Skip to main content

​
How it works

Review cycles measures the average number of completed human review cycles from the PR author’s perspective. A cycle completes when an eligible human non-approval review is followed by a verified PR content change before merge. Several reviews before the same change share one cycle. Further changes add no cycles until another eligible review occurs. Approval alone, feedback without a subsequent verified change, and changes without pending feedback contribute zero. Events at or after merge are excluded; a review and update with identical timestamps do not establish that the update followed the review. For example:
This PR has one completed cycle. Review C has no subsequent content change, so it does not complete another cycle. Immediate approval has zero completed cycles. The mean, median, and cycle drilldown include only merged PRs with at least one completed cycle. Review rate uses the broader merged-PR population independently. Identical-tree rewrites and proven clean base refreshes do not complete cycles. Content changes include binary files, renames, file modes, deletions, symlinks, and submodule revisions. Only verified changes count: incomplete historical transition evidence can undercount completions. Unresolved history is retained in internal evidence coverage. Organizations awaiting rollout retain the previous started-review-group definition until recalculation. The rollout kill switch restores that definition consistently.

​
What it indicates

  • 1 completed cycle means feedback was followed by one verified revision
  • Higher values may indicate PRs that need more rework, unclear requirements, or opportunities to improve code quality before review

​
Common causes of high review cycles

  • Large or complex PRs that require multiple rounds of feedback
  • Unclear requirements leading to revisions
  • Code quality issues that need to be addressed
The following are excluded from this metric:
  • Self-reviews (where the PR author reviews their own PR)
  • Reviews that occur after the PR is merged
  • Reviews from bots and AI tools (e.g., Copilot, CodeRabbit)
Last modified on October 17, 2026