OpenCode Orchestrator

Autonomous multi-agent plugin for OpenCode

Overview

Multi-agent system that autonomously executes complex tasks. Commander delegates work to specialized agents, manages parallel execution, and maintains shared context.

🏛️ Master Session Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                           👤 USER REQUEST                                   │
│                        "/task Build a REST API"                             │
└─────────────────────────────────────────────────────────────────────────────┘
                                     │
                                     ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                    🎯 MASTER SESSION (Commander Agent)                      │
│                                                                             │
│  ┌───────────────────────────────────────────────────────────────────────┐  │
│  │ 1️⃣ THINK     → Analyze request, assess complexity (L1/L2/L3)         │  │
│  │ 2️⃣ PLAN      → Create .opencode/todo.md via Planner                  │  │
│  │ 3️⃣ DELEGATE  → Spawn Worker Sessions via delegate_task              │  │
│  │ 4️⃣ MONITOR   → Watch .opencode/ for progress, handle completions     │  │
│  └───────────────────────────────────────────────────────────────────────┘  │
│                                                                             │
│  ┌─────────────────┐  ┌─────────────────────────────────────────────────┐  │
│  │  Session State  │  │               4 CONSOLIDATED AGENTS             │  │
│  │  Map<id,state>  │  │  🎯 Commander  📋 Planner                       │  │
│  └─────────────────┘  │  🔨 Worker     ✅ Reviewer                       │  │
│                       └─────────────────────────────────────────────────┘  │
│  ┌─────────────────┐
│  │  Plugin Hooks   │                                                       │
│  │  event          │  ┌─────────────────────────────────────────────────┐  │
│  │  chat.message   │  │ 🔄 SessionRecovery (auto error handling)        │  │
│  │  tool.execute   │  │ 📋 TodoContinuation (auto-resume on idle)       │  │
│  └─────────────────┘  │ 📣 TaskToastManager (consolidated notifications)│  │
│                       └─────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────────────────────┘
                                     │
                        delegate_task (async)
                                     │
           ┌─────────────────────────┼─────────────────────────┐
           ▼                         ▼                         ▼
┌─────────────────────┐  ┌─────────────────────┐  ┌─────────────────────┐
│  ⚡ Worker Session 1 │  │  ⚡ Worker Session 2 │  │  ⚡ Worker Session N │
│  Agent: Planner     │  │  Agent: Worker      │  │  Agent: Reviewer    │
│                     │  │                     │  │                     │
│  • Independent exec │  │  • Independent exec │  │  • Independent exec │
│  • Read/Write       │  │  • Read/Write       │  │  • Read/Write       │
│    .opencode/       │  │    .opencode/       │  │    .opencode/       │
│  • Notify parent    │  │  • Notify parent    │  │  • Notify parent    │
│    on complete      │  │    on complete      │  │    on complete      │
└─────────────────────┘  └─────────────────────┘  └─────────────────────┘
           │                         │                         │
           └─────────────────────────┼─────────────────────────┘
                                     ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                    📁 SHARED CONTEXT (.opencode/)                           │
│                                                                             │
│  .opencode/todo.md     ← Master TODO (Planner creates, Reviewer updates)  │
│  .opencode/context.md  ← Adaptive context (shrinks as progress increases)  │
│  .opencode/docs/       ← Cached docs (Planner/Worker save, auto-expire)    │
│  .opencode/archive/    ← Old context for reference                         │
│                                                                             │
│  ⚡ All sessions read/write this shared workspace                           │
│  📊 Progress is tracked via todo.md checkboxes                              │
└─────────────────────────────────────────────────────────────────────────────┘

Installation

npm install -g opencode-orchestrator

Restart OpenCode after installation.

Usage

🚀 Two Modes of Operation

🎯 Commander Mode - /task (Recommended for Real Work)

Use /task when you need the AI to complete a mission autonomously:

/task "Fix the login bug in the docker-compose environment"
/task "Add dark mode support to the entire app"
/task "Refactor the API to use TypeScript"

What Commander Mode Does:

• ♾️ Runs until sealed — Loops until agent outputs <mission_seal>SEALED</mission_seal>
• 🧠 Anti-Hallucination — Researches docs before coding
• ⚡ Parallel Execution — Up to 50 concurrent agents
• 🔄 Auto-Recovery — Handles errors automatically
• 📊 Triage System — Adapts strategy to complexity (L1/L2/L3)

🎖️ Mission Seal Loop:

/task "mission" → Agent works → Idle? → Seal found? 
                       ↑              │
                       │      No      │ Yes
                       └──────────────┴──→ ✅ Complete

When the agent finishes ALL work, it outputs:

<mission_seal>SEALED</mission_seal>

Control Commands:

• /stop or /cancel — Stop the loop manually
• Max 20 iterations (configurable)

_{/task "mission" triggers full Commander mode with Mission Seal loop}

💬 Chat Mode - Regular Conversation (Simple Q&A)

Just type normally without /task for simple questions:

How do I add a loading spinner?
What's the difference between useState and useReducer?

Chat Mode is just regular conversation — no autonomous execution, no parallel agents, no mission tracking.

💡 Pro Tip: Use /task for anything that requires multiple steps, file changes, or verification. Use Chat Mode for quick questions.

The 7 Agents

🛠️ Available Tools

📊 Resource Guarantees

🧪 Test Coverage

Test Files:  19 passed
Tests:       216 passed
Duration:    ~4.3s

🔄 Reliability Features

Session Recovery

Automatic recovery from common errors:

• tool_result_missing - Tool crash recovery
• thinking_block_order - Thinking block issues
• rate_limit - API rate limiting with backoff
• Max 3 recovery attempts per session

Todo Continuation

• Monitors session.idle events
• 2-second countdown before auto-continuation
• Cancels on user interaction
• Skips if background tasks running

noReply Optimization

• Individual task completion: noReply: true (saves tokens)
• All tasks complete: noReply: false (AI processes results)

Uninstall

npm uninstall -g opencode-orchestrator

🔧 Debugging

Log file location:

# Find log path (macOS uses /var/folders/...)
npm run log
# or:
tail -f "$(node -e 'console.log(require("os").tmpdir())')/opencode-orchestrator.log"

# Windows
# C:\Users\<username>\AppData\Local\Temp\opencode-orchestrator.log

Documentation

• System Architecture — Detailed technical docs
• OpenCode SDK Reference — API usage reference
• Release Notes — Version history
• Troubleshooting

License

MIT License. LICENSE