multi-agent

面向 Codex、Cursor、Claude Code、OpenClaw、Hermes、VS Code 的多 Agent 协作协议、Skill 与 MCP 设计。 鏈粨搴撳綋鍓嶅寘鍚涓€鐗?multi-agent-coding Skill锛屼互鍙婂悗缁?MCP/IDE 鎻掍欢鐨勬帴鍙ｈ鏍兼枃妗ｃ€傚畠鐨勭洰鏍囦笉鏄仛涓€涓け鎺х殑 agent swarm锛岃€屾槸璁╁涓?agent 鍦ㄥ鏉傚伐绋嬩换鍔′腑鎸夆€滀换鍔″崱銆佹潈闄愯竟鐣屻€佽瘎瀹°€侀獙璇併€佹渶缁堥泦鎴愨€濈殑鏂瑰紡鍗忎綔銆?

椤圭洰瀹氫綅

Skill = 鍗忎綔鏂规硶璁哄拰鎻愮ず璇嶈绾?MCP = 浠诲姟銆佺姸鎬併€佸鎵广€佽瘎瀹″拰瀹¤鐨勫伐鍏峰悗绔?IDE Plugin = 鍥惧舰鍖栦换鍔￠潰鏉裤€丳rompt 鐢熸垚鍣ㄥ拰鏈湴闆嗘垚鍏ュ彛

褰撳墠宸插疄鐜扮殑鏄?Skill 鍜屾枃妗ｈ鏍硷紱MCP Server 涓?IDE 鎻掍欢灞炰簬鍚庣画瀹炵幇鏂瑰悜銆?

multi-agent-coding Skill

multi-agent-coding 鏄竴涓?prompt-guided 鐨勫 Agent 缂栫爜鍗忎綔 Skill锛岀敤浜庢寚瀵?agent 鍦ㄥ鏉備唬鐮佷换鍔′腑瀹屾垚锛?

闇€姹傜悊瑙?鈫?鐜妫€鏌?鈫?浠诲姟鎷嗚В 鈫?骞惰璋冪爺 鈫?鍙楁帶瀹炵幇 鈫?Diff 瀹¤ 鈫?澶氳瑙掕瘎瀹?鈫?楠岃瘉 鈫?鏈€缁堜氦浠?```

鏍稿績瑙掕壊锛?
- `Main Agent`锛氫富鎺э紝璐熻矗瑙勫垝銆佸垎娲俱€侀泦鎴愩€侀獙璇佸拰鏈€缁堜氦浠樸€?- `Explorer`锛氬彧璇昏皟鐮斾唬鐮併€佹灦鏋勩€佹祴璇曘€佺害瀹氬拰椋庨櫓銆?- `Worker`锛氬湪 `allowed_paths` 鍐呭彈鎺у疄鐜帮紝涓嶅緱瓒婃潈銆?- `Reviewer`锛氬彧璇昏瘎瀹★紝涓嶆敼浠ｇ爜锛屽彲浣跨敤 `ssrd` 绛夎瘎瀹?Skill銆?- `Verifier`锛氳繍琛屾垨鎻忚堪娴嬭瘯銆佹瀯寤恒€乴int銆佸鐜版楠ゃ€?- `Integrator`锛氶€氬父鐢?Main Agent 鎵紨锛岃礋璐ｅ悎骞跺拰鏈€缁堜竴鑷存€с€?
## 鑳借В鍐充粈涔堥棶棰?
- 澶嶆潅浠诲姟鎷嗕笉娓咃紝瀹规槗杈瑰仛杈圭寽銆?- 澶у瀷浠ｇ爜搴撲笂涓嬫枃澶ぇ锛岄渶瑕佸涓?Explorer 骞惰璋冪爺銆?- 鍓嶅悗绔€佸悗绔€佹祴璇曘€佹枃妗ｇ瓑浠诲姟鍙互鍒嗗伐锛屼絾瀹规槗浜掔浉瑕嗙洊銆?- Worker 浣跨敤鍏朵粬 Skill 鏃剁己灏戞巿鏉冭竟鐣屻€?- 澶氫釜 reviewer 鐨勬剰瑙侀渶瑕佸幓閲嶃€佹帓搴忓拰姹囨€汇€?- 鏈€缁堜氦浠樺墠缂哄皯 Diff 瀹¤鍜岄獙璇侀棴鐜€?
## 閲嶈璇存槑

