The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided SMILES strings via CSV files. This represents a theoretical attack surface for indirect prompt injection if the input data were to contain malicious instructions targeting the agent's reasoning. The risk is minimized as the data is processed by specialized chemical informatics tools inside a container, and the agent does not directly reason over the raw file content.
Ingestion points: Input CSV files provided via the --csv parameter and processed by scripts in the KERMT repository.
Boundary markers: Explicit validation steps (check_data.py) and a preparation phase (prepare_data.py) serve as functional boundaries.
Capability inventory: The skill executes shell commands for container management, performs file writes to a dedicated run directory, and manages Docker lifecycles.
Sanitization: The instructions specify that the preparation step strips the input CSV to SMILES-only content, which effectively sanitizes the data of potential non-chemical text columns.
[COMMAND_EXECUTION]: The skill orchestrates its workflow through a series of shell commands. These include running local repository scripts (kermt_container.sh) to perform system audits, data validation, and model inference. These operations are transparent, well-documented, and strictly confined to the skill's primary scientific purpose.

kermt-infer