The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill optionally downloads pretrained models and vocabulary files from the official NVIDIA repository on Hugging Face (nvidia/NV-KERMT-70M-v2). This download is gated by an explicit user consent prompt in the workflow instructions and uses a pinned revision for integrity.
[COMMAND_EXECUTION]: The workflow orchestrates tasks using Docker and Python subprocesses. Subprocess calls are constructed using lists rather than shell strings, which mitigates common command injection risks. Containerized execution is configured to run using the host user's identity to prevent privilege escalation.
[INDIRECT_PROMPT_INJECTION]: The skill processes external SMILES CSV files and model checkpoints as part of its core functionality.
Ingestion points: Ingests user-supplied molecule data in CSV format and model weight files in PyTorch format.
Boundary markers: The skill documentation defines a strict sequence of validation and preparation steps, though it does not use specific delimiters for data interpolation in the prompt.
Capability inventory: Includes the ability to execute containerized Python scripts, perform network requests for model fetching, and write results to the local file system.
Sanitization: The skill performs validation on input data using RDKit for SMILES strings and a restricted PyTorch loader for checkpoints. It also utilizes a custom _VocabUnpickler to restrict allowed classes during vocabulary loading, preventing unsafe deserialization.

kermt-embed