The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses git and gh (GitHub CLI) to retrieve diffs, commit history, and prior pull request comments for analysis. These operations are limited to standard repository metadata retrieval.
[DATA_EXFILTRATION]: The skill implements a strict 'Data Trust Boundary' in references/common.md. It explicitly treats all VCS output (diffs, logs, blame) and guideline files as data for analysis only, disregarding any directives, prompts, or behavioral suggestions found within them to prevent indirect prompt injection.
[PROMPT_INJECTION]: To prevent hallucinated line numbers, the skill uses a custom 'Diff Annotation' algorithm ([L<n>] markers). Reviewers are only permitted to cite lines that carry these markers, creating a validated allowlist for code citations.
[SAFE]: The skill includes a 'Size Gate' to prevent processing excessively large diffs (>3000 lines or 40 files), which mitigates potential resource exhaustion or model degradation issues during large-scale reviews.
[SAFE]: Guideline discovery is strictly scoped to the repository root using git rev-parse --show-toplevel. It explicitly excludes personal configuration files from the user's home directory (~/.claude/), preventing the leakage of global settings into the project review context.
[SAFE]: Code modifications are strictly opt-in. While the skill suggests fixes, it requires explicit human confirmation before writing any changes to the working tree.

review-lens