杩欎釜 Skill 鏄崗浣滆绾︼紝涓嶆槸瀹夊叏娌欑锛屼篃涓嶆槸鑷姩缂栨帓鍣ㄣ€傚畠涓嶈兘鐪熸寮哄埗闅旂鏂囦欢绯荤粺銆佺綉缁溿€丟it 鎴栬繘绋嬫潈闄愩€?
瀹冩彁渚涚殑鏄細

- 浠€涔堟椂鍊欏簲璇ヤ娇鐢ㄥ Agent銆?- 濡備綍鍐欎换鍔″崱銆?- Worker 鑳芥敼鍝簺璺緞銆?- Worker 浠€涔堟椂鍊欏彲浠ヤ娇鐢ㄥ叾浠?Skill銆?- Reviewer 濡備綍鍙璇勫銆?- Main Agent 濡備綍瀹¤ Diff 鍜屾眹鎬荤粨鏋溿€?
鐪熸鐨勭姸鎬佺鐞嗐€佹潈闄愭鏌ャ€佷换鍔￠潰鏉垮拰瀹¤宸ュ叿锛屽缓璁€氳繃鍚庣画 MCP Server 鍜?IDE 鎻掍欢瀹炵幇銆?
## 鐩綍缁撴瀯

```text
SKILL.md                         # Codex/OpenClaw 椋庢牸 Skill 涓绘枃浠?agents/openai.yaml               # Codex UI 鍏冩暟鎹?templates/task-card.md           # 瀛愪换鍔″崱妯℃澘
templates/result-report.md       # 瀛?Agent 缁撴灉鎶ュ憡妯℃澘
templates/final-delivery.md      # 鏈€缁堜氦浠樻ā鏉?checklists/                      # 鍚敤澶?Agent銆佺幆澧冦€佹潈闄愩€佸畨鍏ㄣ€丏iff 瀹¤妫€鏌ユ竻鍗?examples/                        # feature / bugfix / review 绀轰緥娴佺▼
docs/clients.md                  # Codex/Cursor/Claude/OpenClaw/Hermes/VS Code 鏀寔妯″瀷
docs/mcp-format.md               # MCP 宸ュ叿銆佽祫婧愩€丳rompt 鏍煎紡瑙勬牸
adapters/openclaw/               # OpenClaw/Her standalone skill + scripts

OpenClaw Adapter (v1 Mission Control)

For OpenClaw/Her, install only adapters/openclaw/ as the skill openclaw-multi-agent. No MCP required.

File	Purpose
SKILL.md	Session workflow, gates, tool mapping
scripts/create_task_cards.py	Task cards + ownership.json + status.json + run-plan.json
scripts/update_task_status.py	Sync gates, update task status, summarize run
scripts/audit_worker_output.py	Scope audit + optional audit JSON under audits/
scripts/run_local_demo.py	Deterministic local demo + --self-check
QUICKSTART.md	Fresh-checkout install and validation steps
examples/favorite-feature.yaml	Sample multi-module task definition

Local state (gitignored): .codex-multi-agent/ with tasks/, results/, findings/, approvals/, audits/, status.json.

Workflow gates: Explorer → Worker → Reviewer → Verifier → scope audit → final delivery. Main runs explicit commands from each task card; scripts update state only.

Quick start:

cd adapters/openclaw
python3 scripts/validate_all.py
# See adapters/openclaw/QUICKSTART.md for full Golden Path

Roadmap: docs/roadmap.md (v1 scripts → v2 MCP → v3 IDE panel). MCP contract: docs/mcp-format.md.

OpenClaw 用户可直接安装 adapters/openclaw/，无需阅读整个根仓库。

Supported clients (v1 adapters)

All four adapters share the OpenClaw mission-control core (create_task_cards.py, update_task_status.py, audit_worker_output.py, verify_workspace.py, .codex-multi-agent/). Each client folder is a thin launch + docs layer — no duplicated gate logic.

Client	Status	Notes
OpenClaw / Her	Production v1 ✅	Reference implementation; full gate demo — QUICKSTART
Cursor	Usable + real dogfood ✅	Requires agent + tmux; tmux mode is fire-and-forget — use --foreground to debug — QUICKSTART
Codex	Usable + real dogfood ⚠	Requires codex CLI + writable sandbox (CODEX_SANDBOX, default workspace-write) — QUICKSTART
Claude Code	Contract validated ⚠	Local claude --print subject to quota/429; prefer ACP path in OpenClaw — QUICKSTART
