The Agent Skills Directory

Untrusted Data Handling: The skill identifies external pull request content—such as titles, bodies, diffs, and comments—as untrusted data. It implements specific instructions to treat this content as data rather than directives, which serves as a mitigation against indirect prompt injection.
GitHub CLI Integration: The skill uses the official GitHub CLI (gh) and GraphQL API for all repository interactions. It verifies authentication and collaborator status before proceeding, ensuring the agent operates within the maintainer's existing permissions.
Secure Command Execution: When submitting review bodies, the skill writes content to temporary files and uses the --body-file flag with gh pr review. This approach avoids passing large or complex strings directly through shell arguments, effectively mitigating potential shell-escape vulnerabilities.
Human-in-the-Loop Design: All write operations, including posting reviews, closing pull requests, or locking conversations, require explicit confirmation from the maintainer. The skill drafts findings and dispositions but never performs mutations unilaterally.
Parallel Analysis Optimization: The skill utilizes background subagents to pre-fetch and analyze upcoming pull requests in the queue. These subagents are strictly scoped to read-only operations and do not have the capability to perform write mutations, maintaining a clear privilege boundary.
Slop Detection Logic: A structural scan is performed on new pull requests to identify non-genuine contributions (e.g., student projects or low-effort submissions) based on patterns like area sprawl or fork merge-commit floods. This allows for early-exit and manual triaging of suspicious content before deep analysis.

code-review