OpenCode Orchestrator

Multi-Agent Plugin for OpenCode

What is this?

A 5-agent structured architecture that makes weak LLMs perform reliably.

Problem: Models like GLM-4.7 struggle with complex engineering tasks.

Solution: XML-structured prompts + explicit reasoning patterns + evidence-based completion.

Key Features

Structured Prompts — XML tags provide clear boundaries for weak models
Explicit Reasoning — THINK → ACT → OBSERVE → ADJUST pattern
Auto-Fix — Inspector audits AND repairs automatically
Persistent Memory — Context saved to disk, never lost
Parallel Execution — Independent tasks run concurrently
Relentless Loop — Runs until mission complete (no user input needed)

Installation

npm install -g opencode-orchestrator

Restart OpenCode after installation.

Usage

Just Select Commander Agent

Press tab → Select Commander → Type your request!

"Implement user authentication with JWT"

That's it. Commander automatically:

Enters mission mode
Plans tasks in parallel
Executes until complete
Stops when done with evidence

Or Use /task Command

/task "Implement user authentication with JWT"

Both work identically - Commander agent OR /task command.

How It Works

     Select Commander or /task
           │
           ▼
    ┌─────────────┐
    │  COMMANDER  │  ← THINK → ACT → OBSERVE → ADJUST
    └──────┬──────┘
           │
    ┌──────┴──────┐
    ▼             ▼
 ARCHITECT    MEMORY (load context)
    │
    ▼
 ┌──────────────────┐
 │  PARALLEL DAG    │
 │  Group 1: [T1, T2, T3] ← run together
 │  Group 2: [T4] ← wait for group 1
 └──────────────────┘
           │
    ┌──────┴──────┐
    ▼             ▼
 BUILDER       BUILDER  (parallel)
    │             │
    └──────┬──────┘
           ▼
    ┌─────────────┐
    │  INSPECTOR  │  ← AUDIT → [FAIL? → FIX] → PASS
    └──────┬──────┘
           ▼
    ┌─────────────┐
    │   MEMORY    │  ← Save progress
    └──────┬──────┘
           ▼
        ✅ MISSION COMPLETE

Agents (5-Agent Architecture)

Agent	Role	What It Does
Commander	Orchestrator	Controls mission, delegates work, verifies completion
Architect	Planner	Decomposes complex tasks into parallel DAGs
Builder	Developer	Implements code (logic + UI combined)
Inspector	Quality + Fix	Audits code AND auto-fixes problems
Memory	Context	Saves/loads progress across sessions

Why 5 Agents?

Builder = Coder + Visualist (full-stack in one agent)
Inspector = Reviewer + Fixer (auto-fixes on audit failure)
Fewer agents = fewer handoffs = better for weak models

Optimized for Weak Models

Technique	Purpose
XML tags	Clear semantic boundaries
THINK-ACT-OBSERVE-ADJUST	Explicit step-by-step reasoning
Few-shot examples	Correct output format learning
Tables/JSON output	Structured over prose
Evidence required	Concrete verification criteria

Uninstall

npm uninstall -g opencode-orchestrator

The plugin automatically removes itself from OpenCode config.

Troubleshooting

If Commander doesn't appear after install:

npm uninstall -g opencode-orchestrator
npm install -g opencode-orchestrator
# Restart OpenCode

Docs

Architecture & Design — System design, agents, and why it works
Plugin Troubleshooting — Setup issues

License

MIT License. LICENSE

Updates are frequent. Keep your version fresh.