# headroom-docs - 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 headroom-docs 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_0017` supported 0.86

## 它能做什么

- **Context Compression Library**（可做安装前预览）：Inlines compression into any Python or TypeScript application via `compress(messages)` API, reducing context by 60-95%. 证据：`README.md`, `sdk/typescript/README.md`, `sdk/typescript/package.json`, `examples/basic_usage.py` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Local Proxy Server**（需要安装后验证）：Zero-code-change proxy that intercepts LLM traffic, compresses context, and forwards to upstream providers on port 8787. 证据：`README.md`, `examples/deployment/macos-launchagent/README.md`, `docs/observability.md`, `docs/auth-modes.md` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Agent Wrapping CLI**（可做安装前预览）：One-command wrapper for Claude Code, Codex, Cursor, Aider, and Copilot that automatically routes traffic through Headroom. 证据：`README.md`, `plugins/headroom-agent-hooks/README.md` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **MCP Server**（需要安装后验证）：Model Context Protocol server exposing headroom_compress, headroom_retrieve, and headroom_stats tools for any MCP-compatible client. 证据：`README.md`, `examples/mcp_demo/` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Cross-Agent Memory**（需要安装后验证）：Shared compressed context store across Claude, Codex, and Gemini with automatic deduplication. 证据：`README.md`, `sdk/typescript/README.md`, `sdk/typescript/examples/shared-context-multi-agent.ts` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **headroom learn**（需要安装后验证）：Mines failed agent sessions and writes corrections to CLAUDE.md or AGENTS.md for future reference. 证据：`README.md` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **CCR (Context Compression Recovery)**（需要安装后验证）：Reversible compression storing originals locally; LLM retrieves via headroom_retrieve tool on demand. Never deletes originals. 证据：`README.md`, `headroom/evals/README.md`, `sdk/typescript/examples/ccr-retrieve.ts` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Framework Adapters (Vercel AI, OpenAI, Anthropic, Gemini, LangChain, Strands)**（可做安装前预览）：Drop-in adapters for major LLM frameworks and agent libraries enabling zero-code-change compression. 证据：`sdk/typescript/README.md`, `sdk/typescript/package.json`, `examples/langchain_demo/README.md`, `examples/strands_bedrock_demo.py` Claim：`clm_0001` supported 0.86, `clm_0005` supported 0.86, `clm_0008` supported 0.86, `clm_0015` supported 0.86 等
- **Evaluation Framework**（需要安装后验证）：Rigorous OSS benchmark suite proving compression preserves LLM accuracy with Tier 1/2/3 tests. 证据：`headroom/evals/README.md`, `headroom/evals/README.md` Claim：`clm_0007` supported 0.86, `clm_0009` supported 0.86
- **AWS Bedrock Native Integration**（需要安装后验证）：Native Rust handler for Bedrock InvokeModel with SigV4 signing and Anthropic compression support. 证据：`docs/bedrock.md`, `docs/bedrock.md`, `docs/bedrock.md` Claim：`clm_0010` supported 0.86
- **Auth Mode Classification**（可做安装前预览）：Three-mode request classifier (Payg, OAuth, Subscription) driving compression policy decisions. 证据：`docs/auth-modes.md`, `docs/auth-modes.md`, `docs/auth-modes.md`, `docs/auth-modes.md` Claim：`clm_0002` supported 0.86, `clm_0011` supported 0.86
- **Observability (Prometheus Metrics)**（需要安装后验证）：Full Prometheus metric catalogue covering compression ratios, cache hit rates, rate limits, and Bedrock telemetry. 证据：`docs/observability.md`, `docs/observability.md`, `docs/observability.md` Claim：`clm_0002` supported 0.86, `clm_0012` supported 0.86
- **OpenClaw Plugin**（需要安装后验证）：NPM package providing Headroom context compression plugin for OpenClaw with gateway provider routing. 证据：`plugins/openclaw/README.md`, `plugins/openclaw/package.json`, `plugins/openclaw/README.md` Claim：`clm_0013` supported 0.86
- **Claude Code/Copilot Hooks**（可做安装前预览）：Lightweight startup hooks for Claude Code and GitHub Copilot CLI that ensure Headroom proxy is running. 证据：`plugins/headroom-agent-hooks/README.md`, `plugins/headroom-agent-hooks/.claude-plugin/plugin.json` Claim：`clm_0003` supported 0.86, `clm_0014` supported 0.86
- **Compression Simulation (Dry Run)**（可做安装前预览）：Preview compression effects without calling the LLM — shows transforms, waste signals, and cache alignment. 证据：`sdk/typescript/README.md`, `sdk/typescript/examples/simulation-dry-run.ts` Claim：`clm_0001` supported 0.86, `clm_0005` supported 0.86, `clm_0008` supported 0.86, `clm_0015` supported 0.86 等
- **Compression Hooks**（可做安装前预览）：Pre/post compression customization with per-message importance biases and result observation. 证据：`sdk/typescript/README.md`, `sdk/typescript/examples/hooks-custom-compression.ts` Claim：`clm_0001` supported 0.86, `clm_0005` supported 0.86, `clm_0008` supported 0.86, `clm_0015` supported 0.86 等

## 怎么开始

- `pip install "headroom-ai[all]"          # Python` 证据：`README.md` Claim：`clm_0018` supported 0.86, `clm_0020` supported 0.86
- `npm install headroom-ai                 # Node / TypeScript` 证据：`README.md` Claim：`clm_0019` supported 0.86
- `pip install "headroom-ai[all]"          # Python, everything` 证据：`README.md` Claim：`clm_0020` supported 0.86
- `npm install headroom-ai                 # TypeScript / Node` 证据：`README.md` Claim：`clm_0021` supported 0.86
- `pipx install --python python3.13 "headroom-ai[all]"` 证据：`README.md` Claim：`clm_0022` supported 0.86
- `git clone https://github.com/chopratejas/headroom.git && cd headroom` 证据：`README.md` Claim：`clm_0023` supported 0.86
- `pip install -e ".[dev]" && pytest` 证据：`README.md` Claim：`clm_0024` supported 0.86
- `pip install headroom-ai` 证据：`wiki/quickstart.md` Claim：`clm_0018` supported 0.86, `clm_0020` supported 0.86, `clm_0025` unverified 0.25, `clm_0026` unverified 0.25 等
- `pip install "headroom-ai[proxy]"` 证据：`wiki/quickstart.md` Claim：`clm_0026` unverified 0.25
- `pip install "headroom-ai[all]"` 证据：`wiki/quickstart.md` Claim：`clm_0018` supported 0.86, `clm_0020` supported 0.86, `clm_0027` unverified 0.25

