oc-agent-router

An OpenCode plugin that routes each new task subagent to one of your configured models. It asks Jev for a structured choice immediately before OpenCode creates the child session.

Compatibility: This plugin supports OpenCode v2 only. OpenCode v1 is not supported.

Install

opencode plugin add oc-agent-router

Set TYPESAFE_API_KEY in the environment that starts OpenCode, then configure the plugin in ~/.config/opencode/opencode.json or your project opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    [
      "oc-agent-router",
      {
        "models": [
          "openai/gpt-5.4",
          "anthropic/claude-sonnet-4-6",
          "google/gemini-3-pro"
        ],
        "agents": ["general", "explore"],
        "confidenceThreshold": 0.8
      }
    ]
  ]
}

All configured models must already be available in OpenCode. Restart OpenCode after installing or changing its configuration.

TYPESAFE_API_KEY is read only from the environment of the OpenCode process. Do not put it in opencode.json.

Local Testing

This repository includes .opencode/opencode.json, which loads the built dist/ plugin directory. Before starting OpenCode from this directory, build the plugin and update its models to model IDs enabled in your OpenCode configuration.

npm run build
TYPESAFE_API_KEY=... opencode

Ask OpenCode to delegate work with a subagent. A new child session will receive the model selected by Jev; resumed child sessions and calls with an explicit model are intentionally left unchanged. Restart OpenCode after changing the plugin source, build output, or .opencode/opencode.json.

To leave child sessions on their configured or inherited model, omit TYPESAFE_API_KEY or set a threshold higher than Jev's returned confidence.

Options

How It Works

The plugin calls POST https://api.typesafe.ai/v1/systemone with a Choice question whose only choices are your allowed models, then supplies the selected model to OpenCode's native subagent tool immediately before it creates the child session. The original agent ID, system prompt, permissions, task, and display name remain unchanged.

Child-session resumes (sessionID) are deliberately not rerouted, so a resumed session keeps its original model. Calls that explicitly select a model are also left unchanged. If Jev is unavailable, returns an invalid choice, or is below confidenceThreshold, the plugin leaves the model unset so OpenCode uses the configured agent model or parent session model.

The routing request sends only the subagent ID, short task description, and task prompt. It does not send session history, tool outputs, or API credentials.

For example, a task that asks general to implement a parser sends the following payload. Its Authorization header is Bearer $TYPESAFE_API_KEY, which comes from the process environment rather than any config file.

{
  "model": "jev-latest",
  "state": {
    "task": "Implement the parser for the new configuration format.",
    "description": "Implement parser",
    "requested_agent": "general"
  },
  "questions": {
    "model": {
      "type": "choice",
      "instructions": "Choose the configured model most suitable for completing this OpenCode subagent task. Prefer a capable model for implementation, debugging, and complex reasoning; prefer an efficient model for focused exploration or simple tasks.",
      "criteria": {
        "openai/gpt-5.4": "Use the configured OpenCode model openai/gpt-5.4.",
        "anthropic/claude-sonnet-4-6": "Use the configured OpenCode model anthropic/claude-sonnet-4-6.",
        "google/gemini-3-pro": "Use the configured OpenCode model google/gemini-3-pro."
      }
    }
  }
}

Every new child session is eligible for routing, including configured and Markdown-defined subagents. Resumes (sessionID) retain their existing model.