Add Atlas quickstart guide

Co-authored-by: multica-agent <github@multica.ai>

INTEGRATION-hermes.md

@@ -0,0 +1,122 @@
+# Hermes Agent Integration
+
+Atlas can run on **[hermes-agent](https://github.com/NousResearch/hermes-agent)**
+(MIT, by Nous Research) as the host runtime. Hermes provides exactly what Atlas
+needs: agentskills.io-compatible skill loader, cron scheduler, MCP client,
+multi-platform delivery (incl. email/feishu/dingtalk), session memory + FTS5,
+and subagent delegation.
+
+This doc tells the operator how to install Atlas on top of Hermes in ≤ 30 minutes.
+
+## What you need
+
+- Hermes installed: `curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash`
+- Python 3.10+ (for `atlas-extractor` MCP wrapper)
+- This repo (`assistant-claw`) cloned
+
+## Step 1 — Drop Atlas skills into Hermes skills dir
+
+```bash
+# Hermes scans ~/.hermes/skills/ on startup
+mkdir -p ~/.hermes/skills/atlas
+cp -r skills/* ~/.hermes/skills/atlas/
+
+# Check Hermes sees them
+hermes skills list | grep claw-
+```
+
+Each `claw-*/SKILL.md` already has the agentskills.io-compatible frontmatter
+(`name`, `description`, `version`, `author`, `license`, `metadata.hermes.*`)
+so Hermes recognizes them immediately.
+
+## Step 2 — Fetch the nuwa-skill upstream once
+
+```bash
+# Public deployment
+git clone --depth 1 https://github.com/alchaincyf/nuwa-skill.git \
+  ~/.hermes/skills/atlas/claw-boss-distiller/upstream
+
+# Air-gapped deployment (after Moments mirror is up — see ADR-0001)
+git clone --depth 1 https://git.moments.top/Moments.top/nuwa-skill.git \
+  ~/.hermes/skills/atlas/claw-boss-distiller/upstream
+```
+
+## Step 3 — Install atlas-extractor as an MCP server (TBD code)
+
+The `mcp-tools/email-extractor/` Python package currently exposes a CLI
+(`atlas-extract`). Wrapping it as an MCP server is V0.5 work; until then
+Hermes can call the CLI via Bash.
+
+```bash
+cd mcp-tools/email-extractor
+python -m venv .venv
+.venv/bin/pip install -e '.[test]'
+.venv/bin/pytest -q   # confirm 7/7 passing
+```
+
+## Step 4 — Configure the boss's email account
+
+```bash
+hermes env set BOSS_EMAIL boss@clientco.com
+hermes env set ATLAS_IMAP_HOST imap.clientco.com
+hermes env set ATLAS_IMAP_USER atlas-readonly@clientco.com
+hermes env set ATLAS_IMAP_PASSWORD <app-password>     # never main password
+hermes env set ATLAS_STATE_DIR ~/.atlas/state
+```
+
+## Step 5 — Wire the autopilots into Hermes cron
+
+Atlas's 5 autopilot YAMLs live in `autopilots/atlas-*.yaml`. Translate each
+into Hermes natural-language cron:
+
+```bash
+hermes cron add "every day 7:30" "Run Atlas daily-brief skill"
+hermes cron add "every monday 8:00" "Run Atlas weekly-rollup skill"
+hermes cron add "every month on 1st 9:00" "Run Atlas monthly customer-health skill"
+hermes cron add "every quarter on 1st 10:00" "Run Atlas boss_skill quarterly refresh"
+```
+
+Andon (event-triggered) is set up via `hermes hooks add` once Atlas writes a
+sentinel file like `~/.atlas/state/runs/_andon_pending.flag`.
+
+## Step 6 — Smoke test
+
+```bash
+# Manually invoke claw-boss-distiller on the demo input
+hermes "Run claw-boss-distiller on skills/claw-boss-distiller/demo/INPUT/ \
+        and write the result to skills/claw-boss-distiller/demo/OUTPUT/boss_skill.live.md"
+
+# Check it produced a 5-layer doc
+diff skills/claw-boss-distiller/demo/OUTPUT/boss_skill.demo.md \
+     skills/claw-boss-distiller/demo/OUTPUT/boss_skill.live.md
+```
+
+## Step 7 — First Daily Brief
+
+Wait until the next scheduled 7:30, OR force-run:
+
+```bash
+hermes "Run Atlas daily-brief now and email it to ${BOSS_EMAIL}"
+```
+
+The boss receives a 5-minute Brief in their inbox.
+
+## Caveats
+
+1. **Hermes self-improving loop** — Hermes can auto-edit its own skills based
+   on usage. **Lock down `SOUL.md` and the SKILL.md `boundaries:` list** by
+   adding them to Hermes's skill-write deny list. We never want Atlas to
+   auto-relax its own guardrails.
+
+2. **Branding** — Hermes's default identity is "Hermes Agent by Nous Research".
+   Override the system prompt to introduce as "Atlas / 总助 Claw" for client
+   deployments. See `hermes_cli/skin_engine.py` for theming hooks.
+
+3. **Execution environment** — Use `local` or `docker` backends only for
+   client-internal deployments. SSH / Daytona / Modal assume external network.
+
+## Roadmap
+
+- [ ] V0.5: Wrap `atlas-extractor` as proper MCP server (replace Bash CLI invocation)
+- [ ] V0.5: Auto-generate Hermes skill manifest from `skills/*/SKILL.md` on commit
+- [ ] V1: Build a Hermes "skill pack" installer (`hermes skills install moments/atlas`)

