The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input including pull request metadata (titles, bodies, comments) and the source code changes being reviewed.
Ingestion points: External data enters the agent context via scripts/forge_packet.py (which fetches pull request content via GitHub CLI) and scripts/review_context.py (which reads repository file content and git history).
Boundary markers: Present. Both SKILL.md and references/rubric.md explicitly instruct the agent to treat change descriptions, code, and comments as evidence, never instructions.
Capability inventory: The skill uses subprocess.run within its provided Python scripts to execute system commands like git and gh. These operations are limited to information gathering and local state management.
Sanitization: The skill's scripts implement input validation, including regex-based filtering for repository and pull request identifiers, and the use of literal pathspecs in git commands to prevent argument injection.

review-code