The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of local Python scripts (scan_bypasses.py, review_gate.py, post_review.py) and uses the gh CLI tool to interact with GitHub repositories. This provides a functional execution surface for the agent within its environment.
[INDIRECT_PROMPT_INJECTION]: The agent processes untrusted data from pull request diffs, descriptions, and comments. A malicious pull request could contain instructions designed to influence the agent's review logic or attempt to bypass established safety filters.
Ingestion points: Untrusted PR diffs and metadata are fetched via gh pr diff and gh pr view (SKILL.md, Step 0).
Boundary markers: Absent. The diff content is provided to the agent without explicit delimiters or instructions to ignore embedded commands.
Capability inventory: The agent can execute scripts, access the file system, and perform network requests through the authenticated GitHub CLI.
Sanitization: No sanitization is performed on the ingested PR content prior to analysis.
[DYNAMIC_EXECUTION]: The scripts/post_review.py script uses subprocess.run to programmatically construct and execute gh CLI commands based on variables in the agent's context, such as PR numbers and finding content.

the-judge