math_with_judge:
  resources_servers:
    math_with_judge:
      entrypoint: app.py
      judge_model_server:
        type: responses_api_models
        name: policy_model
      judge_responses_create_params: {input: []}
      should_use_judge: false
      domain: math
      verified: false
      description: OpenCode agent harness on OpenMathReasoning math dataset with math-verify
      value: Improve model math capabilities in the OpenCode agent harness

math_with_judge_opencode_agent:
  responses_api_agents:
    opencode_agent:
      entrypoint: app.py
      resources_server:
        type: resources_servers
        name: math_with_judge
      concurrency: 8
      command: opencode
      model: nvinf/nvidia/qwen/qwen3-next-80b-a3b-instruct
      opencode_version: 1.17.11
      thinking: false
      timeout: 600
      system_prompt: |
        You are a precise math assistant with a bash tool. Run python3 to compute or verify numeric
        results instead of doing them by hand, then state the final answer as \boxed{answer}.
      opencode_config:
        permission:
          bash: allow
          edit: allow
          webfetch: allow
        provider:
          nvinf:
            npm: "@ai-sdk/openai-compatible"
            options:
              baseURL: ${policy_base_url}
              apiKey: ${policy_api_key}
            models:
              nvidia/qwen/qwen3-next-80b-a3b-instruct: {}
      datasets:
      - name: example
        type: example
        jsonl_fpath: resources_servers/math_with_judge/data/example.jsonl