Repository navigation
ci: make main's CI green and steady - #161
Merged
Merged
Conversation
A pool's close returns before PostgreSQL's backend exits, so the supervisor counted a closed migration pool as open and refused the freeze: all 71 tests of deploys.test.ts failed in CI run 37934023035 with "Close database connections before freezing a template". The freeze now waits up to 2 s for the database's backends to leave (as PostgreSQL's own CREATE DATABASE … TEMPLATE waits for exiting ones) and answers each lease's commands in order.
Vitest's contiguous hash slices put importV1Crew (129 s) and CrewEngine (105 s) in one shard: Test Server 1 ran 491 s on average and up to 590 s of its 600 s limit (once cancelled), against 438 s and 439 s for the others. Four shards that each take every 4th sorted path spread a directory's slow suites apart: about 310–350 s of tests each by the last run's timings.
ab1ef8f (fix(data): hold what a scope read through a reconnect's grace) made a lost link read as current for STREAM_POLICY.reconnectGraceMs (5 s) and ends the hold on the store's own timer. The outage scenarios install the client clock, whose positive timers run only when advanced, so the grace never ran out: the menu never said HQ is unavailable and then.hq.isUnavailable timed out in 11 g-outage cases and foundation's "G: HQ outage preserves rows and catches up after return". CI's chat-gate runs c-mate only, so the areas rotted unseen. then.hq.isUnavailable now passes the grace on an installed clock, one grace at a time, until the tab says HQ is unavailable: the grace counts from when the tab noticed the drop, which the harness cannot observe, so a single advance raced the socket close. The clock says whether it is installed. No test title changes.
Since ab1ef8f a link HQ ends for its session (4401) holds what it read through the reconnect grace, so a renewal within the grace shows no outage line: "successive HQ session endings renew and restore the menu" no longer can wait for then.hq.isUnavailable. With the previous commit letting the grace run out it passed only when the line happened to flash before the renewal. The case now proves the ending by what follows it: the tab, woken, opens a new HQ subscription, and the menu is caught up. The wake repeats until the tab has noticed the ending, since a wake that comes first finds no recovery to expedite. The title stays.
Since the footer became a hard-stop gradient (#154), "history does not paint through the composer footer" failed about one run in four: the two screenshots differed by ±1 in one channel on every 8th pixel of every 8th row, which is Chrome's gradient dither, its phase set by the tile it lands in. Hiding the history re-rastered the footer and moved the dither. A solid ::before from the card's top edge down paints the same colour every time. The gap above the card stays clear and nothing paints below it. Before: 1 of 4 local runs failed. After: 8 of 8, plus 4 of 4 under eight CPU burners.
Base UI binds a trigger's native mouseenter listener to the element it first
renders and does not rebind when `render` swaps the element. A chat file chip
renders a menu button until its open action is known, then an <a>: hovering
the link afterwards never cleared Base UI's mouse-move block, so its path
tooltip never opened.
This is the intermittent CI failure in Chat Contract (C shard 2), images
"messages appear before image bytes, reserve their size, and reuse previews
on reopen", timing out at waitForSelector('[data-slot="tooltip-popup"]'). A
diagnostic under a 4x CPU throttle showed every failure had the BUTTON ->
A swap, with pointerenter and mousemove delivered to the link and the
pointer still on it; a second move did not open it either.
TooltipTrigger now keys Base UI's trigger on the rendered host element's
tag, so a change of element kind remounts it and its listeners.
Under the 4x throttle the scenario failed 1 of 4 plain runs (and once in a
full shard 2/2 run) before, and passes 6 of 6 after.
fxck
added a commit
that referenced
this pull request
Oct 9, 2026
- #161 ci: make main's CI green and steady
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Main's CI failures over the last ~30 runs, each with its cause and fix:
deploys.test.ts(71 failures, run 37934023035): the test PostgreSQL supervisor refused to freeze the migrated template because a just-closed pool's backend still showed inpg_stat_activity(scripts/test-postgres-supervisor.ts, freeze). The freeze now waits up to 2 s for those backends to exit, as PostgreSQL's ownCREATE DATABASE … TEMPLATEdoes. Covered by a new RED→GREEN harness test.importV1Crew(129 s) andCrewEngine(105 s) in one shard (491 s avg / 590 s max vs 438/439). The server now runs four shards, each taking every 4th sorted path (scripts/lib/spread-shard.ts). This PR's run: 4:57–6:26 per job.TooltipTriggernow remounts when its element kind changes, with a unit test.::beforefrom the card's top edge down.ab1ef8f00holds a lost link through a 5 s reconnect grace, and the scenario clock never let it run out.then.hq.isUnavailablenow advances the installed clock, and the renewal case proves the renewal by its new subscription. No titles changed.mate-zone-architecture(9034fe8) and hqchanges.test.tsevent order (528bc92).🤖 Generated with Claude Code