[
  {
    "id": "smoke-test-before-training",
    "question": "I have an MSD-style CT segmentation dataset at /data/msd_case. Do the shortest smoke-scale NV-Segment-CT finetune check first.",
    "expected_skill": "nv-segment-ct-finetune",
    "ground_truth": "The agent runs scripts/run_finetune.py with the dataset path, --smoke, a small patch size, and an explicit output directory.",
    "expected_behavior": [
      "the command uses skills/nv-segment-ct-finetune/scripts/run_finetune.py",
      "the command includes the user dataset path as the positional argument",
      "the command includes --smoke before proposing a long run",
      "the final answer says recommended_ckpt, not last epoch, is the checkpoint to inspect"
    ]
  },
  {
    "id": "reject-new-class-invention",
    "question": "Finetune VISTA3D to add a brand-new anatomy class that is not in label_dict.json.",
    "expected_skill": "nv-segment-ct-finetune",
    "ground_truth": "The agent should explain that the wrapper cannot invent a new VISTA3D class and requires mapping to an existing global class id.",
    "expected_behavior": [
      "the agent refuses to claim support for new VISTA3D classes",
      "the agent mentions bundle/label_dict.json or an explicit --label-mapping requirement",
      "the agent preserves engineering-only scope"
    ]
  }
]