能回指到 evidence_refs 中的文件路径。
- AI_CONTEXT_PACK.md 没有把预览包装成真实运行。
- 用户能在 3 分钟内看懂适合谁、能做什么、如何开始和风险边界。

---

## Doramagic Context Augmentation

下面内容用于强化 Repomix/AI Context Pack 主体。Human Manual 只提供阅读骨架；踩坑日志会被转成宿主 AI 必须遵守的工作约束。

## Human Manual 骨架

使用规则：这里只是项目阅读路线和显著性信号，不是事实权威。具体事实仍必须回到 repo evidence / Claim Graph。

宿主 AI 硬性规则：
- 不得把页标题、章节顺序、摘要或 importance 当作项目事实证据。
- 解释 Human Manual 骨架时，必须明确说它只是阅读路线/显著性信号。
- 能力、安装、兼容性、运行状态和风险判断必须引用 repo evidence、source path 或 Claim Graph。

- **Overview**：importance `high`
  - source_paths: README.md, komi/__init__.py, examples/demo_loop.py
- **Installation Guide**：importance `high`
  - source_paths: komi/cli.py, komi/adapters/claude_code/setup.py, komi/adapters/codex/setup.py, komi/wizard.py, pyproject.toml
- **System Architecture**：importance `high`
  - source_paths: komi/adapters/base.py, komi/adapters/hooklib.py, komi/engine/__init__.py, komi/engine/store.py, komi/engine/model.py
- **Host Adapters**：importance `high`
  - source_paths: komi/adapters/claude_code/__init__.py, komi/adapters/claude_code/paths.py, komi/adapters/claude_code/hook_recall.py, komi/adapters/claude_code/hook_distill.py, komi/adapters/codex/__init__.py
- **Core Engine Components**：importance `high`
  - source_paths: komi/engine/model.py, komi/engine/store.py, komi/engine/classify.py, komi/engine/embed.py
- **Recall System**：importance `high`
  - source_paths: komi/engine/recall.py, komi/engine/embed.py, komi/adapters/hooklib.py, komi/adapters/claude_code/hook_recall.py
- **Distillation Process**：importance `high`
  - source_paths: komi/engine/distill.py, komi/engine/prompts/distill.md, komi/adapters/claude_code/llm.py, komi/adapters/codex/llm.py, komi/adapters/hooklib.py
- **Curation and Learning Lifecycle**：importance `medium`
  - source_paths: komi/engine/curator.py, komi/engine/classify.py, komi/adapters/claude_code/curate.py

## Repo Inspection Evidence / 源码检查证据

- repo_clone_verified: true
- repo_inspection_verified: true
- repo_commit: `f088e43e9b43ceb303dd5dbb277fbeecc3cd90bd`
- inspected_files: `pyproject.toml`, `README.md`, `examples/demo_codex_host.py`, `examples/demo_loop.py`

宿主 AI 硬性规则：
- 没有 repo_clone_verified=true 时，不得声称已经读过源码。
- 没有 repo_inspection_verified=true 时，不得把 README/docs/package 文件判断写成事实。
- 没有 quick_start_verified=true 时，不得声称 Quick Start 已跑通。

## Doramagic Pitfall Constraints / 踩坑约束

这些规则来自 Doramagic 发现、验证或编译过程中的项目专属坑点。宿主 AI 必须把它们当作工作约束，而不是普通说明文字。

### Constraint 1: 可能修改宿主 AI 配置

- Trigger: 项目面向 Claude/Cursor/Codex/Gemini/OpenCode 等宿主，或安装命令涉及用户配置目录。
- Host AI rule: 列出会写入的配置文件、目录和卸载/回滚步骤。
- Why it matters: 安装可能改变本机 AI 工具行为，用户需要知道写入位置和回滚方法。
- Evidence: capability.host_targets | hn_item:48343216 | https://news.ycombinator.com/item?id=48343216 | host_targets=claude, claude_code
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 2: 能力判断依赖假设

- Trigger: README/documentation is current enough for a first validation pass.
- Host AI rule: 将假设转成下游验证清单。
- Why it matters: 假设不成立时，用户拿不到承诺的能力。
- Evidence: capability.assumptions | hn_item:48343216 | https://news.ycombinator.com/item?id=48343216 | README/documentation is current enough for a first validation pass.
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 3: 维护活跃度未知

- Trigger: 未记录 last_activity_observed。
- Host AI rule: 补 GitHub 最近 commit、release、issue/PR 响应信号。
- Why it matters: 新项目、停更项目和活跃项目会被混在一起，推荐信任度下降。
- Evidence: evidence.maintainer_signals | hn_item:48343216 | https://news.ycombinator.com/item?id=48343216 | last_activity_observed missing
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 4: 下游验证发现风险项

- Trigger: no_demo
- Host AI rule: 进入安全/权限治理复核队列。
- Why it matters: 下游已经要求复核，不能在页面中弱化。
- Evidence: downstream_validation.risk_items | hn_item:48343216 | https://news.ycombinator.com/item?id=48343216 | no_demo; severity=medium
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 5: 存在评分风险

- Trigger: no_demo
- Host AI rule: 把风险写入边界卡，并确认是否需要人工复核。
- Why it matters: 风险会影响是否适合普通用户安装。
- Evidence: risks.scoring_risks | hn_item:48343216 | https://news.ycombinator.com/item?id=48343216 | no_demo; severity=medium
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 6: issue/PR 响应质量未知

- Trigger: issue_or_pr_quality=unknown。
- Host AI rule: 抽样最近 issue/PR，判断是否长期无人处理。
- Why it matters: 用户无法判断遇到问题后是否有人维护。
- Evidence: evidence.maintainer_signals | hn_item:48343216 | https://news.ycombinator.com/item?id=48343216 | issue_or_pr_quality=unknown
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。

### Constraint 7: 发布节奏不明确

- Trigger: release_recency=unknown。
- Host AI rule: 确认最近 release/tag 和 README 安装命令是否一致。
- Why it matters: 安装命令和文档可能落后于代码，用户踩坑概率升高。
- Evidence: evidence.maintainer_signals | hn_item:48343216 | https://news.ycombinator.com/item?id=48343216 | release_recency=unknown
- Hard boundary: 不要把这个坑点包装成已解决、已验证或可忽略，除非后续验证证据明确证明它已经关闭。