Edict Multi Agent Orchestration

1---2name: edict-multi-agent-orchestration3description: Install and use the Edict (三省六部) multi-agent orchestration system with 12 specialized AI agents, real-time kanban dashboard, and audit trails4triggers:5  - set up edict multi agent system6  - configure sansheng liubu agents7  - install edict orchestration8  - use openclaw multi agent dashboard9  - set up AI agent pipeline with review gates10  - configure edict kanban board11  - how to use edict agent workflow12  - deploy three provinces six ministries AI system13---14 15# Edict (三省六部) Multi-Agent Orchestration16 17> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.18 19Edict implements a 1400-year-old Tang Dynasty governance model as an AI multi-agent architecture. Twelve specialized agents form a checks-and-balances pipeline: Crown Prince (triage) → Zhongshu (planning) → Menxia (review/veto) → Shangshu (dispatch) → Six Ministries (parallel execution). Built on [OpenClaw](https://openclaw.ai), it provides a real-time React kanban dashboard, full audit trails, and per-agent LLM configuration.20 21---22 23## Architecture Overview24 25```26You (Emperor) → taizi (triage) → zhongshu (plan) → menxia (review/veto)27             → shangshu (dispatch) → [hubu|libu|bingbu|xingbu|gongbu|libu2] (execute)28             → memorial (result archived)29```30 31**Key differentiator vs CrewAI/AutoGen**: Menxia (门下省) is a mandatory quality gate — it can veto and force rework before tasks reach executors.32 33---34 35## Prerequisites36 37- [OpenClaw](https://openclaw.ai) installed and running38- Python 3.9+39- Node.js 18+ (for React dashboard build)40- macOS or Linux41 42---43 44## Installation45 46### Quick Demo (Docker — no OpenClaw needed)47 48```bash49# x86/amd64 (Ubuntu, WSL2)50docker run --platform linux/amd64 -p 7891:7891 cft0808/sansheng-demo51 52# Apple Silicon / ARM53docker run -p 7891:7891 cft0808/sansheng-demo54 55# Or with docker-compose (platform already set)56docker compose up57```58 59Open http://localhost:789160 61### Full Installation62 63```bash64git clone https://github.com/cft0808/edict.git65cd edict66chmod +x install.sh && ./install.sh67```68 69The install script automatically:70- Creates all 12 agent workspaces (taizi, zhongshu, menxia, shangshu, hubu, libu, bingbu, xingbu, gongbu, libu2, zaochao, legacy-compat)71- Writes SOUL.md role definitions to each agent workspace72- Registers agents and permission matrix in `openclaw.json`73- Symlinks shared data directories across all agent workspaces74- Sets `sessions.visibility all` for inter-agent message routing75- Syncs API keys across all agents76- Builds React frontend77- Initializes data directory and syncs official stats78 79### First-time API Key Setup80 81```bash82# Configure API key on first agent83openclaw agents add taizi84# Then re-run install to propagate to all agents85./install.sh86```87 88---89 90## Running the System91 92```bash93# Terminal 1: Data refresh loop (keeps kanban data current)94bash scripts/run_loop.sh95 96# Terminal 2: Dashboard server97python3 dashboard/server.py98 99# Open dashboard100open http://127.0.0.1:7891101```102 103---104 105## Key Commands106 107### OpenClaw Agent Management108 109```bash110# List all registered agents111openclaw agents list112 113# Add/configure an agent114openclaw agents add <agent-name>115 116# Check agent status117openclaw agents status118 119# Restart gateway (required after config changes)120openclaw gateway restart121 122# Send a message/edict to the system123openclaw send taizi "帮我分析一下竞争对手的产品策略"124```125 126### Dashboard Server127 128```python129# dashboard/server.py — serves on port 7891130# Built-in: React frontend + REST API + WebSocket updates131python3 dashboard/server.py132 133# Custom port134PORT=8080 python3 dashboard/server.py135```136 137### Data Scripts138 139```bash140# Sync official (agent) statistics141python3 scripts/sync_officials.py142 143# Update kanban task states144python3 scripts/kanban_update.py145 146# Run news aggregation147python3 scripts/fetch_news.py148 149# Full refresh loop (runs all scripts in sequence)150bash scripts/run_loop.sh151```152 153---154 155## Configuration156 157### Agent Model Configuration (`openclaw.json`)158 159```json160{161  "agents": {162    "taizi": {163      "model": "claude-3-5-sonnet-20241022",164      "workspace": "~/.openclaw/workspaces/taizi"165    },166    "zhongshu": {167      "model": "gpt-4o",168      "workspace": "~/.openclaw/workspaces/zhongshu"169    },170    "menxia": {171      "model": "claude-3-5-sonnet-20241022",172      "workspace": "~/.openclaw/workspaces/menxia"173    },174    "shangshu": {175      "model": "gpt-4o-mini",176      "workspace": "~/.openclaw/workspaces/shangshu"177    }178  },179  "gateway": {180    "port": 7891,181    "sessions": {182      "visibility": "all"183    }184  }185}186```187 188### Per-Agent Model Hot-Switching (via Dashboard)189 190Navigate to **⚙️ Models** panel → select agent → choose LLM → Apply. Gateway restarts automatically (~5 seconds).191 192### Environment Variables193 194```bash195# API keys (set before running install.sh or openclaw)196export ANTHROPIC_API_KEY="sk-ant-..."197export OPENAI_API_KEY="sk-..."198 199# Optional: Feishu/Lark webhook for notifications200export FEISHU_WEBHOOK_URL="https://open.feishu.cn/open-apis/bot/v2/hook/..."201 202# Optional: news aggregation203export NEWS_API_KEY="..."204 205# Dashboard port override206export DASHBOARD_PORT=7891207```208 209---210 211## Agent Roles Reference212 213| Agent | Role | Responsibility |214|-------|------|----------------|215| `taizi` | 太子 Crown Prince | Triage: chat → auto-reply, edicts → create task |216| `zhongshu` | 中书省 | Planning: decompose edict into subtasks |217| `menxia` | 门下省 | **Review/Veto**: quality gate, can reject and force rework |218| `shangshu` | 尚书省 | Dispatch: assign subtasks to ministries |219| `hubu` | 户部 Ministry of Revenue | Finance, data analysis tasks |220| `libu` | 礼部 Ministry of Rites | Communication, documentation tasks |221| `bingbu` | 兵部 Ministry of War | Strategy, security tasks |222| `xingbu` | 刑部 Ministry of Justice | Review, compliance tasks |223| `gongbu` | 工部 Ministry of Works | Engineering, technical tasks |224| `libu2` | 吏部 Ministry of Personnel | HR, agent management tasks |225| `zaochao` | 早朝官 | Morning briefing aggregator |226 227### Permission Matrix (who can message whom)228 229```python230# Defined in openclaw.json — enforced by gateway231PERMISSIONS = {232    "taizi":    ["zhongshu"],233    "zhongshu": ["menxia"],234    "menxia":   ["zhongshu", "shangshu"],  # can veto back to zhongshu235    "shangshu": ["hubu", "libu", "bingbu", "xingbu", "gongbu", "libu2"],236    # ministries report back up the chain237    "hubu":     ["shangshu"],238    "libu":     ["shangshu"],239    "bingbu":   ["shangshu"],240    "xingbu":   ["shangshu"],241    "gongbu":   ["shangshu"],242    "libu2":    ["shangshu"],243}244```245 246---247 248## Task State Machine249 250```python251# scripts/kanban_update.py enforces valid transitions252VALID_TRANSITIONS = {253    "pending":     ["planning"],254    "planning":    ["reviewing", "pending"],      # zhongshu → menxia255    "reviewing":   ["dispatching", "planning"],   # menxia approve or veto256    "dispatching": ["executing"],257    "executing":   ["completed", "failed"],258    "completed":   [],259    "failed":      ["pending"],  # retry260}261 262# Invalid transitions are rejected — no silent state corruption263```264 265---266 267## Real Code Examples268 269### Send an Edict Programmatically270 271```python272import subprocess273import json274 275def send_edict(message: str, agent: str = "taizi") -> dict:276    """Send an edict to the Crown Prince for triage."""277    result = subprocess.run(278        ["openclaw", "send", agent, message],279        capture_output=True,280        text=True281    )282    return {"stdout": result.stdout, "returncode": result.returncode}283 284# Example edicts285send_edict("分析本季度用户增长数据，找出关键驱动因素")286send_edict("起草一份关于产品路线图的对外公告")287send_edict("审查现有代码库的安全漏洞")288```289 290### Read Kanban State291 292```python293import json294from pathlib import Path295 296def get_kanban_tasks(data_dir: str = "data") -> list[dict]:297    """Read current kanban task state."""298    tasks_file = Path(data_dir) / "tasks.json"299    if not tasks_file.exists():300        return []301    with open(tasks_file) as f:302        return json.load(f)303 304def get_tasks_by_status(status: str) -> list[dict]:305    tasks = get_kanban_tasks()306    return [t for t in tasks if t.get("status") == status]307 308# Usage309executing = get_tasks_by_status("executing")310completed = get_tasks_by_status("completed")311print(f"In progress: {len(executing)}, Done: {len(completed)}")312```313 314### Update Task Status (with validation)315 316```python317import json318from pathlib import Path319from datetime import datetime, timezone320 321VALID_TRANSITIONS = {322    "pending":     ["planning"],323    "planning":    ["reviewing", "pending"],324    "reviewing":   ["dispatching", "planning"],325    "dispatching": ["executing"],326    "executing":   ["completed", "failed"],327    "completed":   [],328    "failed":      ["pending"],329}330 331def update_task_status(task_id: str, new_status: str, data_dir: str = "data") -> bool:332    """Update task status with state machine validation."""333    tasks_file = Path(data_dir) / "tasks.json"334    tasks = json.loads(tasks_file.read_text())335 336    task = next((t for t in tasks if t["id"] == task_id), None)337    if not task:338        raise ValueError(f"Task {task_id} not found")339 340    current = task["status"]341    allowed = VALID_TRANSITIONS.get(current, [])342 343    if new_status not in allowed:344        raise ValueError(345            f"Invalid transition: {current} → {new_status}. "346            f"Allowed: {allowed}"347        )348 349    task["status"] = new_status350    task["updated_at"] = datetime.now(timezone.utc).isoformat()351    task.setdefault("history", []).append({352        "from": current,353        "to": new_status,354        "timestamp": task["updated_at"]355    })356 357    tasks_file.write_text(json.dumps(tasks, ensure_ascii=False, indent=2))358    return True359```360 361### Dashboard REST API Client362 363```python364import urllib.request365import json366 367BASE_URL = "http://127.0.0.1:7891/api"368 369def api_get(endpoint: str) -> dict:370    with urllib.request.urlopen(f"{BASE_URL}{endpoint}") as resp:371        return json.loads(resp.read())372 373def api_post(endpoint: str, data: dict) -> dict:374    payload = json.dumps(data).encode()375    req = urllib.request.Request(376        f"{BASE_URL}{endpoint}",377        data=payload,378        headers={"Content-Type": "application/json"},379        method="POST"380    )381    with urllib.request.urlopen(req) as resp:382        return json.loads(resp.read())383 384# Read dashboard data385tasks    = api_get("/tasks")386agents   = api_get("/agents")387sessions = api_get("/sessions")388news     = api_get("/news")389 390# Trigger task action391api_post("/tasks/pause",  {"task_id": "task-123"})392api_post("/tasks/cancel", {"task_id": "task-123"})393api_post("/tasks/resume", {"task_id": "task-123"})394 395# Switch model for an agent396api_post("/agents/model", {397    "agent": "zhongshu",398    "model": "gpt-4o-2024-11-20"399})400```401 402### Agent Health Check403 404```python405import json406from pathlib import Path407from datetime import datetime, timezone, timedelta408 409def check_agent_health(data_dir: str = "data") -> dict[str, str]:410    """411    Returns health status for each agent.412    🟢 active   = heartbeat within 2 min413    🟡 stale    = heartbeat 2-10 min ago414    🔴 offline  = heartbeat >10 min ago or missing415    """416    heartbeats_file = Path(data_dir) / "heartbeats.json"417    if not heartbeats_file.exists():418        return {}419 420    heartbeats = json.loads(heartbeats_file.read_text())421    now = datetime.now(timezone.utc)422    status = {}423 424    for agent, last_beat in heartbeats.items():425        last = datetime.fromisoformat(last_beat)426        delta = now - last427        if delta < timedelta(minutes=2):428            status[agent] = "🟢 active"429        elif delta < timedelta(minutes=10):430            status[agent] = "🟡 stale"431        else:432            status[agent] = "🔴 offline"433 434    return status435 436# Usage437health = check_agent_health()438for agent, s in health.items():439    print(f"{agent:12} {s}")440```441 442### Custom SOUL.md (Agent Personality)443 444```markdown445<!-- ~/.openclaw/workspaces/gongbu/SOUL.md -->446# 工部尚书 · Minister of Works447 448## Role449You are the Minister of Works (工部). You handle all technical,450engineering, and infrastructure tasks assigned by Shangshu Province.451 452## Rules4531. Always break technical tasks into concrete, verifiable steps4542. Return structured results: { "status": "...", "output": "...", "artifacts": [] }4553. Flag blockers immediately — do not silently fail4564. Estimate complexity: S/M/L/XL before starting457 458## Output Format459Always respond with valid JSON. Include a `summary` field ≤ 50 chars460for kanban display.461```462 463---464 465## Dashboard Panels466 467| Panel | URL Fragment | Key Features |468|-------|-------------|--------------|469| Kanban | `#kanban` | Task columns, heartbeat badges, filter/search, pause/cancel/resume |470| Monitor | `#monitor` | Agent health cards, task distribution charts |471| Memorials | `#memorials` | Completed task archive, 5-stage timeline, Markdown