dyad:multi-pr-review

Multi-agent code review works by spawning independent sub-agents to analyze PR diffs simultaneously. Each agent reviews changed files in a different randomized order to reduce bias, with specific agents focusing on code health, correctness, bugs, and security before aggregating findings through consensus voting.

Automate PR review by running an orchestrator script against a PR diff to trigger the multi-agent review workflow. The system spawns three Claude sub-agents, aggregates their findings, and automatically posts a summary report with inline comments for issues that meet the consensus threshold.

Yes, you can adjust the number of agents, consensus threshold, minimum severity, thinking budget, and max tokens. Only issues identified by two or more agents where at least one rated it medium or higher severity are reported in the final summary.

The consensus voting system deduplicates by checking aggregated findings against existing PR comments before posting. This prevents redundant feedback, ensuring only newly identified issues that meet the consensus threshold are posted as inline comments.

The best way to reduce ordering bias is to provide each independent review agent with the changed files in a different randomized order. This ensures no single agent is influenced by the sequence in which files are presented during the diff analysis.

When no new high or medium severity issues are found, the multi-agent PR review still posts a summary report. Any low priority issues identified are included in a collapsible section at the bottom of the summary.