Edict (三省六部) Multi-Agent Orchestration

Skill by ara.so — Daily 2026 Skills collection.

Edict 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, it provides a real-time React kanban dashboard, full audit trails, and per-agent LLM configuration.

---

Architecture Overview

You (Emperor) → taizi (triage) → zhongshu (plan) → menxia (review/veto)
             → shangshu (dispatch) → [hubu|libu|bingbu|xingbu|gongbu|libu2] (execute)
             → memorial (result archived)

Key differentiator vs CrewAI/AutoGen: Menxia (门下省) is a mandatory quality gate — it can veto and force rework before tasks reach executors.

---

Prerequisites

• OpenClaw installed and running
• Python 3.9+
• Node.js 18+ (for React dashboard build)
• macOS or Linux

---

Installation

Quick Demo (Docker — no OpenClaw needed)

# x86/amd64 (Ubuntu, WSL2)
docker run --platform linux/amd64 -p 7891:7891 cft0808/sansheng-demo

# Apple Silicon / ARM
docker run -p 7891:7891 cft0808/sansheng-demo

# Or with docker-compose (platform already set)
docker compose up

Open http://localhost:7891

Full Installation

git clone https://github.com/cft0808/edict.git
cd edict
chmod +x install.sh && ./install.sh

The install script automatically:

• Creates all 12 agent workspaces (taizi, zhongshu, menxia, shangshu, hubu, libu, bingbu, xingbu, gongbu, libu2, zaochao, legacy-compat)
• Writes SOUL.md role definitions to each agent workspace
• Registers agents and permission matrix in openclaw.json
• Symlinks shared data directories across all agent workspaces
• Sets sessions.visibility all for inter-agent message routing
• Syncs API keys across all agents
• Builds React frontend
• Initializes data directory and syncs official stats

First-time API Key Setup

# Configure API key on first agent
openclaw agents add taizi
# Then re-run install to propagate to all agents
./install.sh

---

Running the System

# Terminal 1: Data refresh loop (keeps kanban data current)
bash scripts/run_loop.sh

# Terminal 2: Dashboard server
python3 dashboard/server.py

# Open dashboard
open http://127.0.0.1:7891

---

Key Commands

OpenClaw Agent Management

# List all registered agents
openclaw agents list

# Add/configure an agent
openclaw agents add <agent-name>

# Check agent status
openclaw agents status

# Restart gateway (required after config changes)
openclaw gateway restart

# Send a message/edict to the system
openclaw send taizi "帮我分析一下竞争对手的产品策略"

Dashboard Server

# dashboard/server.py — serves on port 7891
# Built-in: React frontend + REST API + WebSocket updates
python3 dashboard/server.py

# Custom port
PORT=8080 python3 dashboard/server.py

Data Scripts

# Sync official (agent) statistics
python3 scripts/sync_officials.py

# Update kanban task states
python3 scripts/kanban_update.py

# Run news aggregation
python3 scripts/fetch_news.py

# Full refresh loop (runs all scripts in sequence)
bash scripts/run_loop.sh

---

Configuration

Agent Model Configuration (openclaw.json)

{
  "agents": {
    "taizi": {
      "model": "claude-3-5-sonnet-20241022",
      "workspace": "~/.openclaw/workspaces/taizi"
    },
    "zhongshu": {
      "model": "gpt-4o",
      "workspace": "~/.openclaw/workspaces/zhongshu"
    },
    "menxia": {
      "model": "claude-3-5-sonnet-20241022",
      "workspace": "~/.openclaw/workspaces/menxia"
    },
    "shangshu": {
      "model": "gpt-4o-mini",
      "workspace": "~/.openclaw/workspaces/shangshu"
    }
  },
  "gateway": {
    "port": 7891,
    "sessions": {
      "visibility": "all"
    }
  }
}

Per-Agent Model Hot-Switching (via Dashboard)

Navigate to ⚙️ Models panel → select agent → choose LLM → Apply. Gateway restarts automatically (~5 seconds).

Environment Variables

# API keys (set before running install.sh or openclaw)
export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."

# Optional: Feishu/Lark webhook for notifications
export FEISHU_WEBHOOK_URL="https://open.feishu.cn/open-apis/bot/v2/hook/..."

# Optional: news aggregation
export NEWS_API_KEY="..."

# Dashboard port override
export DASHBOARD_PORT=7891

---

Agent Roles Reference

Agent	Role	Responsibility
taizi	太子 Crown Prince	Triage: chat → auto-reply, edicts → create task
zhongshu	中书省	Planning: decompose edict into subtasks
menxia	门下省	Review/Veto: quality gate, can reject and force rework
shangshu	尚书省	Dispatch: assign subtasks to ministries
hubu	户部 Ministry of Revenue	Finance, data analysis tasks
libu	礼部 Ministry of Rites	Communication, documentation tasks
bingbu	兵部 Ministry of War	Strategy, security tasks
xingbu	刑部 Ministry of Justice	Review, compliance tasks
gongbu	工部 Ministry of Works	Engineering, technical tasks
libu2	吏部 Ministry of Personnel	HR, agent management tasks
zaochao	早朝官	Morning briefing aggregator

Permission Matrix (who can message whom)

# Defined in openclaw.json — enforced by gateway
PERMISSIONS = {
    "taizi":    ["zhongshu"],
    "zhongshu": ["menxia"],
    "menxia":   ["zhongshu", "shangshu"],  # can veto back to zhongshu
    "shangshu": ["hubu", "libu", "bingbu", "xingbu", "gongbu", "libu2"],
    # ministries report back up the chain
    "hubu":     ["shangshu"],
    "libu":     ["shangshu"],
    "bingbu":   ["shangshu"],
    "xingbu":   ["shangshu"],
    "gongbu":   ["shangshu"],
    "libu2":    ["shangshu"],
}

---

Task State Machine

# scripts/kanban_update.py enforces valid transitions
VALID_TRANSITIONS = {
    "pending":     ["planning"],
    "planning":    ["reviewing", "pending"],      # zhongshu → menxia
    "reviewing":   ["dispatching", "planning"],   # menxia approve or veto
    "dispatching": ["executing"],
    "executing":   ["completed", "failed"],
    "completed":   [],
    "failed":      ["pending"],  # retry
}

# Invalid transitions are rejected — no silent state corruption

---

Real Code Examples

Send an Edict Programmatically

import subprocess
import json

def send_edict(message: str, agent: str = "taizi") -> dict:
    """Send an edict to the Crown Prince for triage."""
    result = subprocess.run(
        ["openclaw", "send", agent, message],
        capture_output