The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/audit_codex_skill_surface.py script uses Python's subprocess module to execute the codex command-line utility. It runs codex debug prompt-input to inspect the model's active catalog and codex app-server to retrieve the full unshortened skill inventory. These operations are diagnostic and intended for state verification.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from external SKILL.md files during its audit routines.
Ingestion points: scripts/audit_project_skill_roots.py and scripts/audit_codex_skill_surface.py read and parse YAML frontmatter from skill bundles in .agents/skills, .claude/skills, and user-defined project directories.
Boundary markers: The instructions include a 'Non-negotiable safety boundaries' section that explicitly mandates read-only behavior for audits and requires manual user authorization for any mutations such as moving, retiring, or installing skills.
Capability inventory: The skill leverages subprocess for system interaction and performing filesystem operations including directory traversal, file reads, and SHA256 hashing to detect drift.
Sanitization: The scripts implement regular expression validation for skill names to ensure they follow a strict lowercase-and-hyphen schema, but they do not sanitize description fields or instruction bodies, which could contain instructions meant to influence the auditing agent.

skill-governance