QUICKSTART.zh-CN.md

@@ -0,0 +1,324 @@
+# 总助 Claw 快速安装配置
+
+这份文档只解决一件事：把 assistant-claw 仓库配置成一个可运行的总助 Claw / Atlas 实例。
+
+总助 Claw 是老板视角的项目执行雷达。它读取老板授权范围内的邮件和业务信号，维护项目、人员、客户状态，并输出日 Brief、周 Rollup、客户健康报告和 Andon 告警。
+
+这不是"三万同款 7 Agent / 38 技能"安装包。三万架构只作为参考，不要把写作、交易、社区、多 Bot 等通用团队能力塞进总助 Claw 的快速安装流程。
+
+## 可复制的安装配置引导词
+
+把下面这段发给负责部署的 Agent，让它按步骤引导你完成总助 Claw 配置：
+
+```text
+请帮我配置「总助 Claw」。
+
+这是 Moments 公司出品的老板视角项目执行雷达：1 个 Atlas Agent + 6 个专项技能 + 邮件抽取流水线。
+1. 请先学习 assistant-claw QuickStart：https://git.moments.top/Moments.top/assistant-claw/src/branch/main/QUICKSTART.zh-CN.md
+2. 问清：客户/老板信息、邮件接入方式、首次盘点范围、Brief 时间、告警渠道、数据边界。
+3. 请按步骤引导我完成：安装 Atlas 身份文件和 skills → 初始化 state/memory/boss_skill.md → 配置 OpenClaw agents.list 和 workspace → 配置飞书 bindings → 验证。
+
+不要拉邮件，除非我明确确认邮件权限和盘点范围。
+
+🦞 Atlas 团队能力：
+- 邮件解析
+- 项目追踪
+- 人员行为观察
+- 客户雷达
+- 老板规则蒸馏
+- Brief/报告生成
+
+请开始引导我配置吧！
+```
+
+## 安装前先确认
+
+开始配置前，先向客户或部署负责人确认这些信息：
+
+| 项 | 要确认什么 |
+|----|------------|
+| 客户公司 | 公司名、行业、规模、主要工作语言 |
+| 老板信息 | 老板姓名、职位、Brief 阅读习惯 |
+| 邮件权限 | IMAP / OAuth / 转发邮箱；是否允许拉取过去 12 个月邮件 |
+| 首次盘点范围 | 默认过去 12 个月；如客户邮箱太大，先从 3 或 6 个月开始 |
+| 输出节奏 | 日 Brief 时间，默认老板本地时间 07:30 |
+| 告警渠道 | 邮件、飞书、微信机器人或仅生成报告 |
+| 数据边界 | 原始邮件是否必须留在客户网络内；谁能看到 Brief |
+
+没有这些答案时，不要开始拉邮件，也不要配置自动推送。
+
+## 目标目录
+
+推荐把客户实例放在 OpenClaw workspace 下：
+
+```bash
+export ATLAS_HOME="$HOME/.openclaw/workspace-atlas"
+mkdir -p "$ATLAS_HOME"
+```
+
+每个客户一个独立 workspace。不要让多个客户共用同一套 `state/`、`memory/`、`boss_skill.md`。
+
+## 第 1 步：安装 Atlas 配置资产
+
+在 `assistant-claw` 仓库根目录执行：
+
+```bash
+cp IDENTITY.zh-CN.md "$ATLAS_HOME/IDENTITY.md"
+cp SOUL.zh-CN.md "$ATLAS_HOME/SOUL.md"
+cp AGENTS.zh-CN.md "$ATLAS_HOME/AGENTS.md"
+cp USER.md "$ATLAS_HOME/USER.md"
+cp TOOLS.md "$ATLAS_HOME/TOOLS.md"
+cp MEMORY.md "$ATLAS_HOME/MEMORY.md"
+cp BOOTSTRAP.md "$ATLAS_HOME/BOOTSTRAP.md"
+cp HEARTBEAT.md "$ATLAS_HOME/HEARTBEAT.md"
+
+cp -R skills "$ATLAS_HOME/skills"
+cp -R mcp-tools "$ATLAS_HOME/mcp-tools"
+cp -R state-schemas "$ATLAS_HOME/state-schemas"
+```
+
+说明：
+
+- 中文部署默认使用 `*.zh-CN.md` 作为运行文件。
+- `USER.md` 是客户侧配置，必须在部署时填写。
+- `skills/` 是总助 Claw 专项能力，不是通用技能市场合集。
+
+## 第 2 步：创建运行态目录
+
+```bash
+mkdir -p "$ATLAS_HOME/state/projects"
+mkdir -p "$ATLAS_HOME/state/people"
+mkdir -p "$ATLAS_HOME/state/customers"
+mkdir -p "$ATLAS_HOME/state/runs"
+mkdir -p "$ATLAS_HOME/state/unclustered"
+mkdir -p "$ATLAS_HOME/state/audit"
+mkdir -p "$ATLAS_HOME/state/extracted"
+mkdir -p "$ATLAS_HOME/memory"
+mkdir -p "$ATLAS_HOME/boss_voice"
+```
+
+初始化 `state/index.json`：
+
+```bash
+cat > "$ATLAS_HOME/state/index.json" <<'JSON'
+{
+  "projects": [],
+  "people": [],
+  "customers": [],
+  "updated_at": null
+}
+JSON
+```
+
+初始化当天记忆：
+
+```bash
+TODAY="$(date +%F)"
+cat > "$ATLAS_HOME/memory/$TODAY.md" <<EOF
+# $TODAY
+
+## 本日运行记录
+
+- 部署初始化。
+EOF
+```
+
+初始化 `boss_skill.md`：
+
+```bash
+cat > "$ATLAS_HOME/boss_skill.md" <<'EOF'
+# boss_skill.md
+
+状态：待蒸馏
+
+首次部署后，由 `claw-boss-distiller` 读取老板过去 6 个月发出邮件，生成第一版 draft rules。
+老板审核后，才能把规则标记为 confirmed。
+EOF
+```
+
+## 第 3 步：填写客户配置
+
+编辑：
+
+```bash
+$ATLAS_HOME/USER.md
+```
+
+至少填这些字段：
+
+- Principal name
+- Role
+- Industry
+- Company size
+- Email domain
+- Primary work language
+- Daily email volume
+- Average projects in flight
