# world-model-optimizer - 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 world-model-optimizer 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

- **希望把专业流程带进宿主 AI 的用户**：仓库包含 Skill 文档。 证据：`.claude/skills/improve-judge/SKILL.md`, `.claude/skills/ready-for-merge/SKILL.md` Claim：`clm_0003` supported 0.86

## 它能做什么

- **AI Skill / Agent 指令资产库**（可做安装前预览）：项目包含可被宿主 AI 读取的 Skill 或 Agent 指令文件，可用于把专业流程带入 Claude、Codex、Cursor 等宿主。 证据：`.claude/skills/improve-judge/SKILL.md`, `.claude/skills/ready-for-merge/SKILL.md` 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`, `packages/environment-capture/README.md`, `packages/environment-capture/bird-sql/README.md`, `packages/environment-capture/swe-bench/README.md` 等 Claim：`clm_0002` supported 0.86

## 怎么开始

- `pip install world-model-optimizer` 证据：`README.md` Claim：`clm_0004` supported 0.86, `clm_0005` supported 0.86
- `pip install "world-model-optimizer[e2b]"` 证据：`README.md` Claim：`clm_0005` supported 0.86
- `pip install gdown` 证据：`packages/environment-capture/bird-sql/README.md` Claim：`clm_0006` supported 0.86
- `git clone --depth 1 https://github.com/SWE-agent/mini-swe-agent.git` 证据：`packages/environment-capture/swe-bench/README.md` Claim：`clm_0007` supported 0.86
- `git clone --depth 1 https://github.com/sierra-research/tau2-bench.git` 证据：`packages/environment-capture/tau-bench/README.md` Claim：`clm_0008` supported 0.86
- `pip install environment-capture            # the library: contract, capture driver, hygiene, hub fetch` 证据：`packages/environment-capture/README.md` Claim：`clm_0009` supported 0.86
- `pip install 'environment-capture[fetch]'   # + huggingface_hub, for publishing bundles` 证据：`packages/environment-capture/README.md` Claim：`clm_0010` supported 0.86
- `curl -LO https://huggingface.co/datasets/experiential-labs/wmo-dabstep-traces/resolve/main/traces.otel.jsonl` 证据：`packages/environment-capture/README.md` Claim：`clm_0011` supported 0.86
- `pip install "llm-waterfall[bedrock]"           # extras: bedrock, openai, anthropic, azure, all` 证据：`packages/llm-waterfall/README.md` Claim：`clm_0012` supported 0.86
- `pip install "llm-waterfall[all]"` 证据：`packages/llm-waterfall/README.md` Claim：`clm_0013` supported 0.86

## 继续前判断卡

- **当前建议**：需要管理员/安全审批
- **为什么**：继续前可能涉及密钥、账号、外部服务或敏感上下文，建议先经过管理员或安全审批。

### 30 秒判断

- **现在怎么做**：需要管理员/安全审批
- **最小安全下一步**：先跑 Prompt Preview；若涉及凭证或企业环境，先审批再试装
- **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：希望把专业流程带进宿主 AI 的用户**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`.claude/skills/improve-judge/SKILL.md`, `.claude/skills/ready-for-merge/SKILL.md` Claim：`clm_0003` supported 0.86
- **能力存在：AI Skill / Agent 指令资产库**（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/skills/improve-judge/SKILL.md`, `.claude/skills/ready-for-merge/SKILL.md` 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`, `packages/environment-capture/README.md`, `packages/environment-capture/bird-sql/README.md`, `packages/environment-capture/swe-bench/README.md` 等 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, `clm_0005` 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/skills/improve-judge/SKILL.md`, `.claude/skills/ready-for-merge/SKILL.md`, `AGENTS.md`, `CLAUDE.md`
- **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。
- **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`, `packages/environment-capture/README.md`, `packages/environment-capture/bird-sql/README.md`, `packages/environment-capture/swe-bench/README.md` 等
- **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/skills/improve-judge/SKILL.md`, `.claude/skills/ready-for-merge/SKILL.md`, `AGENTS.md`, `CLAUDE.md`
- **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`, `packages/environment-capture/README.md`, `packages/environment-capture/bird-sql/README.md`, `packages/environment-capture/swe-bench/README.md` 等
- **环境变量 / API Key**：项目入口文档明确出现 API key、token、secret 或账号凭证配置。 原因：如果真实安装需要凭证，应先使用测试凭证并经过权限/合规判断。 证据：`.agents/docs/reference/rag-scaling-methodology.md`, `README.md`, `docs/reference/distill.md`, `docs/reference/failover.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**：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。）
- **不要使用真实生产凭证**：环境变量/API key 一旦进入宿主或工具链，可能产生账号和合规风险。（适用：出现 API、TOKEN、KEY、SECRET 等环境线索时。）
- **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。
- **准备撤销测试 API key 或 token**：测试凭证泄露或误用时，可以快速止损。
- **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_0014` inferred 0.45
- **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`, `packages/environment-capture/README.md`, `packages/environment-capture/bird-sql/README.md`, `packages/environment-capture/swe-bench/README.md` 等 Claim：`clm_0015` 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

