The Agent Skills Directory

[DYNAMIC_EXECUTION]: lib/cli.py dynamically modifies sys.path to include ../../apply-agent-research/lib. This allows the skill to load modules (sanitizer, proposal_gate) from a relative sibling path computed at runtime. This pattern can be risky if the environment is not strictly controlled.
[COMMAND_EXECUTION]: lib/cli.py uses subprocess.run() to execute the GitHub CLI (gh issue create) for filing audit findings. While the command is constructed as a list to mitigate shell injection risks, it represents an automated external mutation performed by the skill.
[EXTERNAL_DOWNLOADS]: The skill instructions (Step 1) direct the agent to clone external repositories (mattpocock/skills) and read from a knowledge base mirror to gather audit data. This data is subsequently processed to generate reports and issue proposals.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external Markdown files to extract pillar names and headings, which are then interpolated into reports and GitHub issues.
Ingestion points: Reads SKILL.md files from mattpocock/skills and the agent-research-knowledge mirror.
Boundary markers: None explicitly used during the initial scanning phase, though a sanitizer is used before output.
Capability inventory: Files GitHub issues through lib/cli.py using subprocess.run().
Sanitization: Employs a sanitizer.check from the sibling skill on the combined title and body text before the gh command is executed.

skill-divergence-audit