Repository navigation
Flaky CI tests: session-history E2E anchor, Windows git-info refresh, Windows setup-config lock timeout #829
Description
Three tests failed intermittently in CI on 2026-10-14. In each case the code under test was untouched, and rerunning only the failed jobs passed. They block the required checks, so each flake costs a rerun before the next merge.
1. tests/web/session-history.e2e.ts:340: prefetched native history loads on upward reading without moving the visible anchor
- Failing run: https://github.com/openpi-dev/openpi/actions/runs/38539593564/job/115796543639 (main @
7ac1f3e, Web E2E shard 2/2) expect(...).toBeLessThanOrEqual(2)at line 451 received13818: the scroll position was not kept near the bottom of the live view.
2. Windows: tests/extensions/git-info/index.test.ts:30: automatic refresh with a spaced fixture path stays local until /pr explicitly requests GitHub data
- Failing run: https://github.com/openpi-dev/openpi/actions/runs/38550635790/job/115807447499 (main @
534d144, Runtime Node Windows parallel 2/2) - The assertion at line 167, "the initial Git fixture refresh must recognize the repository", got
false. This looks like a timing issue on slow Windows runners.
3. Windows: tests/extensions/shared/setup-config.test.ts:661: a reused PID does not impersonate the dead lock owner
- Failing run: https://github.com/openpi-dev/openpi/actions/runs/38558753554/job/115834312456 (PR feat(web): add Configuration, Appearance and Archived chats settings #819, Runtime Node Windows parallel 1/2)
- Error:
Timed out waiting for another Pi process to finish updating ...\my-pi-setup.json. The file-lock wait timed out.
Also observed once today (same pattern, passed on rerun)
tests/web/document-reading-state.spec.ts:234: keeps the selected worksheet across the same tool remount (UI windows-latest, main @7d3640d, run https://github.com/openpi-dev/openpi/actions/runs/38571951985)tests/web/settings-appearance.e2e.ts:371: archived chats query project/sort before paging (PR ci: pin shared automation workflows to a full commit SHA #724, run https://github.com/openpi-dev/openpi/actions/runs/38570344248)
Suggested direction
For each test, fix the root cause: wait on the actual readiness signal instead of wall-clock time or a fixed timeout, and make lock-wait budgets scale on Windows. Don't just add retries or longer timeouts.
Activity
627150795 commentedon Oct 14, 2026
我来处理其中第 3 项：Windows 上 setup-config 锁等待超时及“重用 PID 不会冒充已退出锁持有者”的回归覆盖。范围仅限这条锁等待/所有权路径，不处理 session-history E2E 与 git-info 刷新两项。
test: fix root causes of two flaky tests (#829) (#831)
tt-a1i commentedon Oct 14, 2026
Status update:
settings-appearance.e2e.ts:371was a real store bug: an accepted restore was reported as unconfirmed whenever a concurrent snapshot refresh superseded its own. Locally it failed 4/16 before the fix and passed 16/16 after; a unit test covers it.git-info/index.test.ts:30was a fixture cold start: the first exec of a freshly copiednode.exeexceeded the 3 s Git budget. The fixture is now warmed up before timing.session-history.e2e.ts:340: superseded. The failing bottom-of-view assertion was replaced in feat(web): improve resource settings, usage and conversation reading #807 (b79bb1d), and 8/8 local repeats pass onmain. Leave this open for now and close it if it doesn't recur.setup-config.test.ts:661: diagnosed, not fixed.queryProcessStartedAt(extensions/shared/setup-config.ts) spawnspowershell.exe -Command Get-Processwith a 1 sexecFiletimeout.undefined,processLivenessreportsunknown, and recovery fails closed until the 5 s lock deadline expires.wmic-free query with an appropriate budget, verified on Windows. I didn't change this without being able to verify it.document-reading-state.spec.ts:234: not reproduced (0/15 local runs). Still open.tt-a1i commentedon Oct 15, 2026
New flaky test, observed on Windows:
tests/web/session-sidebar.spec.ts:239: "offers an empty pin drop area during a project drag and suppresses the trailing pointer click" (UI windows-latest Node 22.19.0). Failing run: https://github.com/openpi-dev/openpi/actions/runs/38760204307/job/116509157410 (PR fix(web): follow Anthropic model catalog pagination #722, whose change to provider model discovery is unrelated). The test was added with the sidebar drag work in feat(web): add sidebar actions, ordering and conversation fork UI #822. Not yet investigated.tt-a1i commentedon Oct 15, 2026
Another Windows UI flake, seen on main after #714 merged:
tests/web/provider-models-settings.spec.ts:474: "keeps API drafts across all three add modes and writes nothing when cancelled" (UI windows-latest Node 22.19.0) failed withTest timed out in 5000ms. Run: https://github.com/openpi-dev/openpi/actions/runs/38764433340/job/116523827850 (fix(setup): reject malformed UTF-8 before config mutation #714 only touches setup-config UTF-8 decoding, so it is not the cause). When that job was rerun, it failed ontests/web/session-sidebar.spec.ts:239instead.tt-a1i commentedon Oct 15, 2026
Another Windows UI flake, seen on main (c041b3e) when the job was rerun:
tests/web/workbar-session-position.spec.ts:577: "reattaches a Session's file position after a copied same-ID Session, releasing the old preview" (UI windows-latest Node 22.19.0) failed withTestingLibraryElementError: Unable to find role="button" and name "README.md". Run: https://github.com/openpi-dev/openpi/actions/runs/38764433340The same Windows UI job has now failed three times on one commit, each time on a different test, so the Windows UI suite looks generally timing-sensitive.
tt-a1i commentedon Oct 15, 2026
Another flake, this time in Linux Web E2E, seen on #722 after it was synced with main:
tests/web/openpi-web.e2e.ts:2510: "recovers an unknown prompt admission only after an explicit user decision" (Web E2E shard 1/2) failed withapiResponse.json: Response has been disposed, followed bybrowserContext.close: Protocol error. Run: https://github.com/openpi-dev/openpi/actions/runs/38774705615/job/116559219255. It passed on rerun. fix(web): follow Anthropic model catalog pagination #722 only changes provider model discovery pagination, so it is not the cause.tt-a1i commentedon Oct 15, 2026
Another Windows timing flake, seen on main (4418140) after #722 merged:
tests/extensions/ai-providers/antigravity.test.ts:1530: "streamAntigravity bounds the full SSE lifetime after the first event" (Runtime Node, Windows parallel 1/2) failed withTimed out waiting for the first SSE event; the test expected the/next SSE event/timeout. It looks like a race between the first-event and full-lifetime timers, which are only a few ms apart. Run: https://github.com/openpi-dev/openpi/actions/runs/38776735793/job/116566218404 (fix(web): follow Anthropic model catalog pagination #722 does not touch antigravity).fix: separate readiness from fixed clocks in three CI flakes (#861)