dyad:swarm-pr-review

This skill leverages the swarm functionality of Claude Code to automate and elevate the pull request review process. Instead of relying on a single pass, it spawns a dedicated team of three specialized AI agents: a correctness and debugging expert, a code health specialist, and a UX wizard. These agents independently analyze the provided PR diff, identify potential issues ranging from critical bugs to maintainability concerns and accessibility improvements, and then engage in a collaborative dialogue. They challenge each other's findings, validate cross-references for infrastructure changes, and reach a consensus on the final verdict. The system ultimately aggregates these insights into a professional, actionable summary comment on GitHub, including inline comments for HIGH and MEDIUM severity issues to streamline the feedback loop.

• Multi-perspective analysis covering correctness, security, control flow, code health, duplication, complexity, and UX design consistency.

• Collaborative discussion phase allows agents to endorse or challenge peer findings, reducing false positives and noise.

• Automated integration with GitHub CLI (gh) to fetch diffs, PR metadata, and existing review comments to prevent duplication.

• Structured output format using JSON findings for programmatic clarity and readability.

• Intelligent conflict detection for infrastructure changes like DB migrations where schema updates lack corresponding code implementation.

• Requires GitHub CLI (gh) authentication and permissions to post comments on the target repository.

• Best suited for teams looking to enforce high standards for code quality, accessibility, and architectural integrity in their PR workflows.

• Operates best when the user provides specific PR context or triggers it within a repository workspace where the branch is active.

• The agents are constrained by the provided PR diff scope and should not be expected to perform full repository-wide architecture changes.

• Ensure that sensitive code or environment variables are not inadvertently exposed in public-facing PR comments.