- **AI Skill / Agent 指令资产库**：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：`.claude/skills/improve-judge/SKILL.md`, `.claude/skills/ready-for-merge/SKILL.md` 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`, `packages/environment-capture/README.md`, `packages/environment-capture/bird-sql/README.md`, `packages/environment-capture/swe-bench/README.md` 等 Claim：`clm_0002` supported 0.86

### 上下文规模

- 文件总数：1064
- 重要文件覆盖：40/1064
- 证据索引条目：80
- 角色 / Skill 条目：2

### 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 world-model-optimizer 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 world-model-optimizer 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 world-model-optimizer 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 2 个角色 / Skill / 项目文档条目。

- **improve-judge**（skill）：Iteratively improve the RubricJudge or any LLM scorer against a hand-labeled dataset - run the judge, identify false positives/negatives, diagnose why each failed, propose one experiment prompt, model, or context per failure class, and prove the fix without regressing controls. Use when the user distrusts judge scores, asks to improve/calibrate/tune the judge, or a new corpus type needs judge coverage. 激活提示：当用户任务与“improve-judge”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`.claude/skills/improve-judge/SKILL.md`
- **ready-for-merge**（skill）：Mandatory pre-merge gate for every PR. Runs /code-review --fix at an effort level scaled to the PR's breadth, resolves every review comment Cursor, Greptile, humans , and verifies the diff complies with AGENTS.md. Use whenever the user says a PR is ready to merge, asks to merge, or invokes /ready-for-merge. 激活提示：当用户任务与“ready-for-merge”描述的流程高度相关时，先用它做安装前体验，再决定是否安装。 证据：`.claude/skills/ready-for-merge/SKILL.md`

