Emperor — OpenCode Multi-Agent Collaboration Plugins

Two OpenCode plugins for multi-agent collaboration in software development.

Plugin	Agents	Philosophy	Best For
Emperor (三省六部)	11	Governance — checks & balances	Tasks needing rigorous review
Commander	4	Engineering — fast iteration	Most development tasks

---

Commander Plugin

A lightweight, adaptive multi-agent plugin with 4 agents and a single orchestrator pattern. Optimized for speed and tight feedback loops.

Architecture

graph TD
    User["User Task"] --> Lead["Lead<br/>Analyze & Plan"]
    Lead -->|trivial| Lead2["Lead handles directly"]
    Lead -->|simple/standard/complex| Dispatch["Dispatch Waves"]
    Dispatch --> Coder["Coder<br/>Implement"]
    Coder --> Tester["Tester<br/>Verify"]
    Tester -->|❌ fail| Coder
    Tester -->|✅ pass| Done["Wave Complete"]
    Done -->|complex only| Reviewer["Reviewer<br/>Code Audit"]
    Reviewer --> Report["Lead<br/>Final Report"]
    Done -->|not complex| Report

Loading

Four Agents

Agent	Role	Responsibility
Lead	Orchestrator	Explores codebase, creates plans, classifies complexity, summarizes results
Coder	Implementer	Writes code based on subtask specs from Lead
Tester	Verifier	Runs tests, build verification; triggers fix loop on failure
Reviewer	Auditor	Code review for complex tasks only (security, architecture, quality)

Adaptive Complexity

Lead classifies each task automatically:

Complexity	Condition	Flow
trivial	No subtasks	Lead handles directly
simple	1 low-effort subtask	Coder → Tester
standard	Multiple subtasks	Parallel waves of Coder → Tester
complex	High-effort or risky	Parallel waves + Reviewer audit

Coder↔Tester Fix Loop

The core quality mechanism. When Tester fails verification:

1. Coder gets failure context (same session — context accumulates)
2. Coder fixes → Tester re-verifies (same session)
3. Repeat up to maxFixLoops (default: 3)
4. If still failing → mark subtask as failed

Tools

Tool	Description
cmd_task	Create a task for the Commander team
cmd_status	View task status and history
cmd_halt	Emergency stop an active task

Configuration

Create .opencode/commander.json (optional — all settings have defaults):

{
  "agents": {
    "lead": { "model": "anthropic/claude-sonnet-4-20250514" },
    "coder": { "model": "anthropic/claude-sonnet-4-20250514" },
    "tester": { "model": "anthropic/claude-sonnet-4-20250514" },
    "reviewer": { "model": "anthropic/claude-sonnet-4-20250514" }
  },
  "pipeline": {
    "maxFixLoops": 3,
    "enableReviewer": true,
    "sensitivePatterns": ["删除|remove|delete", "production|deploy", "密钥|secret|credential"]
  },
  "store": {
    "dataDir": ".commander"
  }
}

Usage

Switch to the lead agent in OpenCode:

@lead Add user authentication with JWT tokens, refresh mechanism, and role-based access control

Or use the task tool from any agent:

Use cmd_task tool:
  title: "User Authentication"
  content: "Implement JWT auth, token refresh, RBAC"
  priority: "high"

---

Emperor Plugin (三省六部)

Maps ancient China's Three Departments and Six Ministries governance system into a multi-agent architecture. 11 agents with checks and balances for rigorous task execution.

Architecture

graph TD
    User["User's Edict"] --> Taizi["Crown Prince<br/>Task Triage"]
    Taizi --> Jinyiwei["Jinyiwei<br/>Project Reconnaissance"]
    Jinyiwei --> Zhongshu["Zhongshu Province<br/>Planning"]
    Zhongshu --> Menxia["Menxia Province<br/>Review & Veto"]
    Menxia -->|Reject| Zhongshu
    Menxia -->|Approve| Shangshu["Shangshu Province<br/>Execution Dispatch"]
    Shangshu --> Libu["Libu<br/>Architecture"]
    Shangshu --> Hubu["Hubu<br/>Testing"]
    Shangshu --> Libu2["Libu2<br/>API Docs"]
    Shangshu --> Bingbu["Bingbu<br/>Security Testing"]
    Shangshu --> Xingbu["Xingbu<br/>Security Audit"]
    Shangshu --> Gongbu["Gongbu<br/>Infrastructure"]
    Libu --> Memorial["Memorial Summary"]
    Hubu --> Memorial
    Libu2 --> Memorial
    Bingbu --> Memorial
    Xingbu --> Memorial
    Gongbu --> Memorial
    Memorial --> Taizi

