The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts/simulation_dry_run.py tool generates Python simulation scripts from validated JSON configurations. This generation process is safe as it uses a fixed template and strictly serializes configuration values into literals using repr(), preventing code injection into the generated output. The resulting scripts require manual execution and user acknowledgement of a configuration ID, ensuring a 'human-in-the-loop' security model.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing untrusted simulation data (HDF5, netCDF4, and JSON logs). This is mitigated by comprehensive security controls:
Ingestion points: Data is ingested via scripts/output_inventory.py and scripts/budget_summary.py using h5py for metadata and strict line-by-line parsing for text logs.
Boundary markers: Strict JSON schema enforcement and metadata allowlists (such as SAFE_SCALAR_ATTRIBUTES) are defined in scripts/_schema.py and scripts/output_inventory.py.
Capability inventory: Bundled scripts are limited to local file operations and do not utilize network, subprocess, or arbitrary shell execution capabilities.
Sanitization: Centralized path validation in scripts/_common.py prevents directory traversal and symlink-based attacks by resolving and checking all paths against an explicit root via checked_root and checked_input helpers.

fluidsim