## 证据索引

- 共索引 80 条证据。

- **docs — finished products only, kept deliberately small**（documentation）：docs — finished products only, kept deliberately small 证据：`docs/README.md`
- **.agents/docs — working docs**（documentation）：The unclean side of the documentation AGENTS.md rule 5 : drafts, design notes, experiment logs, raw results, proposals. Committed so it transfers across worktrees and chats; pruned periodically; nothing outside .agents/ may depend on it. When something matures, its cleaned product is promoted to docs/ writeups → docs/research/ , verified how-tos → docs/reference/ and the working copy dies here. The Notion Eng Docs database was migrated here 2026-07-02; files keep their Notion area / status in frontmatter. 证据：`.agents/docs/README.md`
- **benchmark-grid — reference run inputs & artifacts**（documentation）：benchmark-grid — reference run inputs & artifacts 证据：`.agents/docs/research/benchmark-grid/README.md`
- **.agents — the agents' workspace**（documentation）：The unclean side of the work: one-off scripts, experiment runners, plans, scratchpads, drafts. Committed so it transfers across worktrees and chats — but that is the only guarantee. 证据：`.agents/README.md`
- **Agent guide — world-model-optimizer**（documentation）：Agent guide — world-model-optimizer 证据：`AGENTS.md`
- **Claude**（documentation）：See AGENTS.md ./AGENTS.md for project conventions and rules. They apply here in full. 证据：`CLAUDE.md`
- **World Model Optimizer**（documentation）：wmo turns agent traces you already collect into continuous improvement. Start with a model endpoint at frontier quality with 40%+ lower cost. Keep improving it with world model simulations, meta-harness optimization, and model distillation. 证据：`README.md`
- **This is NOT the Next.js you know**（documentation）：This version has breaking changes - APIs, conventions, and file structure may all differ from your training data. Read the relevant guide in node modules/next/dist/docs/ before writing any code. Heed deprecation notices. 证据：`web/AGENTS.md`
- **Claude**（documentation）：@AGENTS.md 证据：`web/CLAUDE.md`
- **web/ - the world-model-optimizer site**（documentation）：web/ - the world-model-optimizer site 证据：`web/README.md`
- **environment-capture**（documentation）：Run agent benchmarks for real and record every agent-environment transition — each action → observation pair, exactly as the environment returned it — as OpenTelemetry GenAI JSONL. Integrating a benchmark is one small adapter; ten are already in 5,900+ real trajectories / 27,000+ real transitions captured and published as license-tagged datasets on the Hugging Face Hub https://huggingface.co/experiential-labs . 证据：`packages/environment-capture/README.md`
- **appworld**（documentation）：A stateful multi-app world. AppWorld drops an agent into a simulated world of nine apps Amazon, Gmail, Venmo, Spotify, phone, file system, Splitwise, Todoist, SimpleNote plus a supervisor app for the account, behind 450+ real Python APIs. A task is a natural-language request — e.g. "what is the title of the most-liked song in my Spotify playlists" — and the agent completes it by writing Python that calls apis. . ... against a live, mutable world , signalling completion with apis.supervisor.complete task ... . This is the first adapter whose world state carries across steps variables and world mutations persist , which is exactly the world-model dynamics this benchmark exists to exercise — s… 证据：`packages/environment-capture/appworld/README.md`
- **bird-sql**（documentation）：Text-to-SQL over real SQLite databases. The environment is a workspace holding a fresh COPY of the task's database as database.db plus its DDL as schema.sql ; the agent explores with the sqlite3 CLI and submits a single SQLite SELECT / WITH query as its answer. Scoring is deterministic EXECUTION MATCH — the predicted and gold SQL are each run against a pristine read-only copy of the database and their result rows compared as an order-insensitive multiset order-sensitive when the question implies ordering — see environment capture/benchmarks/bird sql.py . 证据：`packages/environment-capture/bird-sql/README.md`
- **continual-learning**（documentation）：Database-exploration QA over a large, deliberately obfuscated SQLite database. The environment is a workspace holding one shared products.db ~400 MB of Amazon product/review data with cryptic table/column names, prices in integer cents, timestamps in epoch milliseconds, and drifted/corrupt values ; the agent explores it with real sqlite3 / python3 shell commands and submits a final answer. Scoring is deterministic and LLM-free — numeric match within the gold's absolute tolerance, else normalized text exact-match or containment — see environment capture/benchmarks/continual learning.py . 证据：`packages/environment-capture/continual-learning/README.md`
- **crmarena**（documentation）：Professional CRM work over a realistic Salesforce org. Each task is an analyst question — case routing, handle-time and transfer analytics, top-issue identification, entity disambiguation, policy-violation checks, or knowledge QA — answered by querying a real Salesforce-org database accounts, cases, orders, knowledge articles, case history, ... . The environment stages a fresh read-only copy of the org as crm.db plus a generated schema.md and a small query.py runner into the workspace; the agent explores with python3 query.py "SELECT ..." real rows as JSON and submits the value the question asks for. Scoring is deterministic and LLM-free — exact/contains for the analytical tasks Salesforce… 证据：`packages/environment-capture/crmarena/README.md`
- **dabstep**（documentation）：Data-analysis QA over a shared payments dataset and a business-rules manual. Each task is a question whose correct answer requires reading manual.md it defines what "authorized", "fee", and "fraud rate" mean — the raw columns are ambiguous on their own and computing over the CSV/JSON context files with real shell + pandas. The environment stages the task's context files into a fresh workspace's ./data/ directory; the agent explores, analyzes, and submits an answer. Scoring is deterministic numeric tolerance 0.01, normalized string/list match, accepted alternates — see environment capture/benchmarks/dabstep.py . 证据：`packages/environment-capture/dabstep/README.md`
- **financebench**（documentation）：Financial-document QA over real SEC-filing evidence excerpts. The environment is a workspace whose docs/ holds the task's true evidence doc s plus 4 distractors; the agent retrieves with real shell commands and submits an answer. Scoring is deterministic numeric match, token-F1 fallback — see environment capture/benchmarks/financebench.py . 证据：`packages/environment-capture/financebench/README.md`
