Cookies, session recording and ad conversion measurement improve the product. Before you accept we measure page views and anonymous interactions, with no cookies, no recording and no profile. Reject keeps it that way. Cookie policy
30 merged pull requests, each already reviewed by one of the three, replayed for Vetoo at the exact commit that reviewer saw. Same code, same moment, a second opinion.
Read the elevenVetoo reported 11 confirmed defects that the reviewer already on the pull request had not: 1 critical, 8 high, 2 medium. 8 were fixed by the authors afterwards, 3 are still confirmed in the code.
Every finding, ours and theirs, was checked against what the authors did next.
30 pull requests26 repositories10 languages10 per reviewer
CodeRabbit
Greptile
Cursor BugbotThree habits, each visible in what it caught here.
Four were defects the fix itself left behind: a race closed on one path and still open on the sibling path, a window a rewrite reopened and the second path that kept it open, a buffered response the new shutdown never cancelled.
A soft delete that could not win its own merge because a sibling column was never touched, duplicate ids that pass validation and fail on the key after the parent row is committed, an error matcher waiting for a message the backend never sends, a redirect allowlist that never looked at the userinfo part of the URL, a guard that checked one condition where its siblings check two.
A type declared twice so the check failed on the very commit under review, and a template that pastes spec text into generated code without escaping it.
CodeRabbit posted four comments for every one of Vetoo's on these PRs. The three defects only Vetoo reported were the quiet kind: a shutdown that leaked buffered responses, a re-entrancy window a rewrite reopened, and the second path that kept it open.
CodeRabbit 69 inline comments, Vetoo 17Six defects only Vetoo reported, three of them rated high: a sign-in redirect that trusted userinfo, a partial write on create, spec text pasted unescaped into generated code. On the same PRs, maintainers pushed back on six of Greptile's findings.
Greptile 34 inline comments, Vetoo 20The strongest of the three on this set, and still two confirmed defects it did not have: a dead error-matching branch, and a soft delete that could not win its own merge.
Cursor Bugbot 35 inline comments, Vetoo 14By the primary language of each diff. Public repositories, merged pull requests, September 2026.
Same PRs, same commits. 30 merged pull requests from 26 public repositories, 10 already reviewed by each of CodeRabbit, Greptile and Cursor Bugbot, 10 languages. Each replayed on a private mirror at the exact commit that reviewer saw, original title and description, nothing else.
Adjudicated against the code. Every finding, Vetoo's and theirs, checked against what happened next upstream: fixed by the author, agreed by a maintainer, verified in the code, rebutted, or raised on code the reviewer never saw. Only confirmed defects count. The other reviewers also found defects Vetoo did not; this page shows what Vetoo added.
One rule for the ledger. A row lands only when the defect was confirmed and the reviewer already on that PR had not reported it. Repositories and pull requests are not named on this page; the adjudication sheet with them is shared on request.
11 defects: 3 on pull requests CodeRabbit had reviewed, 6 on pull requests Greptile had reviewed, 2 on pull requests Cursor Bugbot had reviewed. Reviews ran September 2026; these tools change weekly, and so does Vetoo. Severity: critical when normal use crashes, hangs or loses data; high when reachable input misbehaves or a stated defence is missing; medium when the trigger is narrow or the effect is latent. Where the review labelled the finding, that label stands.