Branch: refs/heads/ci-check-flicker-stabilisation Home: https://github.com/xwiki/xwiki-dev-llm Commit: 59b2f2361afa681b16ecb3fc378628a8081a9da0 https://github.com/xwiki/xwiki-dev-llm/commit/59b2f2361afa681b16ecb3fc378628... Author: Vincent Massol <[email protected]> Date: 2026-09-19 (Sat, 19 Sep 2026) Changed paths: M README.md M xwiki/skills/xwiki-ci-check/SKILL.md M xwiki/skills/xwiki-ci-check/routine-prompt.md M xwiki/skills/xwiki-ci-check/tools/ci-check.mjs Log Message: ----------- [Misc] Stabilise one proven flicker a run, and only when nothing blocks a release * ci-check.mjs picks at most one stabilisation candidate per sweep — a proven, already-filed flicker that nothing answers, inside the horizon, not systematic, and the one among those that fails most often, since a fix can only be shown to work on a test a repeat run can catch failing. `summary.stabilise` carries it or the reason there is none, and the paste prints that reason as its own line. * The candidate keeps its evidence and its 28 days of Develocity whether or not it won a deep slot: a filed flicker sits low in the analysis budget by design, and the configuration breakdown is the first thing whoever fixes it reads. * Nothing is proposed while something release-blocking is open. A build break blocks at any age — a quality gate red for a fortnight is a fortnight of blocked releases — while a systematic test breakage stops blocking past the horizon, where the run may not write about it at all and so cannot work on it instead. * SKILL.md §5 gains the procedure: measure the rate on the configuration Develocity names, fix inside Tier C, measure again, and open one unassigned draft PR carrying both rates with their execution counts and what a clean series does not prove — or nothing at all when the second rate is no better. * A failing SonarCloud quality gate stops being `report, never fix`: it makes the build red for everyone, so it is the run's first fix, handed to xwiki-fix-sonarqube-issue and still under nobody's name. On a cycle-2 branch the gate itself is usually the bug, and turning it off there is a proposal to the team. * The digest names a stabilisation PR and is not withheld on the morning it opened one. 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