- **gaia2**（documentation）：A stateful multi-app simulated world. GAIA2 / Meta Agents Research Environments ARE drops an agent into a simulated universe of apps Contacts, Email, Messaging, Calendar, RentAFlat, Shopping, CabApp, CityApp, a sandbox file system, ... pre-populated with fictional user data. A USER message states a task — e.g. "save every apartment in zip codes whose violent-crime rate is 5-10" or "add together the ages of all my contacts in Dublin, then in Galway, and give the absolute difference" — and the agent completes it by calling the apps' real Python tools against a live, mutable world state persists across steps , then answering via AgentUserInterface send message to user . State carries across st… 证据：`packages/environment-capture/gaia2/README.md`
- **kimi-gui-control**（documentation）：Computer-use agent runs that drive macOS GUI apps Safari, Chrome, Notes, Finder, Calculator, … through the macOS Accessibility API plus a shell. Each trajectory is a task like "browse the latest cs.CL listings on arXiv, open the top paper, and report the title, author count, and abstract" : the agent reads the accessibility tree, takes a single targeted action, and re-reads the tree to confirm. 证据：`packages/environment-capture/kimi-gui-control/README.md`
- **SWE-bench trace capture isolated**（documentation）：This directory is a self-contained, local-only capture tool . It runs the real SWE-bench Verified https://www.swebench.com/ benchmark with the standard mini-swe-agent https://github.com/SWE-agent/mini-swe-agent harness and converts the recorded agent trajectories into the world-model-optimizer trace corpus packages/environment-capture/swe-bench/traces.otel.jsonl . 证据：`packages/environment-capture/swe-bench/README.md`
- **tau2-bench trace capture isolated**（documentation）：This directory is a self-contained, local-only capture tool . It runs the real tau²-bench https://github.com/sierra-research/tau2-bench benchmark and converts its trajectories into the world-model-optimizer trace corpus packages/environment-capture/tau-bench/traces.otel.jsonl . 证据：`packages/environment-capture/tau-bench/README.md`
- **RL smoke harness tau-bench**（documentation）：smoke.py exercises every wmo-side data path each downstream training chat will consume end-to-end against the real tau-bench world model on Bedrock, at tiny scale ~30 haiku calls, max steps=3 , 2 scenarios . It exists so we find interface problems in the RL seam here before the transfer prompts spawn four training chats that would each hit the same wall. 证据：`packages/environment-capture/tau-bench/rl/README.md`
- **terminal-tasks trace capture isolated**（documentation）：terminal-tasks trace capture isolated 证据：`packages/environment-capture/terminal-tasks/README.md`
- **Agent guide — llm-waterfall**（documentation）：A stateless, provider-agnostic LLM client that sends each call down an ordered waterfall of backends, failing over only on capacity errors throttling / 5xx / timeouts and propagating real client errors immediately. Every call returns which backend served it, token usage, and USD cost. 证据：`packages/llm-waterfall/AGENTS.md`
- **llm-waterfall**（documentation）：Pool every LLM quota you own behind one client. Rate limits are issued per model, per provider, per account — but a workload pinned to one backend can only ever use one of them, and stalls the moment it throttles. llm-waterfall chains your backends into a single stateless client: each call walks the chain in order, capacity errors throttling, 5xx, timeouts spill to the next backend, and real errors bad request, auth, validation raise immediately. A six-rung chain sustains roughly the sum of six rate limits instead of the minimum of one — capacity you already pay for, actually reachable. 证据：`packages/llm-waterfall/README.md`
- **wmo.distill**（documentation）：Distill a smaller student from a larger teacher on real agent tasks, training a Tinker LoRA and gating whether the adapter is promoted. Three modes, in descending order of maturity. 证据：`wmo/distill/README.md`
- **@earendil-works/pi-agent-core**（documentation）：Stateful agent with tool execution and event streaming. Built on @earendil-works/pi-ai . 证据：`wmo/harness/vendor/pi-agent/README.md`
- **Package**（package_manifest）：{ "name": "web", "version": "0.1.0", "private": true, "scripts": { "dev": "next dev", "build": "next build", "start": "next start", "lint": "eslint .", "typecheck": "tsc --noEmit", "index": "node scripts/build-index.mjs" }, "dependencies": { "next": "16.2.10", "react": "19.2.4", "react-dom": "19.2.4" }, "devDependencies": { "@tailwindcss/postcss": "^4", "@types/node": "^20", "@types/react": "^19", "@types/react-dom": "^19", "eslint": "^9", "eslint-config-next": "16.2.10", "tailwindcss": "^4", "typescript": "^5" } } 证据：`web/package.json`
- **Package**（package_manifest）：{ "name": "@earendil-works/pi-agent-core", "version": "0.80.3", "description": "General-purpose agent with transport abstraction, state management, and attachment support", "type": "module", "main": "./dist/index.js", "types": "./dist/index.d.ts", "exports": { ".": { "types": "./dist/index.d.ts", "import": "./dist/index.js" }, "./node": { "types": "./dist/node.d.ts", "import": "./dist/node.js" }, "./package.json": "./package.json" }, "files": "dist", "README.md" , "scripts": { "clean": "shx rm -rf dist", "build": "tsgo -p tsconfig.build.json", "test": "vitest --run", "test:harness": "vitest --run --config vitest.harness.config.ts", "coverage:harness": "vitest --run --config vitest.harness.c… 证据：`wmo/harness/vendor/pi-agent/package.json`
- **Improve the Judge**（skill_instruction）：An iterative calibration loop against a hand-labeled dataset. Never tweak the judge from intuition: every change starts from a disagreement you can point at and ends with a case that would catch its regression. Deep background and worked example, if it still exists .agents/ is prunable : .agents/docs/reference/judge-meta-eval-playbook.md . 证据：`.claude/skills/improve-judge/SKILL.md`
- **Ready for Merge**（skill_instruction）：This is the mandatory gate before merging any PR in this repository. Do not tell the user a PR is ready to merge until every step below has been completed and passes. 证据：`.claude/skills/ready-for-merge/SKILL.md`
- **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：`packages/environment-capture/LICENSE`
- **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：`packages/llm-waterfall/LICENSE`
- **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：`wmo/harness/vendor/pi-agent/LICENSE`
- **Design note: RAG-aware GEPA why, and how leakage is avoided**（documentation）：Design note: RAG-aware GEPA why, and how leakage is avoided 证据：`.agents/docs/design-decisions/rag-aware-gepa.md`