## 继续前判断卡

- **当前建议**：先做角色匹配试用
- **为什么**：这个项目更像角色库，核心风险是选错角色或把角色文案当执行能力；先用 Prompt Preview 试角色匹配，再决定是否沙盒导入。

### 30 秒判断

- **现在怎么做**：先做角色匹配试用
- **最小安全下一步**：先用 Prompt Preview 试角色匹配；满意后再隔离导入
- **先别相信**：角色质量和任务匹配不能直接相信。
- **继续会触碰**：角色选择偏差、命令执行、宿主 AI 配置

### 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_0017` supported 0.86
- **能力存在：Context Compression Library**（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`, `sdk/typescript/README.md`, `sdk/typescript/package.json`, `examples/basic_usage.py` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86
- **能力存在：Local Proxy Server**（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`, `examples/deployment/macos-launchagent/README.md`, `docs/observability.md`, `docs/auth-modes.md` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86
- **能力存在：Agent Wrapping CLI**（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`, `plugins/headroom-agent-hooks/README.md` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86
- **能力存在：MCP Server**（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`, `examples/mcp_demo/` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86
- **能力存在：Cross-Agent Memory**（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`, `sdk/typescript/README.md`, `sdk/typescript/examples/shared-context-multi-agent.ts` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86

### 现在还不能相信

- **角色质量和任务匹配不能直接相信。**（unverified）：角色库证明有很多角色，不证明每个角色都适合你的具体任务，也不证明角色能产生高质量结果。
- **不能把角色文案当成真实执行能力。**（unverified）：安装前只能判断角色描述和任务画像是否匹配，不能证明它能在宿主 AI 里完成任务。
- **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/marketplace.json`, `plugins/headroom-agent-hooks/.claude-plugin/plugin.json`, `plugins/headroom-agent-hooks/.github/plugin/plugin.json`, `plugins/openclaw/openclaw.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/marketplace.json`, `plugins/headroom-agent-hooks/.claude-plugin/plugin.json`, `plugins/headroom-agent-hooks/.github/plugin/plugin.json`, `plugins/openclaw/openclaw.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。

### 继续会触碰什么

- **角色选择偏差**：用户对任务应该由哪个专家角色处理的判断。 原因：选错角色会让 AI 从错误专业视角回答，浪费时间或误导决策。
- **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`, `wiki/quickstart.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/marketplace.json`, `plugins/headroom-agent-hooks/.claude-plugin/plugin.json`, `plugins/headroom-agent-hooks/.github/plugin/plugin.json`, `plugins/openclaw/openclaw.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：`README.md`, `docs/auth-modes.md`, `docs/bedrock.md`, `docs/observability.md` 等
- **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**：先用交互式试用验证任务画像和角色匹配，不要先导入整套角色库。（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。
- **保留原始角色选择记录**：如果输出偏题，可以回到任务画像阶段重新选择角色，而不是继续沿着错误角色推进。
- **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_0032` 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/marketplace.json`, `plugins/headroom-agent-hooks/.claude-plugin/plugin.json`, `plugins/headroom-agent-hooks/.github/plugin/plugin.json`, `plugins/openclaw/openclaw.plugin.json` Claim：`clm_0033` 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`, `wiki/quickstart.md` Claim：`clm_0034` supported 0.86
- **风险**： 处理方式：
- **风险**： 处理方式：
- **风险**： 处理方式：
- **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

- **Context Compression Library**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`README.md`, `sdk/typescript/README.md`, `sdk/typescript/package.json`, `examples/basic_usage.py` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Local Proxy Server**：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`, `examples/deployment/macos-launchagent/README.md`, `docs/observability.md`, `docs/auth-modes.md` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Agent Wrapping CLI**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`README.md`, `plugins/headroom-agent-hooks/README.md` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **MCP Server**：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`, `examples/mcp_demo/` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Cross-Agent Memory**：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`, `sdk/typescript/README.md`, `sdk/typescript/examples/shared-context-multi-agent.ts` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **headroom learn**：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_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **CCR (Context Compression Recovery)**：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`, `headroom/evals/README.md`, `sdk/typescript/examples/ccr-retrieve.ts` Claim：`clm_0001` supported 0.86, `clm_0002` supported 0.86, `clm_0003` supported 0.86, `clm_0004` supported 0.86 等
- **Framework Adapters (Vercel AI, OpenAI, Anthropic, Gemini, LangChain, Strands)**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`sdk/typescript/README.md`, `sdk/typescript/package.json`, `examples/langchain_demo/README.md`, `examples/strands_bedrock_demo.py` Claim：`clm_0001` supported 0.86, `clm_0005` supported 0.86, `clm_0008` supported 0.86, `clm_0015` supported 0.86 等
- **Evaluation Framework**：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：`headroom/evals/README.md`, `headroom/evals/README.md` Claim：`clm_0007` supported 0.86, `clm_0009` supported 0.86
- **AWS Bedrock Native Integration**：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：`docs/bedrock.md`, `docs/bedrock.md`, `docs/bedrock.md` Claim：`clm_0010` supported 0.86
- **Auth Mode Classification**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`docs/auth-modes.md`, `docs/auth-modes.md`, `docs/auth-modes.md`, `docs/auth-modes.md` Claim：`clm_0002` supported 0.86, `clm_0011` supported 0.86
- **Observability (Prometheus Metrics)**：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：`docs/observability.md`, `docs/observability.md`, `docs/observability.md` Claim：`clm_0002` supported 0.86, `clm_0012` supported 0.86
- **OpenClaw Plugin**：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：`plugins/openclaw/README.md`, `plugins/openclaw/package.json`, `plugins/openclaw/README.md` Claim：`clm_0013` supported 0.86
- **Claude Code/Copilot Hooks**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`plugins/headroom-agent-hooks/README.md`, `plugins/headroom-agent-hooks/.claude-plugin/plugin.json` Claim：`clm_0003` supported 0.86, `clm_0014` supported 0.86
- **Compression Simulation (Dry Run)**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`sdk/typescript/README.md`, `sdk/typescript/examples/simulation-dry-run.ts` Claim：`clm_0001` supported 0.86, `clm_0005` supported 0.86, `clm_0008` supported 0.86, `clm_0015` supported 0.86 等
- **Compression Hooks**：先基于 role_skill_index / evidence_index 帮用户挑选可用角色、Skill 或工作流。 边界：可做安装前 Prompt 体验。 证据：`sdk/typescript/README.md`, `sdk/typescript/examples/hooks-custom-compression.ts` Claim：`clm_0001` supported 0.86, `clm_0005` supported 0.86, `clm_0008` supported 0.86, `clm_0015` supported 0.86 等

