The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands to generate diffs and metadata for the review process.
Evidence: git merge-base, git rev-parse, and git diff commands are used in Step 2 of the protocol in SKILL.md to output data to review-stat.txt and review.diff.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (source code changes) which is then interpreted by an AI subagent, creating a potential injection vector.
Ingestion points: The subagent reads the contents of review.diff and other repository files using git show, git log, and git diff as described in reviewer-template.md.
Boundary markers: The reviewer-template.md uses structural markdown headers to separate instructions from code data, but lacks specific warnings to the subagent to ignore instructions that might be embedded within the code being reviewed.
Capability inventory: Both the primary agent and the subagent have the ability to execute git commands and read arbitrary files within the repository context.
Sanitization: There is no indication that the content of the diff files or repository files is sanitized or filtered before being passed to the subagent's prompt context.

code-review