The reply names the line and the commit#
Code Review reads the commented file at the PR head, 40 lines either side of the comment, or 80 with an "approximate location" note when the comment is outdated. It confirms a fix only when that code shows it, and says where: "Fixed in PR-head commit 229cbe772f9b". When the bug is still there, the reply says what is wrong: "Not fixed at PR-head commit e3a30d03ffd5: line 1530 still uses `1 <= score < 5`".
This works on GitHub and GitHub Enterprise, for every reply to a review comment.
A thread resolves after the code was read#
The flag that can auto-resolve a thread is set only when the current code was loaded. If the file could not be read, the reply says so and leaves the check to the next review. If new commits land while the reply is written, it ends with "Checked at <sha>; the PR has newer commits since" and the thread stays open.
Comments inside the code and the author's message count as claims, and the code decides. In testing, a planted note telling the bot the bug was already fixed left the answer unchanged.
An open major finding scores 3/5#
Any open MAJOR finding now sets Code Quality and Merge Confidence to 3/5, "Review recommended before merging". One or two majors used to score 4, which read as safe to merge. Resolving the thread re-scores the review straight away, so fixing the last major takes it to 5/5.
The rule is the same for every org, and the check conclusion does not depend on the score, so nothing new blocks a merge. The score is the one the Review Scorecard introduced.
Learnings hold rules for the whole repo#
A reply that states a general rule is saved as a learning and shapes every later review of the repo. Fix reports, explanations about one PR, acknowledgements and commit references are now answered as what they are, and when the classifier is unsure it does not save. On 45 hand-labelled real replies, saved learnings fell from 30, 28 of them wrong, to 2, both real rules.
Improvements#
- Code ReviewComment themes are defined once per repo, with fewer model calls per build.
Fixes#
- Code ReviewFindings by theme finishes building, where it could sit on "Learning <repo>'s themes" for hours.
- Code ReviewThe Themes page publishes the coherent themes a build finds.
- Agent InsightsEach project shows its tracker name, the same on every page.