### 上下文规模

- 文件总数：886
- 重要文件覆盖：40/886
- 证据索引条目：80
- 角色 / Skill 条目：40

### 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 headroom-docs 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 headroom-docs 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 headroom-docs 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 40 个角色 / Skill / 项目文档条目。

- **docs**（project_doc）：This is a Next.js application generated with Create Fumadocs https://github.com/fuma-nama/fumadocs . 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：`docs/README.md`
- **What it does**（project_doc）：60–95% fewer tokens · library · proxy · MCP · 6 algorithms · local-first · reversible 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`
- **Headroom Examples**（project_doc）：This directory contains examples demonstrating Headroom's capabilities. 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：`examples/README.md`
- **scripts/**（project_doc）：Utility scripts bundled with the Headroom repo. Most are one-off operator tools; a few are runnable as part of development workflows. 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：`scripts/README.md`
- **signals/ — detection traits**（project_doc）：Cross-cutting classifiers used by transforms. Lives at the crate root because the same classifier feeds many transforms; nesting under transforms/ would imply ownership by one consumer. 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：`crates/headroom-core/src/signals/README.md`
- **macOS LaunchAgent Deployment**（project_doc）：This directory contains templates and scripts for running the headroom proxy server as a persistent background service on macOS using LaunchAgent. 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：`examples/deployment/macos-launchagent/README.md`
- **LangChain + Headroom Demo**（project_doc）：Real-world demonstration of Headroom optimization on LangChain agents. 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：`examples/langchain_demo/README.md`
- **Headroom Evaluation Framework**（project_doc）：Prove that compression preserves LLM accuracy through rigorous OSS benchmarks. 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：`headroom/evals/README.md`
- **Headroom agent hooks**（project_doc）：This plugin exposes lightweight startup hooks for Claude Code and GitHub Copilot CLI. 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：`plugins/headroom-agent-hooks/README.md`
- **@headroom-ai/openclaw**（project_doc）：Context compression plugin for OpenClaw https://github.com/openclaw/openclaw . Compresses tool outputs, code, logs, and structured data — 70-90% token savings with zero LLM calls. 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：`plugins/openclaw/README.md`
- **headroom-ai**（project_doc）：Compress LLM context. Save tokens. Fit more into every request. 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：`sdk/typescript/README.md`
- **headroom-ai examples**（project_doc）：Runnable examples showing how to use the Headroom TypeScript SDK for context compression. 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：`sdk/typescript/examples/README.md`
- **Contributing to Headroom**（project_doc）：Thank you for your interest in contributing to Headroom! This document provides guidelines and instructions for contributing. 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：`CONTRIBUTING.md`
- **Auth Modes**（project_doc）：Headroom classifies every inbound request into one of three auth modes at request entry. The mode drives every downstream compression, cache, and header policy decision. Detection is a pure function of HTTP headers — no I/O, no allocation beyond a single to lowercase of the User-Agent, runs in <10us per call. 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：`docs/auth-modes.md`
- **AWS Bedrock — Operator Guide**（project_doc）：Headroom's Rust proxy ships a native AWS Bedrock InvokeModel surface. After Phase D PRs D1–D3 , Anthropic-on-Bedrock requests are signed, compressed, and observed by the proxy directly — no LiteLLM Python shim on the request path. 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：`docs/bedrock.md`
- **Observability — proxy metrics**（project_doc）：The Headroom Rust proxy exposes Prometheus-format metrics on the /metrics endpoint of every running proxy instance. The metric catalogue below covers Phase D Bedrock route instrumentation and Phase G PR-G3 per-invocation RTK + proxy-wide observability . 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：`docs/observability.md`
- **RTK architecture — why wrap-CLI only**（project_doc）：RTK architecture — why wrap-CLI only 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：`docs/rtk-architecture.md`
- **001. Vision**（project_doc）：Headroom is a context compression proxy for AI provider APIs. It sits between AI coding tools Claude Code, Copilot, Codex, etc. and provider APIs OpenAI, Anthropic, Google, Cohere , compressing context before it reaches the provider to reduce token usage and costs. 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：`docs/spec/001-vision.md`
- **002. Architecture**（project_doc）：Headroom is a context compression proxy for LLM applications, featuring intelligent transforms, semantic caching, and CCR Compress-Cache-Retrieve architecture. 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：`docs/spec/002-architecture.md`
- **003. Architecture Decision Records**（project_doc）：ADR-001: Why a Proxy Instead of SDK Injection 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：`docs/spec/003-adrs.md`
- **004. Domain Model**（project_doc）：A session represents a conversation context between a user and an AI agent. 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：`docs/spec/004-domain-model.md`
- **005. Integrations**（project_doc）：Claude headroom/learn/plugins/claude/ 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：`docs/spec/005-integrations.md`
- **006. Actors**（project_doc）：The developer using an AI coding agent with Headroom. 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：`docs/spec/006-actors.md`
- **007. Behavior**（project_doc）：Headroom forwards requests without modification. 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：`docs/spec/007-behavior.md`
- **008. Capabilities**（project_doc）：Capability Proxy SDK Wrap MCP ASGI LiteLLM ------------ :-----: :---: :----: :---: :----: :-------: Semantic compression ✓ ✓ ✓ ✓ ✓ ✓ Summary compression ✓ ✓ ✓ ✓ ✓ ✓ Token budget management ✓ ✓ ✓ ✓ ✓ ✓ Learn mode ✓ ✓ ✓ ✓ ✓ ✓ Session tracking ✓ ✓ ✓ ✓ ✓ ✓ CCR feedback ✓ ✓ ✓ ✓ ✓ ✓ TOIN tagging ✓ ✓ ✓ ✓ ✓ ✓ Savings tracking ✓ ✓ ✓ ✓ ✓ ✓ Dashboard ✓ - ✓ - - - Health endpoints ✓ - - - - - Metrics Prometheus ✓ - - - - - 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：`docs/spec/008-capabilities.md`
- **009. Compliance**（project_doc）：Data Type Stored Exported Logged ----------- :------: :--------: :------: Prompts Optional cache Never Never Responses Optional cache Never Never Savings metrics Yes Never Never Session metadata Yes Never Never Telemetry Aggregated Never Never 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：`docs/spec/009-compliance.md`
- **010. Data**（project_doc）：Store Location Type Purpose ------- ---------- ------ --------- SQLite headroom.db cwd or ~/.headroom/ Local file Metrics, sessions, compression store Memory DB ~/.headroom/headroom memory.db Local file Memory system optional Vector Index ~/.headroom/headroom memory vectors.db Local file Semantic search optional, requires sqlite-vec Graph Store ~/.headroom/headroom memory graph.db Local file Memory relationships Com… 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：`docs/spec/010-data.md`
- **011. Deployment**（project_doc）：Dockerfile: dockerfile FROM python:3.12-slim 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：`docs/spec/011-deployment.md`
- **012. Diagrams**（project_doc）：Version Date Changes --------- ------ --------- 1.0.0-draft 2026-04-16 Initial diagrams document 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：`docs/spec/012-diagrams.md`
- **013. Disaster Recovery**（project_doc）：Failure Impact Recovery --------- -------- ---------- Proxy crash Service unavailable Restart or failover OOM Service unavailable Restart with more memory Network partition Cannot reach providers Retry with backoff 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：`docs/spec/013-disaster-recovery.md`
- **014. Governance**（project_doc）：- Repository: github.com/JerrettDavis/headroom - License: Apache 2.0 - Main Languages: Python core , TypeScript SDK/dashboard 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：`docs/spec/014-governance.md`
- **015. Interfaces**（project_doc）：Options: Flag Default Description ------ --------- ------------- --host 127.0.0.1 Bind host --port 8787 Bind port --mode token Optimization mode: token or cache --workers 1 Uvicorn worker processes --limit-concurrency 1000 Maximum concurrent connections before 503 --no-optimize false Passthrough mode --no-cache false Disable semantic cache --no-rate-limit false Disable rate limiting --memory false Enable persistent… 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：`docs/spec/015-interfaces.md`
- **016. Observability**（project_doc）：Headroom exposes Prometheus metrics at /metrics . 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：`docs/spec/016-observability.md`
- **017. Operations**（project_doc）：Basic health check. Returns 200 if process is running. 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：`docs/spec/017-operations.md`
- **018. Policies**（project_doc）：Setting Default Description --------- --------- ------------- Compression enabled true Apply compression by default Cache enabled true Use semantic cache Summary enabled true Use summary compression Token budget enforced false No budget limit by default 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：`docs/spec/018-policies.md`
- **019. Quality**（project_doc）：Surface Unit Integration E2E --------- :----: :------------: :---: Proxy Server ✓ ✓ ✓ SDK ✓ ✓ - Compression ✓ ✓ ✓ Cache ✓ ✓ - Learn ✓ ✓ - CCR ✓ ✓ ✓ TOIN ✓ ✓ - Dashboard ✓ - ✓ 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：`docs/spec/019-quality.md`
- **020. Security**（project_doc）：Threat Impact Mitigation -------- -------- ------------ Prompt data exfiltration High No logging, local-only API key theft High Key rotation, secrets management Cache poisoning Medium Input validation DoS via large prompts Medium Token limits SSRF via redirects Medium URL validation 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：`docs/spec/020-security.md`
- **021. Testing**（project_doc）：Unit Tests: - Request routing - Response header injection - Mode switching - Error handling 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：`docs/spec/021-testing.md`
- **022. Rust Migration**（project_doc）：Status: in-progress Stage 0 complete 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：`docs/spec/022-rust-migration.md`
- **Headroom Living Specification**（project_doc）：Version: 1.0.0-draft Date: 2026-04-16 Status: Draft — In Progress Related Issue: GitHub 183 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：`docs/spec/SPEC.md`

