# Leaderboard-admissible Terminal-Bench 2.1 run.
#
# Keep the dataset at the exact digest pinned by the official leaderboard and
# leave environment, verifier, timeouts, and resources unset so Harbor uses the
# canonical task definitions and default execution policy.
jobs_dir: .nanocodex/harbor/jobs

agents:
  - import_path: harbor_adapter.agent:NanocodexAgent
    model_name: openai/gpt-5.6-sol
    env:
      OPENAI_API_KEY: ${OPENAI_API_KEY}
    kwargs:
      effort: high
      web_search: false
      subagents: false
      # Code Mode requires Node. Canonical TB 2.1 images are Debian/Ubuntu
      # derivatives but do not preinstall it, so provision this agent runtime
      # during Harbor's ordinary agent-install phase.
      install_node: true

datasets:
  - name: terminal-bench/terminal-bench-2-1
    ref: sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a
