The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code from external sources (local files, diffs, and GitHub pull requests) and possesses file-write capabilities to apply fixes.
Ingestion points: The agent reads code content via diff, pr <number>, or specific file paths provided as arguments in SKILL.md.
Boundary markers: The instructions explicitly mandate isolation by requiring a "fresh subagent every round" for the review process.
Capability inventory: The agent can modify the local file system (applying fixes), execute shell commands (make, gh), and perform git operations (commit).
Sanitization: No explicit sanitization or filtering of the code content is described before processing.
[COMMAND_EXECUTION]: The skill executes various development-related shell commands to validate and commit fixes.
Evidence: Rule 2 specifies gh pr checkout <number> for PR reviews. Rule 4 specifies running make build, make test, and make lint, or other "native commands" to verify the repository state. Rule 6 involves executing git commit commands.

review-loop