## 证据索引

- 共索引 80 条证据。

- **docs**（documentation）：This is a Next.js application generated with Create Fumadocs https://github.com/fuma-nama/fumadocs . 证据：`docs/README.md`
- **What it does**（documentation）：60–95% fewer tokens · library · proxy · MCP · 6 algorithms · local-first · reversible 证据：`README.md`
- **Headroom Examples**（documentation）：This directory contains examples demonstrating Headroom's capabilities. 证据：`examples/README.md`
- **scripts/**（documentation）：Utility scripts bundled with the Headroom repo. Most are one-off operator tools; a few are runnable as part of development workflows. 证据：`scripts/README.md`
- **signals/ — detection traits**（documentation）：Cross-cutting classifiers used by transforms. Lives at the crate root because the same classifier feeds many transforms; nesting under transforms/ would imply ownership by one consumer. 证据：`crates/headroom-core/src/signals/README.md`
- **macOS LaunchAgent Deployment**（documentation）：This directory contains templates and scripts for running the headroom proxy server as a persistent background service on macOS using LaunchAgent. 证据：`examples/deployment/macos-launchagent/README.md`
- **LangChain + Headroom Demo**（documentation）：Real-world demonstration of Headroom optimization on LangChain agents. 证据：`examples/langchain_demo/README.md`
- **Headroom Evaluation Framework**（documentation）：Prove that compression preserves LLM accuracy through rigorous OSS benchmarks. 证据：`headroom/evals/README.md`
- **Headroom agent hooks**（documentation）：This plugin exposes lightweight startup hooks for Claude Code and GitHub Copilot CLI. 证据：`plugins/headroom-agent-hooks/README.md`
- **@headroom-ai/openclaw**（documentation）：Context compression plugin for OpenClaw https://github.com/openclaw/openclaw . Compresses tool outputs, code, logs, and structured data — 70-90% token savings with zero LLM calls. 证据：`plugins/openclaw/README.md`
- **headroom-ai**（documentation）：Compress LLM context. Save tokens. Fit more into every request. 证据：`sdk/typescript/README.md`
- **headroom-ai examples**（documentation）：Runnable examples showing how to use the Headroom TypeScript SDK for context compression. 证据：`sdk/typescript/examples/README.md`
- **Contributing to Headroom**（documentation）：Thank you for your interest in contributing to Headroom! This document provides guidelines and instructions for contributing. 证据：`CONTRIBUTING.md`
- **Package**（package_manifest）：{ "name": "headroom-docs", "version": "0.0.0", "private": true, "scripts": { "build": "next build", "dev": "next dev", "start": "next start", "types:check": "fumadocs-mdx && next typegen && tsc --noEmit", "postinstall": "fumadocs-mdx" }, "dependencies": { "dotted-map": "^3.1.0", "fumadocs-core": "16.7.10", "fumadocs-mdx": "14.2.11", "fumadocs-twoslash": "^3.1.3", "fumadocs-typescript": "^4.0.3", "fumadocs-ui": "16.7.10", "headroom-ai": "file:../sdk/typescript", "lucide-react": "^1.7.0", "next": "16.2.4", "react": "^19.2.4", "react-dom": "^19.2.4", "recharts": "^3.8.1", "tailwind-merge": "^3.5.0" }, "devDependencies": { "@ai-sdk/openai": "^3.0.51", "@anthropic-ai/sdk": "^0.82.0", "@tailwindc… 证据：`docs/package.json`
- **Package**（package_manifest）：{ "name": "headroom-openclaw", "version": "0.22.4", "description": "Headroom context compression plugin for OpenClaw — 70-90% token savings with zero LLM calls", "type": "module", "main": "./dist/index.js", "types": "./dist/index.d.ts", "files": "dist", "hook-shim", "openclaw.plugin.json", "README.md" , "scripts": { "build": "tsup && node prepare-dist.mjs", "test": "vitest run", "test:watch": "vitest", "typecheck": "tsc --noEmit" }, "dependencies": { "headroom-ai": "^0.22.3" }, "peerDependencies": { "openclaw": " " }, "peerDependenciesMeta": { "openclaw": { "optional": true } }, "devDependencies": { "@types/node": "^22.10.0", "tsup": "^8.0.0", "typescript": "^5.5.0", "vitest": "^4.1.5" }, "… 证据：`plugins/openclaw/package.json`
- **Package**（package_manifest）：{ "name": "headroom-ai", "version": "0.22.4", "description": "Compress LLM context. Save tokens. Fit more into every request.", "type": "module", "main": "./dist/index.cjs", "module": "./dist/index.js", "types": "./dist/index.d.ts", "files": "dist", "README.md", "LICENSE" , "exports": { ".": { "import": { "types": "./dist/index.d.ts", "default": "./dist/index.js" }, "require": { "types": "./dist/index.d.cts", "default": "./dist/index.cjs" } }, "./vercel-ai": { "import": { "types": "./dist/adapters/vercel-ai.d.ts", "default": "./dist/adapters/vercel-ai.js" }, "require": { "types": "./dist/adapters/vercel-ai.d.cts", "default": "./dist/adapters/vercel-ai.cjs" } }, "./openai": { "import": { "ty… 证据：`sdk/typescript/package.json`
