The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied MONAI-style datalist JSON files to manage training data. While the wrapper performs path validation, this represents an ingestion point for untrusted data that could influence agent behavior if the JSON fields are manipulated with natural language instructions.
Ingestion points: The datalist JSON file is ingested in scripts/run_mr_brain_finetune.py via _load_json.
Boundary markers: The skill uses standard JSON parsing but does not implement explicit boundary markers or 'ignore' instructions for the agent when processing the resulting data structures.
Capability inventory: The skill possesses significant capabilities, including local file writes, network access (to HuggingFace/GitHub), and the ability to execute shell commands via subprocess.run.
Sanitization: The script validates modality strings against a fixed whitelist (SUPPORTED_MODALITIES) and verifies the existence of image files on disk.
[COMMAND_EXECUTION]: The script scripts/run_mr_brain_finetune.py uses subprocess.run to orchestrate multi-step training workflows. It executes Python modules such as scripts.diff_model_create_training_data, scripts.diff_model_train, and scripts.diff_model_infer using sys.executable. These commands are integral to the skill's purpose of finetuning medical imaging models.
[EXTERNAL_DOWNLOADS]: The skill is configured to download and execute code from an external repository.
Repository: Fetches the NV-Generate-CTMR codebase from the NVIDIA-Medtech organization on GitHub.
Integrity: The instructions and manifest recommend pinning the download to a specific git commit hash (61c4ec709b84cad468852243c48e250bec732074) to ensure reproducibility and stability.

nv-generate-mr-brain-finetune