- **Transfer prompt — WM scenario-mining → filtered-BC ablation assume no prior context**（documentation）：Transfer prompt — WM scenario-mining → filtered-BC ablation assume no prior context 证据：`.agents/docs/proposals/HANDOFF.md`
- **Closed-loop evaluation future direction**（documentation）：Closed-loop evaluation future direction 证据：`.agents/docs/proposals/closed-loop-eval-spec.md`
- **Charter: inverse distillation — train an agent from mined scenarios + the world model**（documentation）：Charter: inverse distillation — train an agent from mined scenarios + the world model 证据：`.agents/docs/proposals/inverse-distillation-charter.md`
- **Research directions not yet run**（documentation）：A backlog of optimization experiments for the GEPA research harness see GEPA optimization research . Each is "one new Ablation class" away — the framework run ablation , seed aggregation, RubricJudge scoring is already in place. Adding one means writing a conditions list and a run condition, seed that wires the knob, then a scripts/ runner mirroring scripts/run trace scaling.py . 证据：`.agents/docs/proposals/research-directions.md`
- **Routing optimizer v1: benchmark + implementation proposal DISCUSSION DRAFT**（documentation）：Routing optimizer v1: benchmark + implementation proposal DISCUSSION DRAFT 证据：`.agents/docs/proposals/routing-optimizer-v1.md`
- **Sim–real policy-rank agreement research direction**（documentation）：Sim–real policy-rank agreement research direction 证据：`.agents/docs/proposals/sim-real-policy-rank-agreement.md`
- **Architecture**（documentation）：World Model Optimizer turns a frontier LLM into the environment your agent steps against, reconstructed from your own OpenTelemetry traces. This doc is the map: how the packages fit, the data that flows between them, and where to plug in new pieces. 证据：`.agents/docs/reference/architecture.md`
- **Benchmarks → traces: the real trace source**（documentation）：Benchmarks → traces: the real trace source 证据：`.agents/docs/reference/benchmarks-to-traces.md`
- **Embeddings phi for retrieval**（documentation）：Retrieval DreamGym top-k ranks past steps by cosine similarity of an embedding phi state, action . The embedder that produces phi is chosen by HarnessConfig.embed provider an EmbedderKind and sized by HarnessConfig.embed dim . 证据：`.agents/docs/reference/embeddings.md`
- **Playbook: hand-labeled meta-evals for improving automated components the judge method**（documentation）：Playbook: hand-labeled meta-evals for improving automated components the judge method 证据：`.agents/docs/reference/judge-meta-eval-playbook.md`
- **Methodology: trace-scaling + RAG-optimization experiments reproducible runbook**（documentation）：Methodology: trace-scaling + RAG-optimization experiments reproducible runbook 证据：`.agents/docs/reference/rag-scaling-methodology.md`
- **Runbook: building a world model from tau2 traces real Bedrock**（documentation）：Runbook: building a world model from tau2 traces real Bedrock 证据：`.agents/docs/reference/runbook-build-tau2-bedrock.md`
- **Iterating on BASE ENV PROMPT with replay fidelity**（documentation）：Iterating on BASE ENV PROMPT with replay fidelity 证据：`.agents/docs/research/base-env-prompt-iteration.md`
- **Benchmark results: reproducibility**（documentation）：Dated snapshot June 2026 . These numbers were measured on the then-committed 66-trace tau2 corpus; the corpus has since grown to 1033 traces and the repo layout changed PR 38: paths are now examples/tau-bench/... , and the committed per-step report JSONs under benchmarks/results/ were removed . The methodology and the headline conclusion the GEPA lift is stable, not run-to-run luck stand; rerun the commands below for current numbers. 证据：`.agents/docs/research/benchmark-results-reproducibility.md`
- **Concurrency scaling law — provenance data**（documentation）：Concurrency scaling law — provenance data 证据：`.agents/docs/research/concurrency_anthropic/NOTES.md`
- **Fidelity tiers — promoted**（documentation）：The report lives at docs/research/world model findings.md ../../../docs/research/world model findings.md promoted 2026-07-10, PR 55 . Raw per-suite ladder JSONs remain here under fidelity tiers/ ./fidelity tiers/ . 证据：`.agents/docs/research/fidelity-tiers.md`
- **GEPA optimization research**（documentation）：This is the research surface for the harness's optimization trajectory: prompt optimization GEPA today, heavier training methods tomorrow. It exists to try optimization directions empirically — change a knob, measure reconstruction fidelity, record the result — rather than guessing. It is the experimental sibling of Iterating on BASE ENV PROMPT with replay fidelity which hand-tunes the base prompt and Design note: RAG-aware GEPA which explains the RAG-aware, leak-free evaluation every experiment here inherits . Directions not yet run live in Research directions not yet run ; the first completed sweep is written up in Trace scaling law . 证据：`.agents/docs/research/gepa-optimization-research.md`
- **Benchmarking generate / execute / verify GEV**（documentation）：Benchmarking generate / execute / verify GEV 证据：`.agents/docs/research/gev-benchmarks.md`
- **Bench-EXEC scorecard: simulator vs real execution**（documentation）：Bench-EXEC scorecard: simulator vs real execution 证据：`.agents/docs/research/gev_bench_results/exec/exec_scorecard.md`
- **Bench-GEN labeling sheet tau-bench, 100 traces, budget 15**（documentation）：Bench-GEN labeling sheet tau-bench, 100 traces, budget 15 证据：`.agents/docs/research/gev_bench_results/gen/labeling_sheet.md`
- **Bench-GEN blind manual labels: summary**（documentation）：Bench-GEN blind manual labels: summary 证据：`.agents/docs/research/gev_bench_results/gen/labels_summary.md`
- **Bench-GEN: scenario generation, first empirical run**（documentation）：Bench-GEN: scenario generation, first empirical run 证据：`.agents/docs/research/gev_bench_results/gen/report.md`
- **GEV benchmark: consolidated scorecard 2026-07-24**（documentation）：GEV benchmark: consolidated scorecard 2026-07-24 证据：`.agents/docs/research/gev_bench_results/gev_scorecard.md`
- **Bench-VERIFY disagreement attribution**（documentation）：Bench-VERIFY disagreement attribution 证据：`.agents/docs/research/gev_bench_results/verify/disagreements.md`
- **Bench-VERIFY ground-truth inventory**（documentation）：Bench-VERIFY ground-truth inventory 证据：`.agents/docs/research/gev_bench_results/verify/ground_truth_notes.md`
- 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`, `.agents/docs/README.md`, `.agents/docs/research/benchmark-grid/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`, `.agents/docs/README.md`, `.agents/docs/research/benchmark-grid/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。