- **Marketplace**（structured_config）：{ "name": "headroom-marketplace", "owner": { "name": "Headroom Contributors" }, "metadata": { "description": "Headroom marketplace for Claude Code and GitHub Copilot CLI plugins.", "version": "0.22.3" }, "plugins": { "name": "headroom", "source": "./plugins/headroom-agent-hooks", "description": "Headroom startup hooks for Claude Code and GitHub Copilot CLI.", "version": "0.22.3", "author": { "name": "Headroom Contributors", "url": "https://github.com/chopratejas/headroom" }, "homepage": "https://github.com/chopratejas/headroom", "repository": "https://github.com/chopratejas/headroom", "keywords": "headroom", "hooks", "claude-code", "copilot-cli" } } 证据：`.claude-plugin/marketplace.json`
- **Plugin**（structured_config）：{ "name": "headroom", "version": "0.22.3", "description": "Headroom startup hooks for Claude Code and GitHub Copilot CLI.", "author": { "name": "Headroom Contributors", "url": "https://github.com/chopratejas/headroom" }, "homepage": "https://github.com/chopratejas/headroom", "repository": "https://github.com/chopratejas/headroom", "keywords": "headroom", "hooks", "claude-code", "copilot-cli" } 证据：`plugins/headroom-agent-hooks/.claude-plugin/plugin.json`
- **License**（source_file）：Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ 证据：`LICENSE`
- **Auth Modes**（documentation）：Headroom classifies every inbound request into one of three auth modes at request entry. The mode drives every downstream compression, cache, and header policy decision. Detection is a pure function of HTTP headers — no I/O, no allocation beyond a single to lowercase of the User-Agent, runs in <10us per call. 证据：`docs/auth-modes.md`
- **AWS Bedrock — Operator Guide**（documentation）：Headroom's Rust proxy ships a native AWS Bedrock InvokeModel surface. After Phase D PRs D1–D3 , Anthropic-on-Bedrock requests are signed, compressed, and observed by the proxy directly — no LiteLLM Python shim on the request path. 证据：`docs/bedrock.md`
- **Observability — proxy metrics**（documentation）：The Headroom Rust proxy exposes Prometheus-format metrics on the /metrics endpoint of every running proxy instance. The metric catalogue below covers Phase D Bedrock route instrumentation and Phase G PR-G3 per-invocation RTK + proxy-wide observability . 证据：`docs/observability.md`
- **RTK architecture — why wrap-CLI only**（documentation）：RTK architecture — why wrap-CLI only 证据：`docs/rtk-architecture.md`
- **Plugin**（structured_config）：{ "name": "headroom", "version": "0.22.3", "description": "Headroom startup hooks for Claude Code and GitHub Copilot CLI.", "author": { "name": "Headroom Contributors", "url": "https://github.com/chopratejas/headroom" }, "homepage": "https://github.com/chopratejas/headroom", "repository": "https://github.com/chopratejas/headroom", "keywords": "headroom", "hooks", "claude-code", "copilot-cli" , "hooks": "./hooks" } 证据：`plugins/headroom-agent-hooks/.github/plugin/plugin.json`
- **001. Vision**（documentation）：Headroom is a context compression proxy for AI provider APIs. It sits between AI coding tools Claude Code, Copilot, Codex, etc. and provider APIs OpenAI, Anthropic, Google, Cohere , compressing context before it reaches the provider to reduce token usage and costs. 证据：`docs/spec/001-vision.md`
- **002. Architecture**（documentation）：Headroom is a context compression proxy for LLM applications, featuring intelligent transforms, semantic caching, and CCR Compress-Cache-Retrieve architecture. 证据：`docs/spec/002-architecture.md`
- **003. Architecture Decision Records**（documentation）：ADR-001: Why a Proxy Instead of SDK Injection 证据：`docs/spec/003-adrs.md`
- **004. Domain Model**（documentation）：A session represents a conversation context between a user and an AI agent. 证据：`docs/spec/004-domain-model.md`
- **005. Integrations**（documentation）：Claude headroom/learn/plugins/claude/ 证据：`docs/spec/005-integrations.md`
- **006. Actors**（documentation）：The developer using an AI coding agent with Headroom. 证据：`docs/spec/006-actors.md`
- **007. Behavior**（documentation）：Headroom forwards requests without modification. 证据：`docs/spec/007-behavior.md`
- **008. Capabilities**（documentation）：Capability Proxy SDK Wrap MCP ASGI LiteLLM ------------ :-----: :---: :----: :---: :----: :-------: Semantic compression ✓ ✓ ✓ ✓ ✓ ✓ Summary compression ✓ ✓ ✓ ✓ ✓ ✓ Token budget management ✓ ✓ ✓ ✓ ✓ ✓ Learn mode ✓ ✓ ✓ ✓ ✓ ✓ Session tracking ✓ ✓ ✓ ✓ ✓ ✓ CCR feedback ✓ ✓ ✓ ✓ ✓ ✓ TOIN tagging ✓ ✓ ✓ ✓ ✓ ✓ Savings tracking ✓ ✓ ✓ ✓ ✓ ✓ Dashboard ✓ - ✓ - - - Health endpoints ✓ - - - - - Metrics Prometheus ✓ - - - - - 证据：`docs/spec/008-capabilities.md`
- **009. Compliance**（documentation）：Data Type Stored Exported Logged ----------- :------: :--------: :------: Prompts Optional cache Never Never Responses Optional cache Never Never Savings metrics Yes Never Never Session metadata Yes Never Never Telemetry Aggregated Never Never 证据：`docs/spec/009-compliance.md`
- **010. Data**（documentation）：Store Location Type Purpose ------- ---------- ------ --------- SQLite headroom.db cwd or ~/.headroom/ Local file Metrics, sessions, compression store Memory DB ~/.headroom/headroom memory.db Local file Memory system optional Vector Index ~/.headroom/headroom memory vectors.db Local file Semantic search optional, requires sqlite-vec Graph Store ~/.headroom/headroom memory graph.db Local file Memory relationships Compression cache ~/.headroom/cache/ Directory Semantic + summary cache 证据：`docs/spec/010-data.md`
