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: pi agent harness on OpenMathReasoning math dataset with math-verify
      value: Improve model math capabilities in the pi agent harness

math_with_judge_pi_agent:
  responses_api_agents:
    pi_agent:
      entrypoint: app.py
      resources_server:
        type: resources_servers
        name: math_with_judge
      concurrency: 8
      command: pi
      model: nvinf/nvidia/qwen/qwen3-next-80b-a3b-instruct
      pi_version: 0.80.2
      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}.
      models_config:
        providers:
          nvinf:
            baseUrl: ${policy_base_url}
            api: openai-completions
            apiKey: ${policy_api_key}
            compat:
              supportsDeveloperRole: false
              supportsReasoningEffort: false
            models:
            - id: nvidia/qwen/qwen3-next-80b-a3b-instruct
              reasoning: false
      datasets:
      - name: example
        type: example
        jsonl_fpath: resources_servers/math_with_judge/data/example.jsonl