The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted data from pull request diffs, descriptions, and existing comments. Evidence: pr-context-collector.md ingests PR content via the PR_URL. Boundary markers: The skill employs a multi-phase architecture that separates data collection from adjudication and drafting. Capability inventory: The skill can execute shell commands via gh CLI, write files to the workspace, and post to the GitHub API. Sanitization: A dedicated review-verifier.md phase validates all findings and metadata against the source diff before the human approval step.- [EXTERNAL_DOWNLOADS]: The skill is configured to fetch documentation and best practices from trusted services including Google, GitHub, OWASP, and Vercel. These references are used to provide evidence-backed review findings and do not involve execution of remote code.- [COMMAND_EXECUTION]: The skill utilizes the gh (GitHub CLI) tool through encapsulated shell scripts (collect-pr-review-comments.sh, post-pr-review.sh) to safely interact with the GitHub API for data retrieval and posting.- [DATA_EXFILTRATION]: The orchestrator implements robust path validation for the OUTPUT_FILE input. It requires the path to be workspace-relative, end in .md, contain no parent directory (..) segments, and remain outside the .git/ directory, effectively mitigating path traversal and unauthorized file access risks.

review-pull-request