The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and act upon instructions found within code comments, such as markers like IMPORTANT or do not remove. This represents a surface for indirect prompt injection where untrusted content in the code could attempt to influence the agent's behavior.
Ingestion points: The skill reads file contents, diffs, and git metadata in SKILL.md (Step 1 and 2) and the references/comment-reviewer.md instructions.
Boundary markers: The skill utilizes a separate subagent for the initial comment review pass, which provides a layer of isolation for the primary agent context.
Capability inventory: The agent has the ability to execute git commands, run project tests/linters, and perform code reproductions.
Sanitization: The "Trace the claim" logic in references/comment-reviewer.md acts as a verification process, requiring the agent to validate claims against git history and documentation before accepting them.
[COMMAND_EXECUTION]: The skill uses shell commands for repository analysis and verification, including git log -L and git blame as specified in the "Trace the claim" section of references/comment-reviewer.md.
[DYNAMIC_EXECUTION]: In SKILL.md (Step 3), the agent is instructed to fix identified workarounds by modifying code and running project-specific checks, including tests, type checks, and linters, to verify the new implementation.
[EXTERNAL_DOWNLOADS]: The reviewer instructions in references/comment-reviewer.md direct the agent to follow links to external issue trackers, bug trackers, and vendor documentation to gather evidence for comment claims.

no-comments