Develop project from group chat — by Claude Code
接 AI 进群里开发项目 — 飞书（Lark）

IM group  →  Monitor  →  Intent  →  Claude acts  →  Chat reply
_{one plugin, any project, zero backend.}

Inside Claude Code in your target project, run these in order:

/plugin marketplace add WaterTian/cc-bot
/plugin install cc-bot@cc-bot
/reload-plugins
/cc-bot:setup

• Line 1 — register the plugin source (GitHub repo)

• Line 2 — download to ~/.claude/plugins/cache/. CC will pop a scope selector; pick one:

  选项	适用场景	Recommendation
  Install for you (user scope)	You plan to use cc-bot across many projects	✅ 日常推荐
  Install for all collaborators (project scope)	Team project, want teammates to auto-get cc-bot on clone	✅ 团队项目
  Install for you, in this repo only (local scope)	First-time try, don't pollute other projects	✅ 快速试用

• Line 3 — hot-reload plugin list without restarting CC (new since CC 2.1.x; faster than /exit + re-open)

• Line 4 — enter the interactive setup wizard (lark-cli auto-install, OAuth login, chat picker via AskUserQuestion cards, config auto-fill)

When a new release is announced, run these in each project using cc-bot:

/plugin marketplace update cc-bot    # refresh manifest (detects new version)
/plugin update cc-bot@cc-bot         # fetch new version into cache
/reload-plugins                      # apply in current session
/cc-bot:doctor                       # verify: first line prints "cc-bot v<new-version>"

• Before upgrading — run /cc-bot:doctor first; it compares installed version with the latest GitHub release and prints the upgrade hint if drifted, plus flags stale permissions or profile issues.
• After upgrading — run /cc-bot:doctor again to confirm the new version is active; or run /cc-bot:start and the group online notification will show cc-bot v<new-version> 已上线. Setup's greeting + completion lines also carry v<version> (since v0.1.4) for the same reason.

CC's plugin cache is version-indexed (~/.claude/plugins/cache/cc-bot/cc-bot/<version>/). After each upgrade the Monitor launch path points to a new version dir, so CC would otherwise re-prompt for permission.

Since v0.1.3, /cc-bot:setup auto-writes a wildcard rule to <project>/.claude/settings.local.json. As of v0.1.11 the wildcard covers the entire runtime/*.js directory (was just poll.js), so future runtime tools (e.g. check-image-size.js) no longer require re-running setup:

Bash(node C:/Users/*/.claude/plugins/cache/cc-bot/cc-bot/*/runtime/*.js *)

You won't be prompted for any cc-bot runtime tool on future version upgrades.

For v0.1.2 or earlier installs (pre-auto): after you update, re-run /cc-bot:setup — step 9 is idempotent and will append the rule. Or add it manually.

Migrating from v0.1.10 or earlier (had the older runtime/poll.js --project * rule): re-run /cc-bot:setup once — it appends the new runtime/*.js * wildcard while keeping the old single-file rule intact for backwards compatibility.

Since v0.1.6 — /cc-bot:setup step 9 additionally registers a UserPromptSubmit / Stop hook pair into ~/.claude/settings.json so main-window typing automatically pauses group-message handling (poll.js skips emit + sends a one-shot placeholder, randomly picked from a 14-phrase pool since v0.1.9). After upgrading to v0.1.6 you must re-run /cc-bot:setup once to activate the hook — the three §Updating commands above only pull new code; they don't touch user-global settings.json.

git clone https://github.com/WaterTian/cc-bot.git

cd /your/project
claude --plugin-dir /absolute/path/to/cc-bot

Setup prints a version banner on start (cc-bot v<X.Y.Z> setup — <project>, since v0.1.4) then runs through these steps — fully interactive via AskUserQuestion cards, no blind typing:

1. Detect lark-cli — auto-install via npm i -g @larksuite/cli if missing
2. OAuth login — guide you through Lark Open Platform app creation (scope checklist provided), then browser Device Flow login
3. Pick target chat — list bot's chats via AskUserQuestion card; or one-click create a new chat (bot auto-joins, you become owner)
4. Auto-detect IDs — bot_app_id / admin_open_id pulled from lark-cli auth list, zero manual entry
5. Write config — generate .cc-bot/profiles/active.json + state.json + .gitignore
6. Register statusline shim — tees stdin JSON to hud-stdin.json (for bot's HUD intent) + cc-hud rendering (if installed, for status bar)
7. Register Monitor permission (v0.1.3+) — append a wildcard rule to <project>/.claude/settings.local.json, so cc-bot version upgrades never re-prompt for Monitor launch permission

Every step is idempotent — rerun /cc-bot:setup anytime, it skips what's already done.