+- Reads Brief on
+- Tolerance for false positives
+
+如果这些字段为空，总助 Claw 可以启动，但第一周盘点和 Brief 会缺上下文。
+
+## 第 4 步：配置 OpenClaw Agent
+
+修改 OpenClaw 配置前先备份：
+
+```bash
+cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak.$(date +%Y%m%d-%H%M%S)
+```
+
+新增 Atlas Agent：
+
+```bash
+openclaw agents add atlas --workspace "$ATLAS_HOME"
+openclaw agents list
+```
+
+如果当前环境只能有一个默认 Agent，也可以把 agent id 设为 `main`。关键是 `workspace` 必须指向 `$ATLAS_HOME`。
+
+确认 `agents.list` 里有类似结构：
+
+```json
+{
+  "id": "atlas",
+  "name": "总助 Claw",
+  "workspace": "/Users/<you>/.openclaw/workspace-atlas"
+}
+```
+
+总助 Claw 快速安装不需要创建 `writer`、`researcher`、`trader` 等子 Agent，也不需要配置 `subagents.allowAgents`。需要多 Agent 时，等单实例跑通后再做二期扩展。
+
+## 第 5 步：配置飞书入口
+
+如果使用一个飞书 Bot 接入总助 Claw，只需要把该账号绑定到 `atlas`：
+
+```json
+{
+  "bindings": [
+    {
+      "agentId": "atlas",
+      "match": {
+        "channel": "feishu",
+        "accountId": "default"
+      }
+    }
+  ]
+}
+```
+
+保留现有 `channels.feishu` 下的账号配置和访问策略，不要为了新增 binding 删除 `allowFrom`、`dmPolicy`、`groupPolicy` 等字段。
+
+配置后验证：
+
+```bash
+openclaw agents list
+openclaw channels status
+```
+
+多数配置支持热加载。只有状态没刷新、版本要求重启、或已经确认维护窗口时，才执行：
+
+```bash
+openclaw gateway restart
+openclaw gateway status
+```
+
+## 第 6 步：配置邮件抽取工具
+
+总助 Claw 的核心输入是邮件。V0 推荐先跑本地 email-extractor pipeline，确认能稳定解析样例，再接客户邮箱。
+
+```bash
+cd "$ATLAS_HOME/mcp-tools/email-extractor"
+python3 -m venv .venv
+. .venv/bin/activate
+pip install -e .
+pytest
+```
+
+如果测试通过，再接入客户邮件源：
+
+- IMAP：确认 host、port、用户名、授权码、TLS 要求。
+- OAuth：确认回调地址、client id、client secret、refresh token 存放位置。
+- 转发邮箱：确认原始 MIME 是否保留完整 header。
+
+不要把原始邮件复制出客户允许的网络边界。总助 Claw 的输出应引用邮件 ID 和规则，不暴露不相关客户或员工隐私。
+
+## 第 7 步：首次启动验收
+
+在飞书或命令入口问：
+
+```text
+你是谁？请读取你的 IDENTITY、SOUL、USER，并告诉我当前还缺哪些部署字段。
+```
+
+合格回复应满足：
+
+- 自称总助 Claw / Atlas。
+- 能说清自己是老板视角项目执行雷达。
+- 明确 V0 不替老板写信、不发信、不评价员工。
+- 能列出 `USER.md` 中未填写字段。
+
+再问：
+
+```text
+请按 BOOTSTRAP.md 给出首次部署 checklist，不要开始拉邮件。
+```
+
+合格回复应输出：
+
+- 邮件权限确认
+- 历史拉取范围
+- 日 Brief 时间
+- Andon 告警渠道
+- `state/`、`memory/`、`boss_skill.md` 初始化状态
+- W1 一次性盘点计划
+
+## 第 8 步：首次盘点流程
+
+首次盘点不要自动全量跑。先让客户确认范围：
+
+```text
+首次盘点建议拉取过去 12 个月邮件。你的邮箱量级较大时，我建议先从 3 个月试跑，确认抽取质量和误报率后再扩展到 12 个月。请确认范围。
+```
+
+确认后按顺序执行：
+
+1. `claw-email-parser`：拉取、解码、去引用、抽实体、分类意图。
+2. `claw-project-tracker`：聚类项目，生成 `state/projects/*.json`。
+3. `claw-people-observer`：生成行为观察卡，永不输出员工裁定标签。
+4. `claw-customer-radar`：生成客户健康状态和风险信号。
+5. `claw-report-writer`：生成第一份 Inventory Report。
+6. `claw-boss-distiller`：只读取老板发出邮件，生成 `boss_skill.md` draft rules。
+
+第一周交付物：
+
+- 过去 3 / 6 / 12 个月项目清单
+- Stalled / Waiting-For / Next Action 风险图
+- 客户健康清单
+- `boss_skill.md` draft，等待老板审核
+
+## 不要放进快速安装的内容
+
+这些不是总助 Claw 快速安装的一部分：
+
+- 三万同款 7 Agent / 38 技能完整团队。
+- 公众号、博客、小红书、交易、社区运营角色。
+- 多飞书 Bot 企业版配置。
+- 任意通用技能市场清单。
+- 未经确认的 `sessions_spawn` 子 Agent 调度规则。
+
+总助 Claw 的最小可用版本是一个 Atlas 实例，加上邮件抽取、项目追踪、客户雷达、人员观察、报告渲染和 boss_skill 蒸馏。先把这个闭环跑通，再谈多 Agent 扩展。
+
+## 完成话术
+
+安装完成后，给客户这样汇报：
+
+```text
+总助 Claw 已完成基础配置：
+- Agent workspace：<ATLAS_HOME>
+- 已安装身份文件、专项 skills、mcp-tools、state schemas
+- 已初始化 state、memory、boss_skill.md
+- 已绑定入口：飞书 / 其他
+- 已验证：Agent 能读取身份文件，并能输出首次部署 checklist
+
+下一步需要你确认邮件接入方式和首次盘点范围。确认后，我会先试跑小范围邮件抽取，再生成第一份项目盘点报告。
+```