Hermes / VS Code	Docs + scaffolds	Extension scaffolds under ide/extensions/; Hermes adapter planned
MCP coordinator server (multi-agent-coordinator)	v1 in ✅	Stdio MCP over .codex-multi-agent/ — README
Mission-control task panel	v1 in ✅	Local web UI — README

Launchers use pipefail + post-run checks (non-trivial result Markdown, JSON file or extracted sidecar). They exit non-zero and emit "ok": false when the external CLI fails or logs quota/error patterns — not when tee alone succeeds.

Cross-adapter launcher (dependency-free):

python3 scripts/run_multi_agent.py --runtime cursor|codex|claude-code|openclaw --task-card .codex-multi-agent/tasks/T002-worker-backend.md

Project-wide validation:

make validate
# or
python3 scripts/validate_all_adapters.py
./scripts/ci_smoke.sh

Run everything (full smoke suite):

bash scripts/full_validate.sh

Runs all adapter self-checks, OpenClaw demo, both bench harnesses, MCP/IDE checks, tools, memory, and dogfood fixtures. Continues on failure and exits non-zero if any step failed.

See docs/clients.md for the client matrix and docs/roadmap.md for v1 status and caveats.

Tools layer (P0)

Dependency-free stdlib wrappers under tools/ for Workers and Verifiers:

Tool	Purpose
git_tool.py	git status, diff, --name-only changed files
test_runner_tool.py	Discover/run pytest, npm test, pnpm test
lint_tool.py	Best-effort ruff / flake8 / eslint / prettier / mypy / pyright / golangci-lint / cargo clippy / rustfmt
shell_tool.py	Sandboxed shell with allowlist/denylist
repo_index_tool.py	File listing + grep (rg or Python fallback)

Each tool exposes --help, JSON-in/JSON-out via stdin, and --self-check. Task cards declare tools_used:; audits warn on undeclared tool references.

Memory layer (P0)

• MEMORY.md — append-only project decisions (no secrets)
• AGENTS.md — repo-wide role conventions
• adapters/openclaw/scripts/memory_log.py — --append, --from-run (wired into --summarize)

Latest MEMORY.md tail is injected into task card context at generation time.

Local benchmark (P1)

bench/ — lightweight SWE-style cases (not upstream SWE-bench). Self-check:

python3 bench/run_bench.py --self-check --dry-runtime
python3 bench/swebench-lite/run_swebench_lite.py --self-check

See bench/swebench-lite/README.md for SWE-bench Lite-shaped multi-file cases.

MCP enhancements (P1)

• list_framework_tools MCP tool mirrors tools/ entries
• record_finding entries (source: mcp) merge with reviewer sync — not clobbered
• Per-client MCP config snippets: mcp/multi-agent-coordinator/clients/

Examples

• Feature / bugfix / review flows: examples/
• Case study (FizzBuzz, all gates green): examples/case-study-fizzbuzz/
• Case study (multi-file Flask-shaped CLI): examples/case-study-flask-cli/

IDE extension scaffolds (Phase 1)

