The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute repository scripts and the docker CLI to manage containerized training environments. These calls are strictly functional and use list-based argument passing to mitigate injection risks.
[DYNAMIC_EXECUTION]: To handle molecular model state, the skill implements a custom _VocabUnpickler that restricts object creation to a safe whitelist of types. It also uses PyTorch's weights_only=True mode for checkpoint loading to prevent arbitrary code execution via malicious artifacts.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided SMILES datasets in CSV format. The ingestion process includes structural validation using the RDKit library and numerical featurization, which minimizes the risk of the agent interpreting data as instructions.

kermt-add-cmim-pretrain