OpenCode tiered model router

An OpenCode plugin that judges each new human prompt with Qwen and routes the entire agent turn to the least capable configured model tier. Tool-loop continuations remain on the selected model without another judge request.

The defaults are:

• model-router/auto as the trigger model
• easy → sonarllm-dogfooding/Qwen3.6-Sonar
• complex → portkey/gpt-5.6-sol, variant high
• judge errors, timeouts, and invalid responses → complex

• Model switching risks

Reference the plugin from opencode.json and define its three providers. The Portkey provider intentionally uses @ai-sdk/openai and an environment-backed credential.

{
  "plugin": [
    ["file:///absolute/path/to/opencode-model-router/src/index.ts", {
      "diagnostics": { "echo": true }
    }]
  ],
  "provider": {
    "model-router": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "http://127.0.0.1:1/v1", "apiKey": "unused" },
      "models": { "auto": { "name": "Automatic tier" } }
    },
    "sonarllm-dogfooding": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "SonarLLM Dogfooding",
      "options": {
        "baseURL": "https://sonarllm-dogfooding.aws-dev.sonarsource.com/v1"
      },
      "models": {
        "Qwen3.6-Sonar": {
          "name": "Qwen3.6-Sonar",
          "limit": { "context": 260000, "output": 8192 }
        }
      }
    },
    "portkey": {
      "npm": "@ai-sdk/openai",
      "options": {
        "baseURL": "https://api.portkey.ai/v1",
        "apiKey": "{env:PORTKEY_OPENAI_API_KEY}"
      },
      "models": {
        "gpt-5.6-sol": {
          "name": "gpt-5.6-sol",
          "reasoning": true,
          "variants": { "high": { "reasoningEffort": "high" } }
        }
      }
    }
  }
}

Select model-router/auto to route a turn. Selecting any other model manually bypasses the plugin.

Each routed turn writes judge request, judge response, and either the selected tier or fallback error to OpenCode's log under the opencode-model-router service. Request logs contain the user's bounded prompt text, but never request headers or environment-backed credentials.

The default template sets diagnostics.echo to true for concise judge request and response summaries. The selected model or fallback notification is always shown. Interactive OpenCode displays native toasts; opencode run prints one-line messages to stderr. Full requests and raw responses remain in the log.

The selected tier handles the current turn, while subsequent prompts remain on model-router/auto and receive a fresh routing decision.

Plugin options accept a custom judge, trigger, tier list, and fallback:

bun install
bun run typecheck
bun test

The opt-in live test needs internal-network access, OpenCode 1.18.16 or newer, and PORTKEY_OPENAI_API_KEY. It creates isolated XDG config/data directories, uses the read-only plan agent, and retains sanitized diagnostics under artifacts/.

RUN_LIVE_E2E=1 bun run e2e

Phase 1 deliberately does not retry Qwen worker failures or recover silent worker hangs beyond the nine-second judge timeout.