Loading

Eleven Agents

Agent	Role	Responsibility
Crown Prince (taizi)	Triage	Receives requests, communicates only with Three Provinces
Jinyiwei (jinyiwei)	Reconnaissance	Scans project code, generates architecture reports
Zhongshu Province (zhongshu)	Planner	Breaks down tasks, outputs structured JSON plans
Menxia Province (menxia)	Reviewer	Reviews plans, can veto and send back
Shangshu Province (shangshu)	Dispatcher	Parallel execution dispatch, progress monitoring
Libu (libu)	Architect	Code architecture, refactoring, type systems
Hubu (hubu)	Tester	Testing and verification (mandatory)
Libu2 (libu2)	API Officer	API design, documentation
Bingbu (bingbu)	Security Officer	Security testing, performance
Xingbu (xingbu)	Auditor	Security audit, compliance (read-only)
Gongbu (gongbu)	Engineer	Build tools, CI/CD, infrastructure

Tools

Tool	Description
emperor_create_edict	Create an edict and start the workflow
emperor_view_memorial	View execution history and results
emperor_halt_edict	Emergency halt an active edict

Configuration

See .opencode/emperor.json for configuration options. Key settings:

• reviewMode: auto / manual / mixed (recommended)
• sensitivePatterns: Keywords triggering manual review
• mandatoryDepartments: Departments required in every plan (default: ["hubu"])
• maxPlanningRetries: Max retries when Menxia rejects a plan

Usage

@taizi Add user authentication with JWT tokens and role-based access control

Built-in Skills

Skill	Description
taizi-reloaded	Enhanced Crown Prince with judgment-execution separation
quick-verify	Forced pre-delivery verification
hubu-tester	Complete verification report template
menxia-reviewer	Code security review

---

Installation

Plugin Registration

Add plugin paths in .opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "./plugins/emperor/index.ts",
    "./plugins/commander/index.ts"
  ]
}

You can enable one or both plugins.

Project Structure

.opencode/
├── opencode.json                        # Plugin registration
├── emperor.json                         # Emperor config (optional)
├── commander.json                       # Commander config (optional)
└── plugins/
    ├── emperor/                         # Emperor plugin (11 agents)
    │   ├── index.ts
    │   ├── types.ts
    │   ├── config.ts
    │   ├── store.ts
    │   ├── agents/prompts.ts
    │   ├── skills/
    │   ├── engine/
    │   │   ├── pipeline.ts
    │   │   ├── recon.ts
    │   │   ├── reviewer.ts
    │   │   └── dispatcher.ts
    │   └── tools/
    │       ├── edict.ts
    │       ├── memorial.ts
    │       └── halt.ts
    └── commander/                       # Commander plugin (4 agents)
        ├── index.ts
        ├── types.ts
        ├── config.ts
        ├── store.ts
        ├── agents/
        │   ├── lead.ts
        │   ├── coder.ts
        │   ├── tester.ts
        │   └── reviewer.ts
        ├── engine/
        │   ├── pipeline.ts
        │   ├── classifier.ts
        │   └── dispatcher.ts
        └── tools/
            ├── task.ts
            ├── status.ts
            └── halt.ts

Tech Stack

• Runtime: Bun
• Language: TypeScript (strict mode)
• Plugin SDK: @opencode-ai/plugin
• Data Persistence: JSON file storage

Publishing

Each plugin is published independently to npm:

# Emperor release
git tag emperor-v0.5.1 && git push --tags

# Commander release
git tag commander-v0.1.0 && git push --tags

Tag pattern triggers CI to build and publish only the corresponding package.

Package	npm
opencode-plugin-emperor	Emperor plugin
opencode-plugin-commander	Commander plugin

---

中文版 | English