Skip to main content

​
How it works

Weave uses an AI agent to categorize PRs based on the depth of the review they received before being merged. There are three categories: full, rubber stamp, and none. The algorithm is fairly straightforward:
  1. Was there any human review activity on the PR? If not, it’s categorized as none
  2. Was the PR complex enough to warrant a nontrivial review? If not, it’s categorized as a full review.
    • For example, if a PR just fixes a typo, then any review will be categorized as a full review (even if there are no comments at all).
  3. Did the reviewer(s) meaningfully review the PR and provide any meaningful feedback on the PR? If so, it’s categorized as a full review. Otherwise, it’s categorized as a rubber stamp review.

​
When a PR is checked for rubber stamp

Only merged PRs that meet all of these conditions are checked for a rubber stamp review:
  • The PR’s code output is estimated at 1 hour or more
  • The PR has light review activity: at most 2 reviews and at most 5 review comments, with only short review text
On those PRs, an approval with no written feedback (or only a trivial comment such as “LGTM”) is a rubber stamp. Any reviewed PR that doesn’t meet these conditions, including every PR estimated at under 1 hour, is categorized as full.
Only human reviews count. Self-reviews are excluded, as are comments from CI and other non-review automation, such as GitHub Actions and issue-linking bots. Reviews from AI code review tools, such as Cursor Bugbot and Greptile, are also excluded; they’re reported on the AI code review page.
Last modified on October 15, 2026