dyad:multi-pr-review

Automated multi-agent PR review deploys three Claude sub-agents to analyze diffs with randomized file orders, minimizing ordering bias. It aggregates findings through consensus voting, requiring at least two agents to agree on an issue before reporting it.

Consensus voting in code review aggregates findings from multiple sub-agents, only reporting issues when two or more agents agree and at least one rates it MEDIUM or higher. This filters out isolated low-priority noise and reduces false positives.

Yes, multi-agent PR review supports configurable agent counts, consensus thresholds, and extended thinking budgets. You can adjust these parameters to control analysis depth and the strictness required for deduplicating and posting inline comments.

Automated PR review deduplicates findings by checking aggregated consensus results against existing PR comments before posting. This prevents redundant feedback by ensuring only new HIGH or MEDIUM issues are posted as inline comments.

Yes, multi-agent code review relies on the Anthropic dependency to spawn sub-agents that analyze PR diffs. It uses Claude to evaluate code health, maintainability, and clarity, with safety and reproducibility guardrails built in.

Use a multi-agent consensus approach for large PR diffs where ordering bias and single-perspective blind spots are concerns. It provides debiased, deduplicated feedback by requiring multiple agents to validate issues before posting remediation comments.