- **项目概览与快速开始**：importance `high`
  - source_paths: README.md, AGENTS.md, CLAUDE.md, pyproject.toml, justfile
- **Provider 注册、LLM 适配与价格追踪**：importance `high`
  - source_paths: wmo/providers/__init__.py, wmo/providers/registry.py, wmo/providers/models.py, wmo/providers/waterfall.py, wmo/providers/openai.py
- **追踪接入、归一化与路由策略**：importance `high`
  - source_paths: wmo/ingest/__init__.py, wmo/ingest/braintrust.py, wmo/ingest/langfuse.py, wmo/ingest/langsmith.py, wmo/ingest/mastra.py
- **世界模型、模拟引擎与环境捕获**：importance `high`
  - source_paths: wmo/engine/world_model.py, wmo/engine/play.py, wmo/engine/replay.py, wmo/engine/build.py, wmo/engine/knowledge.py
- **Agent Harness、变更提议与 E2B 沙箱执行**：importance `high`
  - source_paths: wmo/harness/create.py, wmo/harness/mutate.py, wmo/harness/proposer.py, wmo/harness/project_proposer.py, wmo/harness/skills.py
- **优化器、判定器与奖励建模**：importance `high`
  - source_paths: wmo/optimize/__init__.py, wmo/optimize/base.py, wmo/optimize/gepa.py, wmo/optimize/judge.py, wmo/optimize/judge_quality.py
- **模型蒸馏流水线**：importance `medium`
  - source_paths: wmo/distill/README.md, wmo/distill/__init__.py, wmo/distill/loop.py, wmo/distill/teacher.py, wmo/distill/rollouts.py
- **本地服务、托管平台与 Web 界面**：importance `medium`
  - source_paths: wmo/serving/__init__.py, wmo/serving/server.py, wmo/serving/chat.py, wmo/serving/builds.py, wmo/serving/endpoint_config.py

