OpenCode Task Router

An OpenCode plugin that classifies a development task with a local Ollama model and recommends the right cost tier and agent to use.

It is designed to keep simple work on free local models and reserve paid models for tasks that actually need them.

What You Get

• A route_task tool for task-to-model routing recommendations
• Cost-tier recommendations: free, cheap, moderate, expensive
• Agent guidance based on the recommended tier
• Lightweight learning from your past routing decisions

Prerequisites

1. OpenCode installed
2. Ollama installed and running
3. The classifier model pulled locally:

ollama serve
ollama pull qwen3:8b

If your global OpenCode config uses enabled_providers, include ollama there too:

{
  "enabled_providers": ["github-copilot", "ollama"]
}

Install From NPM

Add the plugin package to your OpenCode config:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-task-router"]
}

Then restart OpenCode. OpenCode installs npm plugins automatically with Bun at startup.

Minimal Usage

Ask OpenCode to call the tool directly:

Use the route_task tool to analyze this task: refactor the authentication module

Typical output:

## Task Routing Recommendation

| Factor | Assessment |
|--------|-----------|
| Complexity | **moderate** |
| Context needs | **medium** |
| Cost tier | **moderate** |

**Reasoning:** Multi-file refactoring usually needs broader codebase context.

Optional Full Setup

The npm package guarantees the route_task tool.

If you also want the /route shortcut, a local free-tier agent, and an example Ollama provider config, copy the example files from examples/opencode/ into your project:

examples/opencode/
├── opencode.json
└── .opencode/
    ├── agents/
    │   └── local-worker.md
    └── commands/
        └── route.md

Suggested mapping:

• examples/opencode/.opencode/commands/route.md -> .opencode/commands/route.md
• examples/opencode/.opencode/agents/local-worker.md -> .opencode/agents/local-worker.md
• examples/opencode/opencode.json -> merge into your project opencode.json

How It Works

1. route_task sends your task description to a local Ollama classifier
2. The classifier returns task complexity, context estimate, and cost tier
3. The plugin recommends an agent and model tier
4. When the session goes idle, the plugin logs what was recommended to .opencode/router-history.jsonl
5. Recent history is reused as calibration on future routing decisions

Configuration

The plugin currently defaults to:

const OLLAMA_BASE_URL = "http://localhost:11434"
const CLASSIFIER_MODEL = "qwen3:8b"
const MAX_HISTORY_EXAMPLES = 20

These live in src/index.ts.

Local Development

Install dependencies and build the package:

npm install
npm run build
npm test

For local OpenCode development in this repo, the project plugin entrypoint at .opencode/plugins/task-router.ts re-exports the package source from src/index.ts.

The smoke test automates what is practical without depending on a real OpenCode session:

• builds the package
• creates a tarball
• installs the tarball into a temporary directory
• imports the installed package
• executes route_task with a mocked Ollama response
• verifies the history file is written

The unit tests cover the internal routing logic directly:

• cost-tier inference from provider/model ids
• classifier response parsing and normalization
• prompt construction with calibration history
• history file read/write helpers
• plugin recommendation, retry, and idle logging behavior

CI And Publishing

This repo includes GitHub Actions for:

• CI on pushes to main and pull requests
• npm publishing from version tags like v0.1.0
• manual publish dry runs through workflow_dispatch

Publishing can be fully automated through GitHub Actions using npm trusted publishing with GitHub OIDC.

See RELEASE.md for the release checklist and tag-based publishing flow.

Troubleshooting

Cannot connect to Ollama

Make sure Ollama is running and the model is installed:

ollama serve
ollama pull qwen3:8b

Plugin loads but routing falls back to moderate

That usually means the classifier returned invalid output. The plugin retries once, then falls back to a safe default.

No history file created yet

The history file is only written after a routing recommendation and when the session goes idle.

License

MIT