- **011. Deployment**（documentation）：Dockerfile: dockerfile FROM python:3.12-slim 证据：`docs/spec/011-deployment.md`
- **012. Diagrams**（documentation）：Version Date Changes --------- ------ --------- 1.0.0-draft 2026-04-16 Initial diagrams document 证据：`docs/spec/012-diagrams.md`
- **013. Disaster Recovery**（documentation）：Failure Impact Recovery --------- -------- ---------- Proxy crash Service unavailable Restart or failover OOM Service unavailable Restart with more memory Network partition Cannot reach providers Retry with backoff 证据：`docs/spec/013-disaster-recovery.md`
- **014. Governance**（documentation）：- Repository: github.com/JerrettDavis/headroom - License: Apache 2.0 - Main Languages: Python core , TypeScript SDK/dashboard 证据：`docs/spec/014-governance.md`
- **015. Interfaces**（documentation）：Options: Flag Default Description ------ --------- ------------- --host 127.0.0.1 Bind host --port 8787 Bind port --mode token Optimization mode: token or cache --workers 1 Uvicorn worker processes --limit-concurrency 1000 Maximum concurrent connections before 503 --no-optimize false Passthrough mode --no-cache false Disable semantic cache --no-rate-limit false Disable rate limiting --memory false Enable persistent memory --learn false Enable live traffic learning --backend anthropic Backend: anthropic, bedrock, openrouter, anyllm, or litellm- --no-telemetry false Disable anonymous telemetry --stateless false Disable filesystem writes 证据：`docs/spec/015-interfaces.md`
- **016. Observability**（documentation）：Headroom exposes Prometheus metrics at /metrics . 证据：`docs/spec/016-observability.md`
- **017. Operations**（documentation）：Basic health check. Returns 200 if process is running. 证据：`docs/spec/017-operations.md`
- **018. Policies**（documentation）：Setting Default Description --------- --------- ------------- Compression enabled true Apply compression by default Cache enabled true Use semantic cache Summary enabled true Use summary compression Token budget enforced false No budget limit by default 证据：`docs/spec/018-policies.md`
- **019. Quality**（documentation）：Surface Unit Integration E2E --------- :----: :------------: :---: Proxy Server ✓ ✓ ✓ SDK ✓ ✓ - Compression ✓ ✓ ✓ Cache ✓ ✓ - Learn ✓ ✓ - CCR ✓ ✓ ✓ TOIN ✓ ✓ - Dashboard ✓ - ✓ 证据：`docs/spec/019-quality.md`
- **020. Security**（documentation）：Threat Impact Mitigation -------- -------- ------------ Prompt data exfiltration High No logging, local-only API key theft High Key rotation, secrets management Cache poisoning Medium Input validation DoS via large prompts Medium Token limits SSRF via redirects Medium URL validation 证据：`docs/spec/020-security.md`
- **021. Testing**（documentation）：Unit Tests: - Request routing - Response header injection - Mode switching - Error handling 证据：`docs/spec/021-testing.md`
- **022. Rust Migration**（documentation）：Status: in-progress Stage 0 complete 证据：`docs/spec/022-rust-migration.md`
- **Headroom Living Specification**（documentation）：Version: 1.0.0-draft Date: 2026-04-16 Status: Draft — In Progress Related Issue: GitHub 183 证据：`docs/spec/SPEC.md`
- **Example .env file for act local testing.**（source_file）：Example .env file for act local testing. Copy this to .env.act which is gitignored and fill in test values. cp .env.act.example .env.act act automatically reads .env and passes as secrets to workflows. DO NOT commit .env — it contains real or test tokens. 证据：`.env.act.example`
- **Init**（source_file）：version = "0.2.0" ⋮---- all = 证据：`benchmarks/__init__.py`
- **Init**（source_file）：all = ⋮---- LAZY EXPORTS: dict str, tuple str, str = { ⋮---- OPTIONAL EXPORTS = { ⋮---- def getattr name: str - Any ⋮---- module attr = LAZY EXPORTS.get name ⋮---- value = getattr import module module name , attr name ⋮---- optional module attr = OPTIONAL EXPORTS.get name ⋮---- value = None ⋮---- def dir - list str 证据：`headroom/__init__.py`
- **=== Whitespace Normalization ===**（source_file）：class HeadroomMode str, Enum ⋮---- AUDIT = "audit" OPTIMIZE = "optimize" SIMULATE = "simulate" ⋮---- DEFAULT MODEL CONTEXT LIMITS: dict str, int = {} ⋮---- @dataclass class CacheAlignerConfig ⋮---- enabled: bool = False ⋮---- use dynamic detector: bool = True ⋮---- detection tiers: list Literal "regex", "ner", "semantic" = field ⋮---- extra dynamic labels: list str = field default factory=list ⋮---- entropy threshold: float = 0.7 ⋮---- date patterns: list str = field ⋮---- === Whitespace Normalization === normalize whitespace: bool = True collapse blank lines: bool = True ⋮---- Separator used to mark where dynamic content begins in system message Content before this separator is cached; con… 证据：`headroom/config.py`