Then /cc-bot:start (or just say "开bot" / "start bot" in the main session) — bot comes online in ≤ 5s. Group notification shows cc-bot v<X.Y.Z> 已上线.

Main session ── Monitor(persistent) ── node poll.js ── every 30s: IMAdapter.listRecentMessages()
                                                    ├─ dedupe via state.last_processed_time + poll.emitted
                                                    └─ stdout: NEW_MSG|msg_id|sender|text|ts
                                                               ↓ Monitor → notification
                                                         main session → intent routing → adapter.sendText

CC's statusLine ── cc-bot shim ── write hud-stdin.json (for bot's HUD intent)
                                └─ tee cc-hud (optional, for status bar rendering)

Main session also accepts natural-language triggers (开bot / 关bot / switch to xxx).

Each project has its own .cc-bot/profiles/active.json. The intents dict maps natural-language intents to action descriptions Claude executes when the group triggers them. Typical examples (you pick the keys and describe the actions freely):

{
  "intents": {
    "deploy": "Run `bash scripts/deploy.sh production`, report stdout tail",
    "run_tests": "Run `npm test`, report pass/fail counts + first failure trace",
    "query_logs": "Use mcp__cloudbase__logs to fetch last 20 error logs, summarize",
    "build_preview": "Run `npm run build:preview`, upload artifact to <paths.bot_temp_abs>, reply with link"
  }
}

Works for any project type — Web / mini-program / Node service / Python data pipeline / mobile app. If you can script it, you can intent-route it.

• Claude Code — uses Skill / Monitor / TaskStop / AskUserQuestion tools
• lark-cli — npm i -g @larksuite/cli + lark-cli auth login (setup wizard will guide this)
• Shell — Windows: Git Bash required (cmd.exe / PowerShell mangle special characters in argv); macOS / Linux: system bash works out of the box
• Optional: cc-hud — install for prettier status bar (/plugin install cc-hud@WaterTian-cc-hud); cc-bot shim tees it automatically

在目标项目里运行 /cc-bot:setup，开场一行打印当前版本（cc-bot v<X.Y.Z> setup — <project>，v0.1.4 起），然后交互式向导会：

1. 检测 lark-cli — 未装自动 npm i -g @larksuite/cli
2. OAuth 登录引导 — 带你去飞书开放平台建应用（附必需 scope 清单），完成浏览器 Device Flow 登录
3. 选目标群 — 用 AskUserQuestion 卡片列 bot 所在群，或一键新建（bot 自动入群、你成为群主）
4. 自动探测 ID — bot_app_id / admin_open_id 从 lark-cli auth list 直接取，不用手填
5. 写配置 — 生成 .cc-bot/profiles/active.json + state.json + .gitignore
6. 注册 statusline shim — 落盘 stdin JSON（给 bot 用）+ 可选透传 cc-hud（渲染状态栏）
7. 注册 Monitor 通配权限（v0.1.3+）— 向 <project>/.claude/settings.local.json append 通配规则，cc-bot 版本升级不再弹 Monitor 启动权限询问

然后 /cc-bot:start（或主会话直接说「开bot」）。群里上线通知首行是 cc-bot v<X.Y.Z> 已上线。

One plugin, many projects in parallel — zero cross-contamination.

cc-bot ships with Lark. Adding WeCom / DingTalk / Slack / Discord / etc:

1. Add adapters/<im>.js extending IMAdapter (see adapters/base.js) — implement listRecentMessages / sendText / sendImage / downloadResource / getUser
2. Add factory branch in runtime/poll.js: if (im.type === '<im>') { ... }
3. Set im.type in profile.active.json to the new IM name
4. Add skills/<im>-bot/SKILL.md or extend existing one

Repo ships a pre-commit scanner that blocks real IM IDs (cli_* / ou_* / oc_* / om_*) / real-name blocklist / api-secret patterns from entering commits. One-time setup — see CLAUDE.md §Git 提交隐私防护.

_{MIT License © Water}