export |472| Templates | `#templates` | 9 preset edict templates with parameter forms |473| Officials | `#officials` | Token usage ranking, activity stats |474| News | `#news` | Daily tech/finance briefing, Feishu push |475| Models | `#models` | Per-agent LLM switcher (hot reload ~5s) |476| Skills | `#skills` | View/add agent skills |477| Sessions | `#sessions` | Live OC-* session monitor |478| Court | `#court` | Multi-agent discussion around a topic |479 480---481 482## Common Patterns483 484### Pattern 1: Parallel Ministry Execution485 486```python487# Shangshu dispatches to multiple ministries simultaneously488# Each ministry works independently; shangshu aggregates results489edict = "竞品分析：研究TOP3竞争对手的产品、定价、市场策略"490 491# Zhongshu splits into subtasks:492# hubu  → pricing analysis493# libu  → market communication analysis494# bingbu → competitive strategy analysis495# gongbu → technical feature comparison496 497# All execute in parallel; shangshu waits for all 4, then aggregates498```499 500### Pattern 2: Menxia Veto Loop501 502```python503# If menxia rejects zhongshu's plan:504# menxia → zhongshu: "子任务拆解不完整，缺少风险评估维度，请补充"505# zhongshu revises and resubmits to menxia506# Loop continues until menxia approves507# Max iterations configurable in openclaw.json: "max_review_cycles": 3508```509 510### Pattern 3: News Aggregation + Push511 512```python513# scripts/fetch_news.py → data/news.json → dashboard #news panel514# Optional Feishu push:515import os, json, urllib.request516 517def push_to_feishu(summary: str):518    webhook = os.environ["FEISHU_WEBHOOK_URL"]519    payload = json.dumps({520        "msg_type": "text",521        "content": {"text": f"📰 天下要闻\n{summary}"}522    }).encode()523    req = urllib.request.Request(524        webhook, data=payload,525        headers={"Content-Type": "application/json"}526    )527    urllib.request.urlopen(req)528```529 530---531 532## Troubleshooting533 534### `exec format error` in Docker535 536```bash537# Force platform on x86/amd64538docker run --platform linux/amd64 -p 7891:7891 cft0808/sansheng-demo539```540 541### Agents not receiving messages542 543```bash544# Ensure sessions visibility is set to "all"545openclaw config set sessions.visibility all546openclaw gateway restart547# Or re-run install.sh — it sets this automatically548./install.sh549```550 551### API key not propagated to all agents552 553```bash554# Re-run install after configuring key on first agent555openclaw agents add taizi  # configure key here556./install.sh               # propagates to all agents557```558 559### Dashboard shows stale data560 561```bash562# Ensure run_loop.sh is running563bash scripts/run_loop.sh564 565# Or trigger manual refresh566python3 scripts/sync_officials.py567python3 scripts/kanban_update.py568```569 570### React frontend not built571 572```bash573# Requires Node.js 18+574cd dashboard/frontend575npm install && npm run build576# server.py will then serve the built assets577```578 579### Invalid state transition error580 581```python582# kanban_update.py enforces the state machine583# Check current status before updating:584tasks = get_kanban_tasks()585task = next(t for t in tasks if t["id"] == "your-task-id")586print(f"Current: {task['status']}")587print(f"Allowed next: {VALID_TRANSITIONS[task['status']]}")588```589 590### Gateway restart after model change591 592```bash593# After editing openclaw.json models section594openclaw gateway restart595# Wait ~5 seconds for agents to reconnect596```597 598---599 600## Project Structure601 602```603edict/604├── install.sh              # One-command setup605├── openclaw.json           # Agent registry + permissions + model config606├── scripts/607│   ├── run_loop.sh         # Continuous data refresh daemon608│   ├── kanban_update.py    # State machine enforcement609│   ├── sync_officials.py   # Agent stats aggregation610│   └── fetch_news.py       # News aggregation611├── dashboard/612│   ├── server.py           # stdlib-only HTTP + WebSocket server (port 7891)613│   ├── dashboard.html      # Fallback single-file dashboard614│   └── frontend/           # React 18 source (builds to server.py assets)615├── data/                   # Shared data (symlinked into all workspaces)616│   ├── tasks.json617│   ├── heartbeats.json618│   ├── news.json619│   └── officials.json620├── workspaces/             # Per-agent workspace roots621│   ├── taizi/SOUL.md622│   ├── zhongshu/SOUL.md623│   └── ...624└── docs/625    ├── task-dispatch-architecture.md626    └── getting-started.md627```