Not published — starting points for embedding the task panel:

• ide/extensions/vscode/ — VS Code webview → local panel HTTP
• ide/extensions/cursor/ — Cursor-flavored variant
• ide/extensions/hermes/ — future Hermes adapter notes

瑙﹀彂鏉′欢

閫傚悎瑙﹀彂 multi-agent-coding 鐨勬儏鍐碉細

• 鐢ㄦ埛鏄庣‘璇粹€滅敤澶氫釜 agent鈥濃€滃紑 subagent鈥濃€滃苟琛屽鐞嗏€濃€滃涓?reviewer 璇勫鈥濄€?- 浠诲姟娑夊強澶氫釜鐙珛妯″潡鎴栨妧鏈眰锛屼緥濡傚墠绔€佸悗绔€佹暟鎹簱銆佹祴璇曘€?- 闇€瑕佸厛璋冪爺鍐嶅疄鐜帮紝鍐嶈瘎瀹″拰楠岃瘉銆?- 浠ｇ爜搴撹緝澶э紝鍗?agent 涓婁笅鏂囧帇鍔涢珮銆?- 鐢ㄦ埛瑕佹眰澶氳瑙掕瘎瀹°€佸畨鍏ㄥ鏌ャ€佹灦鏋勫鏌ャ€佸鏉?bug 瀹氫綅銆?- Worker 闇€瑕佷娇鐢ㄥ叾浠?Skill锛屼絾蹇呴』鏈夋潈闄愯竟鐣屻€? 涓嶅缓璁Е鍙戠殑鎯呭喌锛?
• 鍗曟枃浠跺皬鏀瑰姩銆?- 鏄庣‘鐨勫皬 bug銆?- 绠€鍗曡В閲婁唬鐮併€?- 鏀规枃妗堛€佹敼閰嶇疆瀛楁绛夎交閲忎换鍔°€?- 娌℃湁瀹夊叏鐨勪换鍔℃媶鍒嗚竟鐣屻€?

Quick Path 涓?Multi-Agent Path

灏忎换鍔¤蛋 Quick Path锛?

Intake 鈫?Plan Lite 鈫?Implement/Answer 鈫?Review Lite 鈫?Verify 鈫?Deliver

澶嶆潅浠诲姟璧?Multi-Agent Path锛?

Intake 鈫?Environment Check 鈫?Task Graph 鈫?Explorer Fan-out 鈫?Synthesis 鈫?Scoped Worker 鈫?Diff Audit 鈫?Reviewer 鈫?Verifier 鈫?Deliver

Worker 浣跨敤鍏朵粬 Skill 鐨勮鍒?

Worker 鍙互浣跨敤鍏朵粬 Skill锛屼絾涓嶈兘鍊?Skill 瓒婃潈銆? Worker 浣跨敤 Skill 蹇呴』婊¤冻锛?

• Skill 鍦ㄤ换鍔″崱鐨?may_use_skills 涓紝鎴?Main Agent 鏄庣‘鎵瑰噯銆?- Skill 涓庡綋鍓?Worker 鐩爣鐩存帴鐩稿叧銆?- 涓嶆墿澶?allowed_paths銆?- 涓嶈繍琛?blocked_commands銆?- 涓嶈闂?secret銆乼oken銆乧ookie銆丼SH key銆佷簯鍑嵁绛夋晱鎰熸枃浠躲€?- 涓嶆敼鍙?Worker 鐨勮鑹诧紝渚嬪浠庡疄鐜拌€呭彉鎴愰儴缃茶€呮垨 Git 鎿嶄綔鑰呫€? 濡傛灉涓嶆弧瓒虫潯浠讹紝Worker 蹇呴』鍋滄骞舵彁浜?Skill Use Request銆? 绀轰緥锛氬涓?agent 浣跨敤 ssrd 璇勫鏂规鏃讹紝搴斿垱寤哄涓彧璇?Reviewer锛岃€屼笉鏄啓鏉冮檺 Worker锛?

mode: review
role: Reviewer
may_use_skills:
  - ssrd
write_permission: false
may_spawn_subagents: false

璺ㄥ鎴风鏀寔

鏈粨搴撶洰鏍囨敮鎸侊細

• Codex
• Cursor
• Claude Code
• OpenClaw
• Hermes
• VS Code

涓嶅悓瀹㈡埛绔殑 agent 鍚姩鏂瑰紡銆佹彃浠舵満鍒躲€丮CP 閰嶇疆涓嶅悓锛屼絾鍏变韩鐨勫崗璁簲璇ヤ繚鎸佷竴鑷达細

• Task Card
• Result Report
• Role Permission
• Skill Use Approval
• Review Finding
• Scope Audit
• Final Delivery

璇︾粏璁捐瑙侊細

• docs/clients.md
• docs/mcp-format.md

MCP 璁捐鏂瑰悜

鍚庣画鍙互瀹炵幇涓€涓?multi-agent-coordinator-mcp锛岀敤浜庢彁渚涳細

• create_task
• list_tasks
• get_task
• update_task_status
• record_result
• check_path_allowed
• record_touched_paths
• request_skill_use
• approve_skill_use
• record_finding
• summarize_review
• audit_scope
• generate_final_report

MCP 璐熻矗鐘舵€佸拰宸ュ叿鍖栨搷浣滐紝Skill 璐熻矗鍗忎綔娴佺▼鍜岃涓鸿竟鐣屻€?

IDE 鎻掍欢鏂瑰悜

鍚庣画鍙互鏀寔 VS Code / Cursor 鎻掍欢锛屾彁渚涳細

• Task Board锛氭煡鐪?pending/running/blocked/completed 浠诲姟銆?- Create Task锛氬浘褰㈠寲鍒涘缓 Explorer/Worker/Reviewer/Verifier 浠诲姟鍗°€?- Findings View锛氬睍绀?reviewer findings锛屽苟鏀寔璺宠浆鏂囦欢琛屽彿銆?- Skill Approval Center锛氱鐞?Worker/Reviewer 浣跨敤鍏朵粬 Skill 鐨勫鎵广€?- Scope Audit锛氫竴閿鏌?Worker 鏄惁瓒婄晫銆佹槸鍚︽湁鍐茬獊銆?- Prompt Generator锛氫负 Codex銆丆ursor銆丆laude Code銆丱penClaw銆丠ermes 鐢熸垚閫傞厤 prompt銆?- Final Report锛氫竴閿敓鎴愭渶缁堜氦浠樻憳瑕併€?

浣跨敤鏂瑰紡

灏嗘湰鐩綍澶嶅埗鍒?Codex Skill 鐩綍锛屾垨浣滀负椤圭洰鍐?Skill 浣跨敤锛岀劧鍚庢樉寮忚皟鐢細

Use $multi-agent-coding to coordinate this coding task with scoped roles, review, and verification.

涓枃绀轰緥锛?

浣跨敤 $multi-agent-coding锛屾妸杩欎釜鍔熻兘鎷嗘垚 Explorer銆乄orker銆丷eviewer 鍜?Verifier 鏉ュ仛銆?```