- **Shared Context**（source_file）：logger = logging.getLogger name ⋮---- @dataclass class ContextEntry ⋮---- key: str original: str compressed: str original tokens: int compressed tokens: int agent: str None timestamp: float transforms: list str = field default factory=list ⋮---- @property def savings percent self - float ⋮---- @dataclass class SharedContextStats ⋮---- entries: int total original tokens: int total compressed tokens: int total tokens saved: int savings percent: float ⋮---- class SharedContext ⋮---- messages = {"role": "tool", "content": content} result = compress messages, model=self. model ⋮---- compressed = result.messages 0 .get "content", content ⋮---- compressed = json.dumps compressed ⋮---- entry = Cont… 证据：`headroom/shared_context.py`
- **Core: lightweight compression SmartCrusher, ContentRouter, CCR, TOIN**（source_file）：build-system requires = "maturin =1.5,<2.0" build-backend = "maturin" 证据：`pyproject.toml`
- **Init**（source_file）：all = 证据：`benchmarks/scenarios/__init__.py`
- **In Memory**（source_file）：use std::collections::VecDeque; use std::sync::Mutex; ⋮---- use dashmap::DashMap; ⋮---- pub struct InMemoryCcrStore { ⋮---- struct Entry { ⋮---- impl InMemoryCcrStore { ⋮---- pub fn new - Self { ⋮---- pub fn with capacity and ttl capacity: usize, ttl: Duration - Self { ⋮---- fn evict until under capacity &self { let mut guard = self.order.lock .expect "ccr order mutex poisoned" ; while self.map.len = self.capacity { let Some oldest = guard.pop front else { ⋮---- self.map.remove &oldest ; ⋮---- impl Default for InMemoryCcrStore { fn default - Self { ⋮---- impl CcrStore for InMemoryCcrStore { fn put &self, hash: &str, payload: &str { ⋮---- if let Some mut existing = self.map.get mut hash { ex… 证据：`crates/headroom-core/src/ccr/backends/in_memory.rs`
- **Mod**（source_file）：pub mod in memory; ⋮---- pub mod redis; pub mod sqlite; ⋮---- use std::path::PathBuf; ⋮---- use thiserror::Error; ⋮---- use crate::ccr::CcrStore; ⋮---- pub use self::redis::RedisCcrStore; pub use in memory::InMemoryCcrStore; pub use sqlite::SqliteCcrStore; ⋮---- pub enum CcrBackendConfig { ⋮---- impl CcrBackendConfig { ⋮---- pub fn sqlite default path: PathBuf - Self { ⋮---- ttl seconds: crate::ccr::DEFAULT TTL.as secs , ⋮---- pub fn in memory default - Self { ⋮---- pub enum CcrBackendInitError { ⋮---- fn from err: ::redis::RedisError - Self { ⋮---- pub fn from config config: &CcrBackendConfig - Result , CcrBackendInitError { ⋮---- Ok Box::new store ⋮---- Some prefix = RedisCcrStore::open w… 证据：`crates/headroom-core/src/ccr/backends/mod.rs`
- **Redis**（source_file）：use redis::Commands; ⋮---- use crate::ccr::CcrStore; ⋮---- pub struct RedisCcrStore { ⋮---- impl RedisCcrStore { ⋮---- pub fn open url: &str, default ttl seconds: u64 - redis::RedisResult { Self::open with prefix url, DEFAULT KEY PREFIX.to string , default ttl seconds ⋮---- pub fn open with prefix ⋮---- let mut conn = client.get connection ?; let : String = redis::cmd "PING" .query &mut conn ?; Ok Self { ⋮---- fn key for &self, hash: &str - String { format! "{}:{}", self.key prefix, hash ⋮---- pub fn default ttl seconds &self - u64 { ⋮---- impl CcrStore for RedisCcrStore { fn put &self, hash: &str, payload: &str { let key = self.key for hash ; let mut conn = match self.client.get connection… 证据：`crates/headroom-core/src/ccr/backends/redis.rs`
- **Sqlite**（source_file）：use std::sync::Mutex; ⋮---- use crate::ccr::CcrStore; ⋮---- pub struct SqliteCcrStore { ⋮---- impl SqliteCcrStore { ⋮---- pub fn open path: impl AsRef , default ttl seconds: u64 - rusqlite::Result { let path buf = path.as ref .to path buf ; ⋮---- conn.pragma update None, "journal mode", "WAL" ?; conn.pragma update None, "synchronous", "NORMAL" ?; ⋮---- conn.execute ⋮---- Ok Self { ⋮---- pub fn path &self - &Path { ⋮---- pub fn default ttl seconds &self - u64 { ⋮---- fn purge expired conn: &Connection, now: u64 - rusqlite::Result { let purged = conn.execute ⋮---- params! now as i64 , ⋮---- Ok purged ⋮---- fn now unix seconds - u64 { ⋮---- .duration since UNIX EPOCH ⋮---- .map d d.as secs .un… 证据：`crates/headroom-core/src/ccr/backends/sqlite.rs`
- **Mod**（source_file）：pub mod backends; ⋮---- use std::time::Duration; ⋮---- pub trait CcrStore: Send + Sync { ⋮---- fn is empty &self - bool { self.len == 0 ⋮---- pub fn compute key payload: & u8 - String { ⋮---- let hex = h.to hex ; ⋮---- hex.as str ..24 .to string ⋮---- pub fn marker for hash: &str - String { format! " " ⋮---- mod tests { ⋮---- fn compute key is 24 hex chars { let k = compute key b"hello world" ; assert eq! k.len , 24 ; assert! k ⋮---- fn compute key is deterministic { let a = compute key b"the same payload" ; let b = compute key b"the same payload" ; assert eq! a, b ; ⋮---- fn compute key diverges for different payloads { let a = compute key b"alpha" ; let b = compute key b"beta" ; assert ne… 证据：`crates/headroom-core/src/ccr/mod.rs`
- **Lib**（source_file）：pub mod auth mode; pub mod cache control; pub mod ccr; pub mod compression policy; pub mod relevance; pub mod signals; pub mod tokenizer; pub mod transforms; ⋮---- pub use cache control::compute frozen count; ⋮---- pub fn hello - &'static str { ⋮---- mod tests { ⋮---- fn hello returns crate name { assert eq! hello , "headroom-core" ; 证据：`crates/headroom-core/src/lib.rs`
- 其余 20 条证据见 `AI_CONTEXT_PACK.json` 或 `EVIDENCE_INDEX.json`。

