The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a 'dry-run' approach for simulation execution. The scripts/simulation_dry_run.py generator produces a Python script that requires an explicit --execute flag and a specific --acknowledge-config-id to run, preventing accidental or unauthorized launches.
[DATA_EXFILTRATION]: Path validation in scripts/_common.py (checked_input, checked_output) explicitly rejects URI schemes (http, https, s3, etc.) and parent directory traversal (..), ensuring that file operations are restricted to the local filesystem and intended directories.
[DYNAMIC_EXECUTION]: While the skill generates a simulation script (run.py), it does so using a static key-to-module map (SOLVER_IMPORTS in scripts/_schema.py) and validated JSON scalars, preventing arbitrary code injection during script generation.
[REMOTE_CODE_EXECUTION]: The skill provides clear guidance on environment management using uv for reproducible builds. All suggested installations target official repositories and well-known services (PyPI), and the skill explicitly warns against using upstream CLI options (like --modify-params) that could execute untrusted text.
[INDIRECT_PROMPT_INJECTION]: The tools for analyzing HDF5 and netCDF files (scripts/output_inventory.py, scripts/budget_summary.py) use a 'lazy' metadata-only approach. They inspect dataset shapes and attributes without loading large field arrays and explicitly detect/report external HDF5 links rather than following them, mitigating potential data-poisoning attacks.

fluidsim