README.md

@@ -1,91 +1,192 @@
-# assistant-claw（总助Claw）
+# 总助 Claw / Atlas 🦅
 
-总助 Claw 的"配置中心"——同时承载两个 agent profile：
+> **老板视角项目执行雷达**——把散落在邮件 / 微信 / 飞书里的工作安排，自动织成一张"人 × 项目 × 客户"三维盘子，每天告诉老板：谁在跟、跟得怎样、有没有卡点、有没有不了了之的事、哪些客户要亲自出面。
 
-- **Vega** —— 通用助理（general-purpose teammate，原 README 设定）
+> "把太多不了了之的事，重新装回老板的视野里。"
-- **Atlas** —— 老板视角项目执行雷达（new；详见 [`atlas/README.md`](./atlas/README.md)）
-
-两个 profile 共用本仓库的目录约定（prompts / skills / tools / autopilots / docs），但 Atlas 是按客户实例化部署的纵向产品，所有 Atlas 特有资产（OpenClaw 8 件套 + 6 个子 skill + email-extractor + boss_skill 蒸馏 demo + 客户演示 deck）封装在 [`atlas/`](./atlas/) 子目录里。Vega 的内容继续按原计划长在顶层。
-
-Atlas 部署时建议直接 `cd atlas && cat README.md` 走它的部署 SOP；本仓库根目录的下面这份 Vega 配置说明保持不变。
 
 ---
 
-总助 Claw 的"配置中心"——管理 Vega（通用助理 Agent）的全部可托管内容：
+## 📍 想做什么 → 看哪里
 
-- **prompts/** — 系统提示词、人格设定、风格 guideline
+| 目标 | 路径 |
-- **skills/** — 可复用的 skill 定义（Markdown + frontmatter）
+|------|------|
-- **tools/** — Vega 调用的工具：MCP 服务配置、自定义 CLI 脚本、HTTP 工具适配
+| 快速安装配置 | [`QUICKSTART.zh-CN.md`](./QUICKSTART.zh-CN.md) |
-- **autopilots/** — 定时 / 触发式自动任务（对应 Multica `multica autopilot` 的配置）
+| 看是个什么产品（5 分钟读完） | [`AGENTS.md`](./AGENTS.md) + [`SOUL.md`](./SOUL.md) |
-- **docs/** — 设计文档、ADR、运行手册
+| 给客户讲（2 页文档） | [`client-deck/`](./client-deck/) |
+| 看 6 个 sub-skill | [`skills/`](./skills/) |
+| 看 boss_skill 蒸馏 demo（10 封邮件 → 5 层老板 OS） | [`skills/claw-boss-distiller/demo/`](./skills/claw-boss-distiller/demo/) |
+| 跑邮件抽取（Python 实现，pytest 通过） | [`mcp-tools/email-extractor/`](./mcp-tools/email-extractor/) |
+| 装到 Hermes Agent 上跑 | [`INTEGRATION-hermes.md`](./INTEGRATION-hermes.md) |
+| 项目卡 / 人员卡 / 客户卡 JSON schema | [`state-schemas/`](./state-schemas/) |
+| Autopilot 定义（5 个：daily / weekly / monthly / quarterly + andon） | [`autopilots/`](./autopilots/) |
+| 设计决策记录 | [`docs/decisions/`](./docs/decisions/) |
 
-> Vega 在 Multica workspace 里的 agent ID 是 `877b1ea4-d096-4323-9ccc-cda31a3a7b3d`，归属 project **总助Claw**（`fdc928a7-d204-4a9f-929e-2c59bea9cda3`）。她和其他垂直 Claw（`research-claw`、`trainer-claw`、`geo-claw`）并列，但 scope 故意更宽：日常杂活、轻量编码、写作、调研、规划、跨工具协作。
+---
 