## Repo Inspection Evidence / 源码检查证据

- repo_clone_verified: true
- repo_inspection_verified: true
- repo_commit: `1d2a54719356b3b7be61daf7ac7fefa9acd42e7b`
- inspected_files: `README.md`, `pyproject.toml`, `uv.lock`, `docs/README.md`, `docs/reference/closed_loop.md`, `docs/reference/connect-library.md`, `docs/reference/cost_quality_dial.md`, `docs/reference/distill.md`, `docs/reference/eval_grid.md`, `docs/reference/eval_suites.md`, `docs/reference/failover.md`, `docs/reference/harness_delta.md`, `docs/reference/ingest.md`, `docs/research/world_model_findings.md`, `packages/environment-capture/INTEGRATION.md`, `packages/environment-capture/README.md`, `packages/environment-capture/appworld/README.md`, `packages/environment-capture/appworld/backend/fetch_data.py`, `packages/environment-capture/appworld/backend/smoke.py`, `packages/environment-capture/appworld/backend/world_backend.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: 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 | https://news.ycombinator.com/item?id=49063454 | 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 2: 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 | https://news.ycombinator.com/item?id=49063454 | 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。

- Trigger: no_demo
- Evidence: downstream_validation.risk_items | https://news.ycombinator.com/item?id=49063454 | 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 4: 存在评分风险

- Trigger: no_demo
- Why it matters: 风险会影响是否适合普通用户安装。
- Evidence: risks.scoring_risks | https://news.ycombinator.com/item?id=49063454 | 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: 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 | https://news.ycombinator.com/item?id=49063454 | 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 6: 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 | https://news.ycombinator.com/item?id=49063454 | 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。