commands::tests::list_documents_merges_assistant_and_meeting_sources_by_recency (tauri/src-tauri) fails when the full minutes-app suite runs, and passes when run in isolation.
Confirmed on macOS (Apple Silicon), with controls to rule out a dependency change:
| Branch |
How run |
Result |
| origin/main |
full suite |
FAIL |
| origin/main |
isolation |
pass |
| desktop dep batch (#590) |
full suite |
FAIL |
| desktop dep batch (#590) |
isolation |
pass |
So it is pre-existing on main and independent of the dependency bumps that surfaced it.
The pass-alone / fail-together pattern points at shared mutable state or an ordering assumption rather than a logic bug: the test merges assistant and meeting sources and asserts recency ordering, so a likely cause is another test writing into the same directory or timestamps colliding within filesystem granularity when tests run concurrently.
This is the third flake of this shape, after #529 (capture relay reconnect) and #566 (bump-version readiness marker). Both of those were reddening unrelated contributor PRs before they were fixed, which is the real cost here.
Wanted: make the test independent of sibling tests (own temp directory, explicit timestamps rather than wall-clock ordering) rather than serializing the suite.
commands::tests::list_documents_merges_assistant_and_meeting_sources_by_recency(tauri/src-tauri) fails when the fullminutes-appsuite runs, and passes when run in isolation.Confirmed on macOS (Apple Silicon), with controls to rule out a dependency change:
So it is pre-existing on main and independent of the dependency bumps that surfaced it.
The pass-alone / fail-together pattern points at shared mutable state or an ordering assumption rather than a logic bug: the test merges assistant and meeting sources and asserts recency ordering, so a likely cause is another test writing into the same directory or timestamps colliding within filesystem granularity when tests run concurrently.
This is the third flake of this shape, after #529 (capture relay reconnect) and #566 (bump-version readiness marker). Both of those were reddening unrelated contributor PRs before they were fixed, which is the real cost here.
Wanted: make the test independent of sibling tests (own temp directory, explicit timestamps rather than wall-clock ordering) rather than serializing the suite.