[chriscoveries]

SettingsModal.guidance.test.tsx › "keeps expanded Guidance 0 keyboard focus and Escape in the child dialog" fails intermittently on CI with AssertionError: expected undefined to be defined at line 98 after about 1.39 s. Only case 0 fails. Case 1, the same code path, passes in about 100 ms. The failing case logs React's "A suspended resource finished loading inside a test, but the event was not wrapped in act(...)".

Root cause. SettingsModal loads GuidanceTab with React.lazy. In case 0 the chunk is still loading when the render's act() scope exits. When it resolves, no act() scope is open. React then commits the tab, runs its mount effect and its prompt load on the Scheduler, while the test polls with vi.waitFor. The test's readiness check (two textareas) passes before the tab has settled: the textareas render before the prompt load finishes, and the expand button stays disabled while the load runs. Its follow-up renders also stay off act(). On a loaded runner the portaled expanded dialog is then not in the document within vi.waitFor's 1000 ms default. Case 1 never suspends because the lazy component is already resolved.

Fix (test-only). Wait for the tab in back-to-back act() scopes instead of vi.waitFor. An act() scope is then open whenever the chunk resolves, and the tab's retry render, effects and prompt load are flushed by act() before the click. A single act() around vi.waitFor does not work, because act() flushes queued work only when its outermost scope exits. The DOM would never update while polling. Readiness now has a 4000 ms deadline followed by an exact two-textarea assertion, so a lazy-chunk failure reaches that assertion before the default 5 s test timeout.

Why not a longer timeout. Raising the vi.waitFor timeout only widens the window for Scheduler work to land. The test would still depend on wall-clock timing, would fail more slowly when it fails, and would hide that the tab mounts outside act().

Verification. The test-only red run (39806277470, using a reference commit based on upstream main at f1d27838) forces the lazy chunk to resolve after render act(): current fails and fixed passes (the Web job reports 1 failed and 2,912 passed). The exact fix head de391350e9383881fd412d6de7a0efad12248b53 passed all six GitHub Actions jobs on the fork, including rust-linux (run 39820886384). The gated tests remain on the fork, not in this PR.

Test environment: GitHub Actions CI on the fork (ubuntu-latest, Node 24.19.0, npm -C web run check and npm -C web test).

[chriscoveries]

CI note: on current main, rust-linux fails the five real-Chrome fixtures added in 1e5651b2 on every pull request. #175 moves them to the nightly job.