4 months ago
896b9aeincident (dither-kit#8): a blocked contributor fixed the change and the comment was edited in place to "passed" — the receipts vanished and the resolution landed at the top of a long thread where nobody reads. this supersedes the "one comment, always upsert" decision. - SAME verdict on a re-run ⇒ EDIT in place (unchanged, the common path). - a verdict TRANSITION ⇒ a NEW comment after the commit, AND the previous comment marked superseded (struck text + a pointer; button/details/marker dropped). only transitions create comments — never a thread of them. - RUN HISTORY is the single source of truth: the worker computes previousVerdict (getPreviousVerdict) and passes it to the adapter, which decides edit-vs-transition SOLELY from it. the marker is used ONLY to locate the active comment, never to infer what happened — so a human deleting/editing the comment can't desync the adapter from history. if a transition's old comment is gone, the resolution posts anyway with nothing to supersede, never silently a "first verdict". - resolution copy knows the previous verdict (blocked→passed "that's cleared. good to merge.", etc.), from the same worker-supplied value. same incident, second fix: a cleared block left a CHANGES_REQUESTED review that can't self-edit and kept gating merge on required-review repos. on a block→(pass|sent-to-review) transition the worker dismisses the outstanding review (a new best-effort, idempotent dismiss-review action; adapter PUTs the dismissal). comment ownership unchanged and now also gates the dismiss. tests: same-verdict edits (ten re-runs = one comment) · transition supersedes old + posts new · a transition whose old comment was DELETED still posts (no supersede, no crash) · superseded golden · dismiss-review PUT · the two decision queries over real postgres. goldens regenerated. checks: typecheck all, biome, boundaries, 229 tests.
Parent8b250b5