-## 仓库结构
+## 🚀 快速启动
 
-```
+### 浏览设计
-.
+
-├── README.md
+```bash
-├── prompts/         # 系统提示词、人格、风格
+git clone https://git.moments.top/Moments.top/assistant-claw.git
-│   └── system.md         # 单一可信源（待落地）
+cd assistant-claw
-├── skills/          # 可触发的 skill 集合
+cat AGENTS.md            # 入口模式 + skill 名册 + autopilot 排班
-│   └── <skill-name>/
+cat SOUL.md              # 边界：永不评级员工 / 永不替老板写信 / 永带引用
-│       └── SKILL.md      # 含 frontmatter: name / description / type
+open client-deck/        # 客户演示
-├── tools/           # Vega 调用的工具
-│   ├── mcp/              # MCP server 配置
-│   └── scripts/          # 自定义 CLI / shell 脚本
-├── autopilots/      # 定时 / 触发的 autopilot 定义
-│   └── <name>.yaml       # title / agent / mode / schedule / payload
-└── docs/
-    └── decisions/        # ADR
 ```
 
-各目录现在用 `.gitkeep` 占位，会随后补齐。
+### 跑邮件抽取（本地，无需任何外部凭据）
 
-## 内容类型说明
+```bash
+cd mcp-tools/email-extractor
+python -m venv .venv && .venv/bin/pip install -e '.[test]'
+.venv/bin/pytest -q                       # 7/7 测试通过
+.venv/bin/atlas-extract dir \
+  --input-dir ../../skills/claw-boss-distiller/demo/INPUT \
+  --state-dir /tmp/atlas-test
+```
 
-### prompts/
+### 拉 nuwa-skill 上游（claw-boss-distiller 依赖）
-Vega 的 system prompt 与人格设定，唯一可信源。线上 agent 的 prompt 应该从这里发布，而不是直接在 Multica 后台手动编辑。
 
-### skills/
+```bash
-每个 skill 一个子目录，至少包含 `SKILL.md`，frontmatter 必须有：
+# 公网部署
-- `name` — 唯一短名（kebab-case）
+git clone --depth 1 https://github.com/alchaincyf/nuwa-skill.git \
-- `description` — 一行触发说明（用于决定何时调用，写得越具体越好）
+  skills/claw-boss-distiller/upstream
-- `type` — `rigid`（必须严格遵循） 或 `flexible`（按情境调整）
 
-Skill 必须能给出一个具体的 trigger 场景；写不出来的就不要加。
+# 客户内网（air-gapped）部署 —— 用 Moments 私有 Gitea 镜像
+# 镜像建立流程见 docs/decisions/0001-mirror-nuwa-skill.md
+git clone --depth 1 https://git.moments.top/Moments.top/nuwa-skill.git \
+  skills/claw-boss-distiller/upstream
+```
 
-### tools/
+### 接到 Hermes Agent 跑
-- `tools/mcp/` — MCP server 的连接配置（command、env、scope 等）
-- `tools/scripts/` — Vega 通过 Bash 调用的自定义脚本，每个脚本顶部一段说明 + 用法示例
 
-### autopilots/
+详见 [`INTEGRATION-hermes.md`](./INTEGRATION-hermes.md)。一句话：6 个 SKILL.md 已经按 agentskills.io / Hermes 兼容格式写好，`cp -r skills/* ~/.hermes/skills/atlas/` 即可被识别。
-对应 `multica autopilot create` 的配置文件，每个 autopilot 一个 YAML/JSON：
-- `title`、`description`、`agent`、`mode`（默认 `create_issue`）
-- `schedule`（cron 或 trigger 描述）
-- `payload` / `template`（每次触发时落到 issue 的内容模板）
 
-发布时通过脚本读这些文件批量调 `multica autopilot create/update`，避免漂移。
+---
 
-### docs/
+## 📐 仓库结构
-设计决策（ADR）、运行手册、复盘记录。改 `prompts/` 或 `skills/` 之前先写一段 ADR，比事后追溯便宜。
 
-## 工作约定
+```
+assistant-claw/                     ← 本仓库 = 总助 Claw / Atlas
+│
+├── README.md                       本文件
+├── .gitignore
+│
+├── 🆔 OpenClaw 8 件套（agent 身份层，根目录直接读）
+│   ├── IDENTITY.md  / .zh-CN.md    🦅 Atlas 身份卡：老板视角项目执行雷达
+│   ├── SOUL.md      / .zh-CN.md    边界与原则
+│   ├── USER.md                     谁是 user（老板，按客户实例化）
+│   ├── AGENTS.md    / .zh-CN.md    入口模式 / skill 名册 / 工具 / autopilot 排班
+│   ├── TOOLS.md                    工具配置（IMAP/Gmail/Exchange + 报告渲染）
+│   ├── MEMORY.md                   长期记忆模板
+│   ├── BOOTSTRAP.md                首次部署 7 步清单
+│   ├── HEARTBEAT.md                每日 / 周 / 月 / 季 + 事件触发任务表
+│   └── CLAUDE.md → AGENTS.md       软链
+│
+├── 📚 INTEGRATION-hermes.md        Hermes Agent 集成 SOP（30 min）
+│
+├── 🛠️ skills/                      6 个 Atlas sub-skill
+│   ├── DESCRIPTION.md              Hermes category 描述
+│   ├── README.md                   skill 协作流图
+│   ├── claw-email-parser/SKILL.md      邮件抽取入口
+│   ├── claw-project-tracker/SKILL.md   项目状态判定（GTD/RACI）
+│   ├── claw-people-observer/SKILL.md   人员观察（硬约束：永不打员工等级）
+│   ├── claw-customer-radar/SKILL.md    客户健康度（CHS + Andon）
+│   ├── claw-boss-distiller/            ⭐ nuwa-skill 衍生
+│   │   ├── SKILL.md
+│   │   ├── ADAPTER.md              nuwa→Atlas 改造说明
+│   │   ├── boss_skill.seed.md      23 条种子规则跨 5 层
+│   │   ├── upstream/README.md      部署时 git clone nuwa 的 SOP
+│   │   └── demo/                   完整蒸馏 demo（10 邮件 → 5 层）
+│   └── claw-report-writer/SKILL.md     日 / 周 / 月报渲染
+│
+├── 🔌 mcp-tools/                   工具层
+│   ├── README.md                   工具清单 + 授权矩阵
+│   ├── email-extractor.md          7 阶段抽取流水线 spec
+│   └── email-extractor/            Python V0 实现（pytest 7/7 通过）
+│       ├── atlas_extractor/        fetch + decode + dequote + pipeline + cli
+│       ├── tests/                  fixture + test_dequote
+│       └── pyproject.toml          → 安装后得到 atlas-extract CLI
+│
+├── 📦 state-schemas/               三张卡的正式 JSON schema
+│   ├── README.md
+│   ├── project.md
+│   ├── person.md
+│   └── customer.md
+│
+├── ⏰ autopilots/                  5 个 Multica autopilot 配置
+│   ├── atlas-daily-brief.yaml          cron 30 7 * * *
+│   ├── atlas-weekly-rollup.yaml        cron 0 8 * * 1
+│   ├── atlas-monthly-customer-health.yaml  cron 0 9 1 * *
+│   ├── atlas-quarterly-boss-skill-refresh.yaml  cron 0 10 1 1,4,7,10 *
