Repository navigation
feat: review every changed file and stop dropping real findings - #32
Merged
Merged
Conversation
A single review preloaded eight changed files and relied on the model to read the rest, so most files went unread and reviews reported no issues. Files the main review cannot hold are now reviewed in coverage passes of their own, implementation first, until the account limit or a time budget. Reviewers are asked for every distinct defect with a confidence floor of 0.8, verification drops a candidate only on a quoted refutation, and @-names in findings become code instead of failing the review. A recall replay of past commits shows 7 of 30 known issues found, from none before.
Mentions get a word joiner after @, which no markdown escape, backtick run, fence or HTML can undo. A file whose coverage candidate fails validation stays not fully reviewed, so a seen defect never becomes an approval. Provider and billing errors end the optional passes instead of discarding the finished review, coverage requests start only within five minutes of the action's start, and batched groups skip coverage passes. A refutation must quote the candidate's own files beyond its evidence.
Character references that GitHub decodes into a mention (@, @, @o) get a word joiner after the ampersand. A GitHub read failure in a coverage pass ends the passes instead of failing the review, the time budget is rechecked right before each request with a four-minute cutoff, a candidate without a string path keeps the whole pass unreviewed, and uncertain candidates are filtered before the cap.
Joiners the model wrote itself are removed before mentions are quieted, so length limits count real text. The recall runner records a case the CLI gave up on and keeps the fixture funded for the remaining cases.
|
Reviewed 0e11cbd · see the review |
hyochan
commented
Nov 18, 2026
|
Visit the preview URL for this PR (updated for commit 0e11cbd): https://codeyou-df251--pr32-feat-review-coverage-19yrziff.web.app (expires Wed, 25 Nov 2026 01:47:29 GMT) 🔥 via Firebase Hosting GitHub Action 🌎 Sign: 844dad9c62537d028cabc21008e4e2e4031ce4e5 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Gitdog found none of the 74 issues that Opus and Fable later found on PRs 25–30. Every review said "No issues found", and most changed files were "not fully reviewed", even on 3-file follow-ups. A single review preloaded 8 changed files and relied on the model to read the rest. A replay of the reviewer on those commits, with a candidate trace, showed three causes:
What changes
The finished main review is never discarded because of a pass, and batched reviews keep covering files through their groups.
@types/nodeor a bot command used to fail the whole review. Now a word joiner follows every@nameand every&that starts a character reference. I checked with GitHub's markdown API that this survives escapes, backtick runs, fences, HTML and@-style references, and that no one is notified.bun run test:e2e:recallreplays the reviewer on 4 past commits (e2e/recall/cases.json, 30 known issues) on the development deployment. An optionaltracehook shows what each step proposed and what verification discarded.Measured
On the replay of those 4 commits, before this change Gitdog found 0 of the 30 known issues. After it, all 7 published findings are real known issues, with no false positives:
resolveover-reportingOnly README.md, which is also a truncated guide, stayed "not fully reviewed" in each case. Runs vary, and over several replays Gitdog surfaced 9 distinct known issues.
Production impact
Verification
bun run lint,typecheck,build,test(2073 tests),format:check,check:api,typecheck:cli,test:cli🤖 Generated with Claude Code