Repository navigation
test(e2e): onboarding flakes waiting for the three enabled free models (#2431 follow-up) #2436
Description
Symptom
e2e_shard (1/2) fails in e2e/onboarding.spec.ts at the enabled-models poll:
expect(received).toEqual(expected)
"enabledModelIds": [
"nemotron-3-ultra-free",
- "mimo-v2.5-free",
- "deepseek-v4-flash-free",
],
Call Log: Timeout 10000ms exceeded while waiting on the predicate
Observed three times today:
- on main at
ede1f12f3(run 31171345408) — passed on re-run - twice in a row on fix(desktop): treat a blank UA-CH platform as absent in Astryx's mod-key probe #2418's merge refs (both after rebases onto current main; the PR touches only Astryx's Kbd/useHotkeys platform probe, patches/, and a keyboard-help e2e)
The state is not slow-arriving — the poll runs 10s and enabledModelIds stays ["nemotron-3-ultra-free"], i.e. some writer produced a nemotron-only connection and nothing upgrades it afterwards.
Suspicion
A first-run race between the writers of opencode-free:
- background bootstrap (
app-lifecycle.ts— seeds v3 with the three models when the store is empty, or migrates viaresolveOpenCodeFreeBootstrapMigration) - whatever creates/updates the connection during the onboarding save flow (
保存供应商on a fresh profile)
resolveOpenCodeFreeBootstrapMigration only recognizes two untouched shapes (v1 big-pickle, v2 nemotron + extras.version 2). If the onboarding-era writer lands a nemotron-only connection after the migration pass has already run (or in a shape the matcher doesn't recognize — e.g. no extras), the connection is permanently stuck at one model and the test's expectation can never be met — matching the 10s-flat poll.
Notes
- Flake rate looks high under CI load (3 hits today across main + one PR).
- fix(desktop): treat a blank UA-CH platform as absent in Astryx's mod-key probe #2418 is currently red on this; unrelated to its content.
Activity
GabrielDrapor commentedon Aug 7, 2026
Update — this is not a timing flake; it reproduces deterministically in isolation on macOS at main tip (
npx playwright test e2e/onboarding.spec.ts --workers=1fails 3/3 locally). CI's occasional green is the race landing the other way.Root cause, pinned
Probed the connection right after onboarding's
保存供应商(fresh profile):{ "slug": "opencode-free", "defaultModel": "nemotron-3-ultra-free", "enabledModelIds": ["nemotron-3-ultra-free"], // no extras — no makaBootstrap version marker // createdAt === updatedAt — created in one shot by the save }The writer is the generic add-provider path, not the bootstrap seed:
provider-add-form.tsxcallsbridge.create({ slug, name, providerType, defaultModel, … })with noenabledModelIdsand noextras, and the create handler completes the shape to[defaultModel]. Two consequences:OPENCODE_FREE_DEFAULT_ENABLED_MODELS, three models) never applies —app-lifecycle.tsonly seeds when the store is empty, and by then it isn't.resolveOpenCodeFreeBootstrapMigrationcan never repair it: the shape matcher requires either v1 (big-pickle) or v2 (nemotron +extras.makaBootstrap.version === 2), and this connection has no extras. It is permanently stuck at one model.So whenever the user (or the e2e) reaches the save button before the background bootstrap wins the race, the three-model expectation from #2431 can never be met.
Fix options
connections:createpath createsproviderType: 'opencode-free', source the full v3 seed shape (enabledModelIds+extrasmarker) frombootstrap-connections.ts— one exported helper, single source of truth.Happy to send the PR for the first option if useful.
GabrielDrapor commentedon Aug 7, 2026
Fixed by #2443 (fills the omitted enabledModelIds at the connection-store create seam). Verified locally on main tip: the previously deterministic failure now passes 4/4 (
npx playwright test e2e/onboarding.spec.ts --workers=1), and #2418's CI — which was blocked on this — is green after rebasing. Closing.