+│   └── atlas-andon-event.yaml          event_triggered（非定时）
+│
+├── 🎤 client-deck/                 给客户的演示主材料
+│
+├── 📝 docs/
+│   └── decisions/                  ADR
+│       └── 0001-mirror-nuwa-skill.md
+│
+└── 🔧 scripts/
+    └── mirror-nuwa-to-moments.sh   一键镜像 nuwa-skill 到 Moments Gitea
+```
 
-- **Issue / 任务流转走 Multica**：所有给 Vega 的任务、反馈、复盘走 Multica 的 issue + comment，**不在仓库里开 issue**。
+---
-- **改动走 PR**：`prompts/`、`skills/`、`tools/`、`autopilots/` 下的内容都会进 production，必须 PR + review，不直接 push main。
-- **保持 lean**：宁可少加，也不堆没人用的脚手架。每加一个 skill / tool / autopilot 都要能回答："谁在什么场景下会触发它"。
 
-## 路线图（暂定）
+## 🎯 产品定位
 
-- [ ] 把当前 Vega 的 system prompt 落到 `prompts/system.md`
+| 维度 | Atlas |
-- [ ] 抽出最常用的 3–5 个 skill 到 `skills/`（issue triage、comment drafting、研究汇总等）
+|------|-------|
-- [ ] 把现有 autopilot 反向导出到 `autopilots/`，建立"配置即代码"基线
+| **服务对象** | 企业一把手（董事长 / 总经理 / 创始人）一人 |
-- [ ] `docs/decisions/0001-why-general-agent.md`：为什么是通用 agent 而不是又一个垂直 claw
+| **核心痛点** | 太多事情不了了之 |
-- [ ] CI：对 `prompts/` 和 `skills/` 做 schema / lint 校验
+| **数据源** | 邮件（V0）→ 微信（V0.5）→ 飞书（V1） |
-- [ ] 发布脚本：从 `autopilots/*.yaml` 批量同步到 Multica
+| **学习成本** | 0（老板照常发邮件，AI 在后台读） |
+| **首份交付** | W1 末："过去一年项目盘点 + 落地率诊断"报告 |
+| **日常节奏** | 每天 7:30 一封 5 分钟看完的 Brief |
+| **北极星指标** | 老板看完 Brief 能立刻说出今日 3 件事 |
+| **部署模型** | 100% 客户私有部署，每客户一个独立实例 |
 
-## 相关链接
+引用 10 套主流管理框架（GTD / RACI / Andy Grove / 9-Box / BARS / CCAR / ONA / CHS / VoC / WBR），每条 AI 判定都标依据规则号，老板 override 一行邮件就生效。
 
-- Multica project：总助Claw（`fdc928a7-d204-4a9f-929e-2c59bea9cda3`）
+---
-- 兄弟仓库：`research-claw` / `trainer-claw` / `geo-claw`
+
+## ⚙️ 工作约定
+
+- **Issue / 任务流转走 Multica**：所有给 Atlas 的需求、反馈、复盘走 Multica 的 issue + comment，不在仓库里开 issue
+- **改动走 PR**：根目录文件、`skills/`、`mcp-tools/`、`autopilots/`、`state-schemas/` 下的内容都进 production，必须 PR + review
+- **每客户实例隔离**：每个客户的 `state/`（projects/people/customers/runs/audit）永远在客户机器上，不进本仓库
+- **保持 lean**：宁可少加，也不堆没人用的脚手架
+- **永不放宽硬边界**：SOUL.md + 各 SKILL.md `boundaries:` 字段是死线（不评级员工 / 不写邮件给非老板 / 跨客户隔离 / 不编造证据）
+
+---
+
+## 🔗 相关
+
+- **Multica project**：总助 Claw（`fdc928a7-d204-4a9f-929e-2c59bea9cda3`）
+- **上游开源依赖**：
+  - [`alchaincyf/nuwa-skill`](https://github.com/alchaincyf/nuwa-skill)（MIT）—— 心智蒸馏方法论
+  - [`NousResearch/hermes-agent`](https://github.com/NousResearch/hermes-agent)（MIT）—— 推荐运行时
+- **agentskills.io 标准**：https://agentskills.io
+- **Mirror（air-gapped 客户部署用）**：`https://git.moments.top/Moments.top/nuwa-skill`（待建）
+
+---
+
+## 🗺️ 路线图
+
+### 已完成
+- [x] OpenClaw 8 件套 + 6 sub-skill 设计
+- [x] email-extractor V0（Stage 1-3 实现 + 7 测试通过）
+- [x] boss-distiller demo run（10 封邮件 → 5 层蒸馏验证）
+- [x] 客户演示 deck v1
+- [x] Hermes 集成 SOP
+- [x] state schemas（project / person / customer）
+- [x] 5 个 autopilot YAML
+
+### 待办
+- [ ] **nuwa-skill 镜像到 Moments Gitea**（仓库已建则 `bash scripts/mirror-nuwa-to-moments.sh`）
+- [ ] email-extractor Stage 4-7（threading / entities / intent / canonical 输出）
+- [ ] 6 个 sub-skill 的可执行实现（当前是设计文档）
+- [ ] `atlas-extractor` 包成 MCP server
+- [ ] dogfood：装到内部 Hermes 实例端到端跑一次
+- [ ] 第一个客户 W1 部署

atlas/README.md

@@ -1,40 +0,0 @@
-# 总助 Claw (Atlas) — Plan & Scaffold
-
-Boss-perspective project execution radar. Built on the OpenClaw agent framework (mirrors `zero-cca/` structure).
-
-## Quick Map
-
-| Layer | Files |
-|-------|-------|
-| **Identity (8 OpenClaw core files)** | `IDENTITY.md` `SOUL.md` `USER.md` `AGENTS.md` `TOOLS.md` `MEMORY.md` `BOOTSTRAP.md` `HEARTBEAT.md` (+ `CLAUDE.md` symlink) |
-| **Skills (6 sub-skills)** | `skills/{claw-email-parser, claw-project-tracker, claw-people-observer, claw-customer-radar, claw-boss-distiller, claw-report-writer}/SKILL.md` |
-| **MCP tools** | `mcp-tools/email-extractor.md` (the deep-dive 7-stage extraction pipeline) + `mcp-tools/README.md` |
-| **State (runtime)** | created at deployment under `state/` per-client |
-| **Memory (runtime)** | created at deployment under `memory/` |
-
-## Counts
-
-- **1 Agent** (Atlas / 总助)
-- **6 Skills** (1 master would re-enter via entry points A/B/C/D/E)
-- **6 MCP tools** (email-fetch, email-extractor, state-io, report-render, email-send, rule-audit)
-- **5 Autopilot schedules** (daily / weekly / monthly / quarterly + event-triggered)
-- **3 State entity types** (project / person / customer) + audit + runs + extracted
-- **5 Report modes** (daily / weekly / monthly / inventory / ad-hoc)
-- **5 Layers** in `boss_skill.md` (Expression DNA / Mental Models / Decision Heuristics / Anti-Patterns / Honest Boundaries)
-- **1 Open-source dependency** (`nuwa-skill` MIT, vendored into `claw-boss-distiller`)
-
-## Read Order for a New Atlas Operator
-
-1. `BOOTSTRAP.md` (first run only)
-2. `SOUL.md` (boundaries — read once, internalize)
-3. `IDENTITY.md` (role)
-4. `USER.md` (boss profile, filled at deployment)
-5. `AGENTS.md` (entry points + skill roster + tool roster + autopilot)
-6. `mcp-tools/email-extractor.md` (the unsexy but critical extraction layer)
-7. `skills/README.md` then individual `SKILL.md`s
-
-## Status
-
