Re-checking checks a release cut (#12594)
A check the forge itself cut short (a restart, a release, a crash loop) proved nothing about the change. It is recorded as cut, never failed: the pull is not red, and no builder is woken for it.
- A release or restart queues the same check again. Releases are counted apart (
CHECK_RELEASE_LIMIT). - A cut within a minute of starting (
CRASH_LOOP_MS) is a crash loop's: it does not spend the long-run allowance (CHECK_RESTART_LIMIT, 3), but every death, long or short, spendsCHECK_CRASH_LIMIT(10). Past either, the check is settled as cut, saying "ask for a re-check", and is not queued again. - A summary is read as cut only when it is exactly one the forge writes (
isCutinsrc/core/runs.ts); a failure that merely begins with the same words stays red. - On boot after an unclean stop, one log line says how many checks were cut and queued again.
Ask again for a window
After a bad release, one call re-checks every open pull whose latest check on its head was cut or failed and ended inside a window.
- API (a repository admin,
adminscope):POST /api/v1/repos/:owner/:name/checks/recheckwith{"since": "<ISO instant>", "until": "<ISO instant>"};untildefaults to now. Asincethat is not a string instant, or afteruntil, is400 invalid_window. Answers{queued: [pull numbers]}, at most 50 per call, withmorewhen some were left. - Terminal (a repository's admins): open the repository, then the palette's "Re-check cut checks"; type
3h(hours ago) or an instant.
A pull whose latest check is not cut or failed, whose check ended outside the window, or whose cut was on an older commit than its head is left alone.