[hyochan]

Why

• Every push re-reviewed the whole PR. On feat: add shared external purchase link flow openiap#612, three reviews of the same 91–114-file PR used 2.0M, 2.7M and 2.8M tokens ($5.28, $7.21 and $7.34). Two of them posted nothing.
• One unusable batch discarded everything. The third review failed at batch 7 because one model response could not be processed. Every batch already paid for was thrown away.

Changes

Review only what a push could have changed

• Anchor. A new review action records reviewSince = {headSha, mergeBaseSha, uncoveredPaths} from the last review that finished with a complete coverage record.

  • "Finished" means published, or completed with nothing new.
  • "Complete" means it recorded uncoveredFiles and its uncovered list was not cut short.
  • Older reviews have no such record, so the next push reviews the whole PR.

• Scope. On the next push, the scope is:

  • the files changed on the head since the anchor (it must be an ancestor);
  • the files changed between the old and new merge base, when a merge of the base branch moved it forward;
  • the files the anchor left unreviewed.

  It is reviewed as one review when it fits 60 files and the single diff budget. The summary says "Reviewed the N files affected since <sha>".

• When the whole PR is reviewed again:

  • a force push;
  • a retargeted or rewound base;
  • a comparison of 300 or more files;
  • any comparison failure;
  • a merge-base mismatch between the scope and the snapshot.

• No model call is made when nothing reviewable changed:

  • A push that only touched files outside the PR diff, or only deleted files, completes as no_new_changes without posting.
  • A binary or oversized file the push changed is reported as not reviewed.

• Publication of a follow-up review re-reads only the files its findings and evidence need.

Keep paid batches when one batch fails

• These failures leave the group's files reported as not reviewed, and the review continues:
  • a billed provider response that could not be processed;
  • invalid JSON;
  • evidence or finding validation failures;
  • review limits.
• At most two groups may fail this way. The third failure stops the review.
• These still stop the whole review: rate limits and other provider HTTP errors, credentials, exhausted credit, ambiguous outcomes, billing, quota and a changed PR.

Review

• Three independent Opus review rounds ran.
• Round 1: 2 high, 3 medium.
• Round 2: 1 medium (a base-branch merge could revert base fixes unseen).
• Round 3: no high or medium issues.

Tests

• bun run lint, typecheck, format:check, build and all 1914 tests pass.
• New tests cover:
  • a 61-file review followed by a one-file push (one single review, an inline comment and the since note);
  • a force push;
  • a push outside the PR diff;
  • merge-base movement, retarget and rewind;
  • comparison failure and size limits;
  • anchor eligibility;
  • group isolation, the two-failure cap and isolation rules.

🤖 Generated with Claude Code

[gitdogbot]

Gitdog found no supported defects in f1e4b2f9aed5. 4 of 8 changed files were not fully reviewed. This is not a merge approval; tests were not run.

Gitdog

[gitdogbot]

Gitdog found no supported defects in 98d66af55db7. 6 of 9 changed files were not fully reviewed. This is not a merge approval; tests were not run.

Gitdog