-Scaffold v0.4 — generated 2026-05-09 from PRD v0.3 + nuwa-skill review + OpenClaw framework adoption.
-
-Next: write `state-schemas/{project,person,customer}.md` formal JSON schemas, write the boss_skill.md seed template, and begin the W1 implementation sprint.

docs/decisions/0001-mirror-nuwa-skill.md

@@ -0,0 +1,63 @@
+# ADR-0001: Mirror nuwa-skill into Moments private Gitea
+
+**Date:** 2026-05-09
+**Status:** Accepted
+**Authors:** Atlas team
+
+## Context
+
+Atlas's `claw-boss-distiller` skill (under `skills/claw-boss-distiller/`) is
+adapted from the open-source [`alchaincyf/nuwa-skill`](https://github.com/alchaincyf/nuwa-skill)
+(MIT). It uses upstream's triple-verification methodology and skill template
+(see `skills/claw-boss-distiller/ADAPTER.md`).
+
+For deployment, each Atlas instance needs the upstream code:
+
+```bash
+git clone --depth 1 https://github.com/alchaincyf/nuwa-skill.git \
+  skills/claw-boss-distiller/upstream
+```
+
+This works for environments with public internet. **It does not work for many
+enterprise client deployments** which Atlas explicitly targets — most of our
+customer base operates inside corporate networks that block direct GitHub access.
+
+## Decision
+
+Mirror `alchaincyf/nuwa-skill` to `git.moments.top/Moments.top/nuwa-skill` and
+treat the Moments mirror as the canonical fetch source for client deployments.
+
+The bot account (`multica-bot`) cannot create org repos via API (lacks
+`write:organization` scope). Repo creation is a one-time human step:
+
+> https://git.moments.top/repo/create → owner: `Moments.top` → name: `nuwa-skill`
+> → empty (no auto-init)
+
+After repo exists, run `./scripts/mirror-nuwa-to-moments.sh` to populate it.
+
+## Consequences
+
+### Positive
+- Air-gapped client deployments work (no GitHub access required)
+- Stable fetch URL even if upstream is unavailable
+- Allows pinning to a specific commit per-deployment if needed
+- We retain MIT license attribution (no IP impact)
+
+### Negative
+- Adds a sync responsibility — when upstream evolves, we need to re-mirror
+- Slight risk of staleness if mirror cron fails silently
+
+## Mitigation
+
+- Run `./scripts/mirror-nuwa-to-moments.sh` on a weekly cron (TBD: add to
+  `autopilots/` once decided)
+- Each Atlas deployment records the upstream commit SHA at install time so
+  drift is auditable
+- README in `skills/claw-boss-distiller/upstream/` clearly notes the
+  mirror is a downstream copy, with last-mirrored timestamp
+
+## References
+
+- Upstream: https://github.com/alchaincyf/nuwa-skill (MIT)
+- Mirror script: `scripts/mirror-nuwa-to-moments.sh`
+- Adapter notes: `skills/claw-boss-distiller/ADAPTER.md`

scripts/mirror-nuwa-to-moments.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+# Mirror github.com/alchaincyf/nuwa-skill (MIT) into Moments private Gitea.
+#
+# Why we mirror: Atlas's claw-boss-distiller depends on nuwa-skill methodology
+# (`references/extraction-framework.md`). For air-gapped client deployments where
+# the client network cannot reach github.com, our private Gitea must serve nuwa.
+#
+# Prereq: a destination repo on git.moments.top must already exist (empty).
+# Default target: https://git.moments.top/Moments.top/nuwa-skill.git
+#
+# Create it once via Gitea UI (the bot account doesn't have org-create scope):
+#   https://git.moments.top/repo/create  →  owner: Moments.top  →  name: nuwa-skill  →  empty
+#
+# Usage:
+#   ./scripts/mirror-nuwa-to-moments.sh
+#   ./scripts/mirror-nuwa-to-moments.sh https://git.moments.top/your-org/nuwa-skill.git
+#
+# After first mirror, schedule a weekly cron to re-mirror upstream changes.
+
+set -euo pipefail
+
+UPSTREAM="https://github.com/alchaincyf/nuwa-skill.git"
+TARGET="${1:-https://git.moments.top/Moments.top/nuwa-skill.git}"
+WORKDIR="$(mktemp -d -t nuwa-mirror-XXXXXX)"
+
+echo "Mirror $UPSTREAM → $TARGET"
+echo "Workdir: $WORKDIR"
+
+# Bare clone preserves all branches/tags/refs
+git clone --mirror "$UPSTREAM" "$WORKDIR"
+cd "$WORKDIR"
+
+# Push everything (--mirror = all refs incl. delete on remote)
+git push --mirror "$TARGET"
+
+cd /
+rm -rf "$WORKDIR"
+
+echo "✅ Mirror complete: $TARGET"
+echo
+echo "Next: update skills/claw-boss-distiller/upstream/README.md to point"
+echo "      air-gapped deployments at the Moments mirror instead of GitHub."

skills/DESCRIPTION.md

@@ -0,0 +1,36 @@
+---
+description: >
+  Atlas (总助 Claw) — boss-perspective project execution radar. Six sub-skills
+  that turn raw email into a 3D map of "people × project × customer", surface
+  what's slipping, and write a 5-minute daily Brief in the boss's voice.
+  Compatible with Hermes Agent (agentskills.io standard) and OpenClaw runtime.
+metadata:
+  hermes:
+    category: atlas
+    tags: [boss-radar, project-tracking, gtd, raci, 9-box, customer-health, ona]
+---
+
+# Atlas Sub-Skills
+
+Six skills work in pipeline:
+
+```
+[email-parser] → [project-tracker] → [people-observer] → [customer-radar]
