harbor_agent:
  responses_api_agents:
    harbor_agent:
      # Python module entrypoint loaded by NeMo Gym.
      entrypoint: app.py
      domain: agent
      description: Harbor integration for agent harnesses and environments.
      value: Improve models in popular agentic environments supported by Harbor such as Terminus2.
      # Max concurrent requests handled by this agent server process.
      concurrency: 50

      # Dataset routing map.
      # Each alias must define exactly one source:
      # - local_dataset_path
      # - or dataset_name (+ optional dataset_version)
     # Optional per-alias workdir can override global harbor_environment_kwargs.workdir.
      # Requests must provide instance_id as "<dataset_alias>::<task_name>".
      harbor_datasets:
        scientific:
          local_dataset_path: "responses_api_agents/harbor_agent/data/nemotron_terminal_synthetic_tasks/skill_based/mixed/scientific_computing"
          workdir: "/app"

      # Agent – which Harbor agent to use (set name or import_path)
      harbor_agent_name: null
      # Python import path for a custom agent class (e.g. "my_pkg.my_mod:MyAgent").
      harbor_agent_import_path: "responses_api_agents.harbor_agent.custom_agents.terminus_2_nemo_gym:Terminus2NemoGym"

      # Extra kwargs forwarded to the Harbor AgentConfig
      harbor_agent_kwargs:
        # Hard stop on number of turns per episode. Remove to enable unlimited turns.
        max_turns: 20
        # Include reasoning content in chat history for each turn.
        interleaved_thinking: true
        # Keep off for RL so the model is not asked to compress context mid-trajectory.
        enable_summarize: false
        # Required for RL: includes per-token ids/logprobs used downstream by training.
        collect_rollout_details: true
        trajectory_config:
          # Preserve raw JSON model response output.
          raw_content: true
        model_info:
          max_input_tokens: 49152
          max_output_tokens: 49152
          input_cost_per_token: 0.0
          output_cost_per_token: 0.0

      # Environment – which Harbor environment to use (set name or import_path)
      harbor_environment_type: null
      # Use custom Singularity backend for HPC clusters.
      harbor_environment_import_path: "responses_api_agents.harbor_agent.custom_envs.singularity.singularity:SingularityEnvironment"
      harbor_environment_kwargs:
        # Shared host cache for converted .sif images (avoid repeated pulls/conversions across jobs).
        singularity_image_cache_dir: "responses_api_agents/harbor_agent/singularity_cache"

      # Model routing: points agent LLM calls at the NeMo Gym model server alias.
      model_server:
        type: responses_api_models
        name: policy_model

      # Timeout controls (null = defer to task.toml defaults):
      #
      # *_timeout:     Override — replaces the task's timeout entirely.
      #                All tasks get exactly this value.
      # *_max_timeout: Cap — uses the task's own timeout but clamps it
      #                to this maximum. Shorter per-task timeouts are respected.
      # timeout_multiplier: Multiplied after override/cap resolution.
      harbor_agent_override_timeout: null               # Override agent timeout (seconds)
      harbor_agent_max_timeout: null           # Cap agent timeout (seconds)
      harbor_verifier_override_timeout: null            # Override verifier timeout (seconds)
      harbor_verifier_max_timeout: null        # Cap verifier timeout (seconds)
      harbor_timeout_multiplier: null          # Multiplier (null = 1.0)

      # Output root directory for Harbor job artifacts.
      harbor_jobs_dir: "/path/to/Gym/responses_api_agents/harbor_agent/jobs"