Branch: refs/heads/ci-check-forward-fix Home: https://github.com/xwiki/xwiki-dev-llm Commit: 4719d20d62a9e73166c51a0e389cafb0f13a794b https://github.com/xwiki/xwiki-dev-llm/commit/4719d20d62a9e73166c51a0e389caf... Author: Vincent Massol <[email protected]> Date: 2026-09-18 (Fri, 18 Sep 2026) Changed paths: M xwiki/skills/xwiki-ci-check/SKILL.md M xwiki/skills/xwiki-ci-check/tools/ci-check.mjs Log Message: ----------- [Misc] Stay quiet about a CI incident the branch has already fixed * Look forward as well as backward: blame reads the commits inside the regression window, and nothing read the ones pushed after CI last built the branch — so a break fixed at 22:27 was still being root-caused, and its author still pinged, at 06:00 the next morning. * Add `fixState` to an incident, computed before the deep budget is allocated: a settled incident spends no slot, gets one line instead of an analysis, and is written about nowhere. * Match on the failing file alone, never the package around it. Backward, a commit touching the test's package is a fair suspect and the cost of being wrong is a comment that asks; forward, the cost is silence about a live breakage. Measured on master: with the package key, a commit touching `VersionIT.java` silenced a systematic failure of its neighbour `DocExtraTabsIT`. * Count coverage in tests, not in keys: one incident spans the N tests that broke together, and only a full house goes quiet. A commit answering some of them is reported beside an incident that still stands. * Cache each commit's record for the run — one push otherwise costs a request per job per commit, now that every job walks the unbuilt range. Verified against xwiki-platform/master, where the two commits of `#6449` and `#6450` answer one incident between them and it went quiet correctly. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> To unsubscribe from these emails, change your notification settings at https://github.com/xwiki/xwiki-dev-llm/settings/notifications