+                          ↓                  ↓                 ↓
+                          └──────────────────┴─────────────────┘
+                                            ↓
+                                  [boss-distiller]   (quarterly)
+                                            ↓
+                                   [report-writer]   (daily/weekly/monthly)
+```
+
+| Skill | Job |
+|-------|-----|
+| `claw-email-parser` | Fetch + extract canonical Email JSON |
+| `claw-project-tracker` | Cluster threads → projects, GTD/RACI, state |
+| `claw-people-observer` | 9-Box / BARS / CCAR / ONA — never assigns ratings |
+| `claw-customer-radar` | CHS / VoC / churn — fires Andon alerts |
+| `claw-boss-distiller` | nuwa-skill derivative, distill boss mental model |
+| `claw-report-writer` | Daily / Weekly / Monthly briefs in boss's voice |
+
+See each `claw-*/SKILL.md` for inputs/outputs, judgment rules, and failure modes.

skills/claw-boss-distiller/SKILL.md

@@ -1,6 +1,18 @@
 ---
 name: claw-boss-distiller
 description: Distill the boss's mental model from outgoing email corpus into a 5-layer boss_skill.md (Expression DNA / Mental Models / Decision Heuristics / Anti-Patterns / Honest Boundaries). Derived from nuwa-skill (MIT). Drafts only — boss confirms every change.
+version: 0.1.0
+author: Moments / Atlas team (adapted from alchaincyf/nuwa-skill, MIT)
+license: MIT
+metadata:
+  hermes:
+    category: atlas
+    tags: [mental-model, distillation, nuwa, boss-os, atlas]
+    related_skills: [claw-email-parser, claw-report-writer]
+    upstream_dependency: alchaincyf/nuwa-skill
+    boundaries:
+      - all_new_rules_default_to_status_draft
+      - never_apply_changes_without_explicit_boss_confirm
 ---
 
 # claw-boss-distiller

skills/claw-customer-radar/SKILL.md

@@ -1,6 +1,16 @@
 ---
 name: claw-customer-radar
 description: Compute Customer Health Score, sentiment trend, RFM, decision-chain map, churn signals. Fire Andon alerts when health crosses thresholds.
+version: 0.1.0
+author: Moments / Atlas team
+license: MIT
+metadata:
+  hermes:
+    category: atlas
+    tags: [customer-success, chs, voc, churn, rfm, andon, atlas]
+    related_skills: [claw-email-parser, claw-project-tracker, claw-report-writer]
+    boundaries:
+      - cross_customer_data_isolation
 ---
 
 # claw-customer-radar

skills/claw-email-parser/SKILL.md

@@ -1,6 +1,14 @@
 ---
 name: claw-email-parser
 description: Wraps the email-extractor MCP tool; orchestrates fetch → extract → write canonical Email JSON. The thin LLM layer that decides what to do with low-confidence extractions.
+version: 0.1.0
+author: Moments / Atlas team
+license: MIT
+metadata:
+  hermes:
+    category: atlas
+    tags: [email, imap, extraction, atlas, intake]
+    related_skills: [claw-project-tracker, claw-people-observer, claw-customer-radar]
 ---
 
 # claw-email-parser

skills/claw-people-observer/SKILL.md

@@ -1,6 +1,18 @@
 ---
 name: claw-people-observer
 description: Compute observation cards for each person — project density, response time, BARS behavior tags, CCAR critical incidents, ONA centrality, 9-Box performance dimension. NEVER assigns potential or final ratings — those are boss-only.
+version: 0.1.0
+author: Moments / Atlas team
+license: MIT
+metadata:
+  hermes:
+    category: atlas
+    tags: [people-analytics, 9-box, bars, ccar, ona, atlas]
+    related_skills: [claw-email-parser, claw-project-tracker, claw-report-writer]
+    boundaries:
+      - never_assign_9box_potential
+      - never_assign_topgrading_grade
+      - never_recommend_pip_or_termination
 ---
 
 # claw-people-observer

skills/claw-project-tracker/SKILL.md

@@ -1,6 +1,14 @@
 ---
 name: claw-project-tracker
 description: Cluster threads into project entities, apply GTD/RACI logic, judge project state (Active/Stalled/Completed/Someday/Dropped) with rule citations.
+version: 0.1.0
+author: Moments / Atlas team
+license: MIT
+metadata:
+  hermes:
+    category: atlas
+    tags: [project-management, gtd, raci, atlas, boss-perspective]
+    related_skills: [claw-email-parser, claw-people-observer, claw-customer-radar, claw-report-writer]
 ---
 
 # claw-project-tracker

skills/claw-report-writer/SKILL.md

@@ -1,6 +1,18 @@
 ---
 name: claw-report-writer
 description: Render Daily Brief / Weekly Rollup / Monthly Customer Health / one-time Inventory in Amazon WBR style with rule citations. Speaks in the boss's voice (Layer 1 of boss_skill.md).
+version: 0.1.0
+author: Moments / Atlas team
+license: MIT
+metadata:
+  hermes:
+    category: atlas
+    tags: [reporting, wbr, daily-brief, atlas, boss-voice]
+    related_skills: [claw-project-tracker, claw-people-observer, claw-customer-radar, claw-boss-distiller]
+    delivery_channels: [email, feishu, dingtalk, wechat]
+    boundaries:
+      - send_only_to_boss
+      - cite_every_judgment
 ---
 
 # claw-report-writer