Browser spec "another step opens at its own tail" fails now and then under the full lane #193

Closed
opened 2026-10-18 01:35:00 -04:00 by cmoriarty · 3 comments
Owner

ui/e2e/ui-polish-103.spec.ts:106, "another step opens at its own tail, however far up the last one was read", failed once in the full lane on 61f84b0 (2026-10-18):

Error: expect(locator).toHaveCount(expected) failed
Locator:  locator('[data-jump-latest]')
Expected: 0
Received: 1
Timeout:  10000ms
  > 116 |   await expect(page.locator('[data-jump-latest]')).toHaveCount(0)

After Escape closes the second step, the "jump to latest" button is still there 10 s later. The merge under test changed no transcript or scroll code (one Settings label only).

Since then:

  • the test passed 8 of 8 on its own (--repeat-each 8 --workers 4);
  • its file passed 11 of 11;
  • the next full lane on the same code (after an unrelated test fix) passed all 324 browser specs.

So it fails only under the full lane's load, which points at a race between the step closing and the transcript's follow state, not at a timeout to raise. Worth finding: the same family of "flake" (the keyboard specs, 2026-10-16) turned out to be a real race in the console's route handling.

To do

  1. Reproduce under load (the full lane's browser stage, or --repeat-each with more workers than cores) and read what keeps [data-jump-latest] after Escape.
  2. Fix the cause in the console, or the spec if it asserts a moment the console doesn't promise.
`ui/e2e/ui-polish-103.spec.ts:106`, "another step opens at its own tail, however far up the last one was read", failed once in the full lane on 61f84b0 (2026-10-18): ``` Error: expect(locator).toHaveCount(expected) failed Locator: locator('[data-jump-latest]') Expected: 0 Received: 1 Timeout: 10000ms > 116 | await expect(page.locator('[data-jump-latest]')).toHaveCount(0) ``` After Escape closes the second step, the "jump to latest" button is still there 10 s later. The merge under test changed no transcript or scroll code (one Settings label only). Since then: - the test passed 8 of 8 on its own (`--repeat-each 8 --workers 4`); - its file passed 11 of 11; - the next full lane on the same code (after an unrelated test fix) passed all 324 browser specs. So it fails only under the full lane's load, which points at a race between the step closing and the transcript's follow state, not at a timeout to raise. Worth finding: the same family of "flake" (the keyboard specs, 2026-10-16) turned out to be a real race in the console's route handling. **To do** 1. Reproduce under load (the full lane's browser stage, or `--repeat-each` with more workers than cores) and read what keeps `[data-jump-latest]` after Escape. 2. Fix the cause in the console, or the spec if it asserts a moment the console doesn't promise.
Author
Owner

Two more specs that fail only under load, seen 2026-10-18 while other test lanes ran on the same laptop (load average 12-49):

  • e2e/follow-transition.spec.ts:547 "the stage › a click on a script output's fold during the arrival folds it": [data-transcript] button[aria-expanded] is not on the page. Failed once in a full lane on 1330513, once in a fast lane on ef05d28 (a tree without the top-bar change), and 1 of 5 in --repeat-each 5 --workers 3. It passes on its own and in quieter lanes. The fold button is looked for while the stage's arrival transition is still running, so it may not be there yet.
  • e2e/keyboard.spec.ts:35 "the go-to sequences reach every page": g i → inbox read runs after 10 s, once in that full lane. Then 5 of 5 passed on their own. The same spec was fixed for a real route race in aa31a26, so worth checking it is only load this time.

All three in this issue share a shape: a spec waits on a moment of an animation or a route change that a loaded machine reaches late.

Two more specs that fail only under load, seen 2026-10-18 while other test lanes ran on the same laptop (load average 12-49): - `e2e/follow-transition.spec.ts:547` "the stage › a click on a script output's fold during the arrival folds it": `[data-transcript] button[aria-expanded] is not on the page`. Failed once in a full lane on 1330513, once in a fast lane on ef05d28 (a tree without the top-bar change), and 1 of 5 in `--repeat-each 5 --workers 3`. It passes on its own and in quieter lanes. The fold button is looked for while the stage's arrival transition is still running, so it may not be there yet. - `e2e/keyboard.spec.ts:35` "the go-to sequences reach every page": `g i → inbox` read `runs` after 10 s, once in that full lane. Then 5 of 5 passed on their own. The same spec was fixed for a real route race in aa31a26, so worth checking it is only load this time. All three in this issue share a shape: a spec waits on a moment of an animation or a route change that a loaded machine reaches late.
Author
Owner

One more of the same shape, in the deploy's image self-check (braid-selfcheck) on the slow runner, 2026-10-18: the solo simulated run failed on 1330513 after 166 s (21 s on a laptop): no event osf.session.forked, step openspec.apply.solo, from_step='openspec.propose #1' and "openspec.apply.solo does not continue openspec.propose". It passed 3 of 3 locally, and the next deploy's self-check (ead9f7e, a superset) passed it. A fork is skipped without an event only when no parent session is found for openspec.propose #1, which fits propose needing a second attempt on a machine that slow. The scenario should name the parent step, not its first attempt, or wait for the attempt that succeeded, so one slow propose does not fail a deploy.

One more of the same shape, in the deploy's image self-check (braid-selfcheck) on the slow runner, 2026-10-18: the `solo` simulated run failed on 1330513 after 166 s (21 s on a laptop): `no event osf.session.forked, step openspec.apply.solo, from_step='openspec.propose #1'` and "openspec.apply.solo does not continue openspec.propose". It passed 3 of 3 locally, and the next deploy's self-check (ead9f7e, a superset) passed it. A fork is skipped without an event only when no parent session is found for `openspec.propose #1`, which fits propose needing a second attempt on a machine that slow. The scenario should name the parent step, not its first attempt, or wait for the attempt that succeeded, so one slow propose does not fail a deploy.
Author
Owner

Closing as obsolete. The three browser specs named here (ui-polish-103.spec.ts, follow-transition.spec.ts, keyboard.spec.ts) were deleted when the suite was rebuilt (#241); ui/e2e/ now has 18 journeys, which have run green on every CI run since. The solo simulated run from the last comment passes in ./tools/sim.sh (only e2e-lane-inferred fails, #296). If a journey fails under load again, it gets a new issue.

Closing as obsolete. The three browser specs named here (`ui-polish-103.spec.ts`, `follow-transition.spec.ts`, `keyboard.spec.ts`) were deleted when the suite was rebuilt (#241); `ui/e2e/` now has 18 journeys, which have run green on every CI run since. The `solo` simulated run from the last comment passes in `./tools/sim.sh` (only `e2e-lane-inferred` fails, #296). If a journey fails under load again, it gets a new issue.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#193
No description provided.