## 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：`docs/README.md`, `README.md`, `examples/README.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：`docs/README.md`, `README.md`, `examples/README.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。

- **Headroom 概述**：importance `high`
  - source_paths: README.md, headroom/__init__.py, headroom/compress.py
- **安装指南**：importance `high`
  - source_paths: pyproject.toml, Dockerfile, docs/content/docs/installation.mdx
- **快速开始**：importance `high`
  - source_paths: headroom/cli/main.py, headroom/cli/wrap.py, headroom/cli/proxy.py
- **代理服务器**：importance `high`
  - source_paths: headroom/proxy/server.py, headroom/proxy/handlers/anthropic.py, headroom/proxy/handlers/openai.py, crates/headroom-proxy/src/lib.rs, crates/headroom-proxy/src/main.rs
- **CCR 可逆压缩**：importance `high`
  - source_paths: headroom/ccr/context_tracker.py, headroom/ccr/response_handler.py, headroom/ccr/tool_injection.py, crates/headroom-core/src/ccr/mod.rs, crates/headroom-core/src/ccr/backends/in_memory.rs
- **压缩算法**：importance `medium`
  - source_paths: headroom/transforms/smart_crusher.py, headroom/transforms/code_compressor.py, headroom/transforms/content_router.py, crates/headroom-core/src/transforms/smart_crusher/mod.rs, crates/headroom-core/src/transforms/mod.rs
- **MCP 服务器集成**：importance `high`
  - source_paths: headroom/mcp_registry/__init__.py, headroom/mcp_registry/claude.py, headroom/mcp_registry/install.py, headroom/integrations/mcp/server.py, headroom/ccr/mcp_server.py
- **CLI 代理包装器**：importance `medium`
  - source_paths: headroom/cli/wrap.py, headroom/providers/claude/__init__.py, headroom/providers/codex/__init__.py, headroom/providers/cursor/__init__.py, headroom/providers/aider/__init__.py

## Repo Inspection Evidence / 源码检查证据

- repo_clone_verified: true
- repo_inspection_verified: true
- repo_commit: `65b59de999facc3d74c768e8bcc7d3be0db0c746`
- inspected_files: `pyproject.toml`, `Dockerfile`, `README.md`, `docker-compose.yml`, `uv.lock`, `docs/package-lock.json`, `docs/bedrock.md`, `docs/source.config.ts`, `docs/auth-modes.md`, `docs/package.json`, `docs/proxy.ts`, `docs/README.md`, `docs/tsconfig.json`, `docs/observability.md`, `docs/rtk-architecture.md`, `docs/spec/008-capabilities.md`, `docs/spec/015-interfaces.md`, `docs/spec/022-rust-migration.md`, `docs/spec/021-testing.md`, `docs/spec/005-integrations.md`

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: 来源证据：docs: clarify MCP setup when proxy /mcp is unavailable

- Trigger: GitHub 社区证据显示该项目存在一个安装相关的待验证问题：docs: clarify MCP setup when proxy /mcp is unavailable
- Host AI rule: 来源显示可能已有修复、规避或版本变化，说明书中必须标注适用版本。
- Why it matters: 可能增加新用户试用和生产接入成本。
- Evidence: community_evidence:github | cevd_109db57bc201482abc7bf318a0ee4792 | https://github.com/chopratejas/headroom/issues/460 | 来源讨论提到 linux 相关条件，需在安装/试用前复核。
- 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: 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 | github_repo:1129940957 | https://github.com/chopratejas/headroom | host_targets=mcp_host, claude, claude_code, cursor
- 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: 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 | github_repo:1129940957 | https://github.com/chopratejas/headroom | 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 4: 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 | github_repo:1129940957 | https://github.com/chopratejas/headroom | 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 5: 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 | github_repo:1129940957 | https://github.com/chopratejas/headroom | 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: 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 | github_repo:1129940957 | https://github.com/chopratejas/headroom | 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 7: 来源证据：[BUG] CCR proactive expansion blocks corrupt message attribution in multi-agent threads

- Trigger: GitHub 社区证据显示该项目存在一个安全/权限相关的待验证问题：[BUG] CCR proactive expansion blocks corrupt message attribution in multi-agent threads
- Host AI rule: 来源问题仍为 open，Pack Agent 需要复核是否仍影响当前版本。
- Why it matters: 可能影响授权、密钥配置或安全边界。
- Evidence: community_evidence:github | cevd_5ab74a3c022a4924998aaa72cc334c04 | https://github.com/chopratejas/headroom/issues/503 | 来源讨论提到 python 相关条件，需在安装/试用前复核。
- 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 8: 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 | github_repo:1129940957 | https://github.com/chopratejas/headroom | 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 9: 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 | github_repo:1129940957 | https://github.com/chopratejas/headroom | 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。