```text
浣跨敤 $multi-agent-coding锛屽紑澶氫釜 Reviewer锛屽苟璁╁畠浠敤 ssrd 璇勫杩欎釜鏂规銆?```

```text
浣跨敤 $multi-agent-coding锛屽垽鏂繖涓换鍔℃槸鍚﹀€煎緱寮€澶氫釜 agent锛涘鏋滀笉鍊煎緱锛屽氨璧?Quick Path銆?```

## 褰撳墠鐘舵€?
- 宸插畬鎴愶細`multi-agent-coding` Skill v0.1銆?- 宸插畬鎴愶細浠诲姟鍗°€佺粨鏋滄姤鍛娿€佹渶缁堜氦浠樻ā鏉裤€?- 宸插畬鎴愶細鐜銆佹潈闄愩€佸畨鍏ㄣ€丏iff 瀹¤妫€鏌ユ竻鍗曘€?- 宸插畬鎴愶細Codex/Cursor/Claude Code/OpenClaw/Hermes/VS Code 鏀寔妯″瀷鏂囨。銆?- 宸插畬鎴愶細MCP 鏍煎紡鏂囨。銆?- 寰呭疄鐜帮細MCP Server銆?- 寰呭疄鐜帮細VS Code / Cursor IDE 鎻掍欢銆?
## 璺嚎鍥?
```text
v0.1 Skill锛氬崗浣滄祦绋嬨€佷换鍔″崱銆佹潈闄愯竟鐣屻€佽瘎瀹″拰楠岃瘉妯℃澘
v0.2 MCP Server锛氫换鍔＄姸鎬併€佸鎵广€乫inding銆乻cope audit銆乫inal report
v0.3 VS Code / Cursor Plugin锛氫换鍔￠潰鏉裤€丳rompt 鐢熸垚鍣ㄣ€佸璁¤鍥?v0.4 Client Adapters锛欳laude Code銆丱penClaw銆丠ermes 閫傞厤
v0.5 Worktree / PR / CI锛氬苟琛屽垎鏀€丳R review銆丆I 澶辫触鍥炴祦