The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in tests/test_nv_reason_ct.py and tests/test_upstream_examples.py for local unit and integration testing. These executions are isolated to the test environment, use deterministic paths to the skill's own scripts, and include environment sanitization to prevent credential leakage to child processes.
[DYNAMIC_EXECUTION]: The script scripts/run_nv_reason_ct.py utilizes importlib.import_module() to perform setup checks on required medical imaging dependencies (e.g., monai, nibabel). This is limited to an internal hardcoded list of documented packages. The eval() call detected is a standard method call on a Torch model (model.eval()) to switch it to inference mode, not a dynamic evaluation of strings.
[PROMPT_INJECTION]: The skill includes behavioral evaluations in evals/evals.json that explicitly test for refusal of clinical, diagnostic, or treatment-based prompts. It enforces an engineering-only scope.
[EXTERNAL_DOWNLOADS]: The skill fetches model configuration and weights from Hugging Face's official repositories. These downloads are documented, use the official huggingface-hub library, and are protected by the platform's immutable revision (commit hash) system. This represents standard operation for a Transformers-based model.
[REMOTE_CODE_EXECUTION]: The skill correctly identifies that Hugging Face trust_remote_code=True allows execution of custom model code. It mitigates this by requiring an explicit --trust-model-code CLI flag from the user and providing clear warnings about the risks before loading any model assets.
[DATA_EXFILTRATION]: No evidence of unauthorized data transfer was found. The skill includes a dedicated environment isolation helper (LIVE_ENVIRONMENT_KEYS) to ensure that credentials like HF_TOKEN are not inherited by subprocesses during live inference parity tests.

nv-reason-ct