The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external search data through a records.json file to generate structured tables. While this is an ingestion point for untrusted data, the skill implements comprehensive security controls.\n
Ingestion points: The _load_records function in scripts/provider.py reads data from a JSON file.\n
Boundary markers: The script enforces a strict data schema using dataclasses (Record, Field, Evidence) and validates all inputs.\n
Capability inventory: The skill generates shell commands for the parallax tool and manages local snapshots in a state directory.\n
Sanitization: It uses shlex.join to prevent shell injection, validates all table and column names against reserved characters (,, :, #) used in the downstream CLI grammar, and employs pathlib.resolve() with is_relative_to() to prevent path traversal attacks during artifact verification.\n- [UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill generates a command for the parallax proposal tool. This generation is secured by strict input validation of table and field names and uses shlex.join to ensure proper shell escaping.\n- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill manages a local state directory for run metadata and evidence snapshots. Access to these files is governed by strict path resolution and containment checks that prevent directory traversal attacks.

data-provider