pinchbench_agent:
  responses_api_agents:
    pinchbench:
      entrypoint: app.py
      # One sandbox per task (the isolation boundary, mirrors swe_agents/harbor). Each
      # sandbox runs its OWN in-sandbox OpenClaw gateway — the only mode (at openclaw
      # 2026.6.5 `openclaw agent` needs a gateway to persist transcripts).
      openclaw_mode: gateway
      gateway_token: pinchbench-local
      # Per-task sandbox via Gym's provider-neutral Sandbox API (PR #1377). Provider is
      # config-selected: apptainer (Slurm/HPC, shown) or opensandbox (cluster). Build a
      # .sif from the image with setup_scripts/build_image.sh --apptainer; pass its path
      # (or a docker://<image> ref) via +sandbox_image=... at run time.
      # Defaults are fine: the .sif rootfs is read-only, but the provider bind-mounts a
      # writable, per-sandbox-isolated host dir at mount_point (/sandbox) — run_task.sh
      # writes the skill copy, OpenClaw $HOME, $TMPDIR and the benchmark run-root there.
      sandbox_provider:
        apptainer: {}
      sandbox_spec:
        image: ${sandbox_image}
        ready_timeout_s: 600
        resources:
          cpu: 4
          memory_mib: 8192
      task_timeout_s: 1800
      # Policy model OpenClaw runs against (streaming-capable endpoint; secrets
      # supplied via CLI/env overrides at run time — never committed).
      model_base_url: ${model_base_url}
      model_api_key: ${model_api_key}
      model_name: ${model_name}
      # Judge for hybrid / llm_judge tasks (OpenAI-compatible endpoint).
      judge_model: ${judge_model}
      judge_base_url: ${judge_base_url}
      judge_api_key: ${judge_api_key}
      web_search_provider: brave
      brave_api_key: ${brave_api_key}
      max_concurrent: 2
      max_tokens: 16384
      context_window: 131072
      timeout_multiplier: 3.0
      datasets:
      - name: example
        type: example
        jsonl_fpath: responses_api_agents/pinchbench/data/example.jsonl