The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses the gh CLI and standard Unix utilities (find, grep, awk, jq) to fetch metadata, diffs, and issue details. Shell commands are used to inventory repository files and symbols to facilitate reusability checks. Command construction follows safety patterns, such as using jq for JSON payloads and gh api flags for variable binding.
[DATA_EXPOSURE]: The skill reads repository content, PR diffs, and project-specific suppression files (.claude/review-suppressions.yml). It maintains a local cache of review results in the user's home directory ($HOME/.claude/skills/review-pr/cache) to optimize subsequent runs.
[PROMPT_INJECTION]: The skill has a risk surface for indirect prompt injection as it processes untrusted content from GitHub PR titles, bodies, and issue descriptions. Evidence of mitigation includes:
Ingestion points: Data enters the context via gh pr view, gh pr diff, and gh issue view commands specified in Phase 1 of SKILL.md.
Boundary markers: The skill uses structured prompts with explicit headers like ## Ground truth and ## Your task to separate untrusted data from instructions.
Capability inventory: The skill uses gh CLI for repository interaction, filesystem tools for searching, and invokes subagents for analysis.
Sanitization: It implements a mandatory "Grounding Pass" for subagents and a detailed "Critic Pass" (Phase 3) that verifies every finding against the mechanical diff and the stated intent.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata and code content from GitHub via the official gh CLI. This is the intended behavior for a remote PR review tool and targets well-known infrastructure.

review-pr