`Build & E2E` is red on `main`: the console smoke test finds neither app shell nor loading screen

[yinlianghui]

Found while verifying PR #4084 (issue #4069). Filed unassigned — not caused by
that PR, and outside its scope, but it turns every open PR's Build & E2E gate
red, so it is reported rather than absorbed.

What fails

[chromium] › e2e/smoke.spec.ts:78:3 › Console App – Smoke › should show the app shell or loading screen
  1 failed
  23 skipped
  9 passed (1.8m)

Error: expect(locator).toBeVisible() failed
Locator: locator('nav').first().or(getByText(/Initializing|Loading|Connecting/i).first())
Expected: visible
Error: element(s) not found
  - Test timeout of 30000ms exceeded.
  at e2e/smoke.spec.ts:88:7

Neither the app shell's nav nor the loading placeholder ever appears, through
all three Playwright retries — the console app does not reach a rendered state at
all, rather than rendering something unexpected.

It is on main, not on a branch

The identical failure — same job, same spec line, same
1 failed / 23 skipped / 9 passed tally — appears on main itself:

Run	Branch / commit	Build & E2E
2026-08-10T04:55Z	main	success
2026-08-10T05:28Z	main	success
2026-08-10T06:15Z	main @ 9154d9e90 (run 31361301040)	failure
2026-08-10T07:30Z	PR #4084, branched off 9154d9e90 (job 93384727221)	failure

9154d9e90 is the merge commit of #4081 (fix(plugin-grid): object-grid publishes the filter key it reads — filter, singular). It is the first run where
this goes red, and the two runs before it are green, so that merge is the prime
suspect — though the smoke test only proves the app never boots, not why, so the
cause is worth confirming rather than assuming.

On PR #4084 that job was the only red of twenty checks; Lint, Type Check and all
four test shards were green.

Why it matters beyond one PR

Build & E2E is not informational (Live E2E is the informational one, and it
passed). While this is red on main, every PR branched from it inherits a red
gate, which is exactly the condition that trains reviewers to merge past reds —
and a genuine regression then rides main's merge ref into every later PR's run
with nothing distinguishing it from the known red.

Suggested first step

Pull the failing run's error-context.md / screenshot attachment (the job
uploads none — No files were found with the provided path: playwright-report/,
which is worth fixing on its own so the artifact survives), or run
pnpm exec playwright test e2e/smoke.spec.ts locally against main and read the
browser console at boot.

---

Generated by Claude Code

[yinlianghui]

Claim: PM loop round 4 — main-health, dispatched at top priority (objectui whole-repo seat). target:v17 added: a red required gate on main during the release window blocks every open PR (auto-merge on #4084 is armed and waiting on exactly this) and trains merges past reds.

Session: session_017Qqyix2QcnpUC9XeYVDzx3
Branch: claude/issue-4086-main-smoke-red
Worktree: objectui-issue-4086
Container & model: M, mode:subagent, model: opus

Working hypothesis for the dev to CONFIRM OR REFUTE, not assume: first red is 9154d9e90 (#4081's merge); the two runs before are green; the failure reproduces identically on a second independent run (PR #4084's, branched off that SHA) — so flake is unlikely and #4081 is the prime suspect. #4081's runtime half routes ObjectGrid's $filter through toFilterNode; unit tests, type-check and all four vitest shards are green while the BUILT console never reaches paint — consistent with a build-graph-only failure (e.g. a module-level import cycle or a deep import that the bundler resolves differently than vitest). Disposition ladder: minimal fix-forward if root cause is small and provable; otherwise revert #4081's merge on main (this seat accepted #4081 and explicitly authorizes its revert — reopening #4041 — rather than leaving main red while a fix is engineered).

---

Generated by Claude Code

[yinlianghui]

ACCEPT — with the premise corrected on the record (objectui seat PM, session session_017Qqyix2QcnpUC9XeYVDzx3) — PR #4095.

#4081 is exonerated. The evidence is conclusive: main went green at 8497579db with every line of 9154d9e90 still present, then red again at 47737ecb3 (an unrelated dependabot bump) — green/red/green/red across four unrelated commits, the suspect present on both sides. The import-cycle hypothesis this seat put in the claim comment is also structurally impossible (toFilterNode joined an import that already pulled @object-ui/core; the module graph never changed). This seat's working hypothesis was wrong; the dev proved it wrong before touching anything — exactly the discipline the dispatch asked for.

Real cause: a race in the smoke test itself. The signed-out console boots → paints the splash at ~70 ms → redirects to the sign-in screen at ~105 ms and stays there. nav is never entitled to render (no session); the loading text lives ~35 ms. The old assertion (nav OR loading text) only matched a state that exists for a coin-flip window, so runner speed decided the color — which is why it flipped twice today with no related code change. Fix: the assertion now names all three recognised terminal boot states (shell / splash / sign-in), matched structurally (login-config-loading testid, #login-email, #login-phone), keeping the set closed so a blank page or crashed render still fails. Reverse-verified both directions with the settle forced (the honest construction, since the bug IS the race): old assertion red with the CI text verbatim, new one green. Rider fixed: CI's github reporter writes no playwright-report/, so every red run discarded its own evidence — test-results/ (screenshot, trace, error-context) is now uploaded.

On PR #4095: Build & E2E itself is GREEN (the gate this card exists for), along with Type Check / Lint / Control Byte / Changeset Declaration / Build Docs / Live E2E. Shards 1/4 and 2/4 are red — inherited from main's NEXT breakage (this morning's dependabot wave broke two generator-template dependency ratchets), proven unrelated (a vitest shard runs neither e2e/ nor .github/), filed as #4098 and being dispatched at top priority now. #4095 is flipped ready with auto-merge armed and lands when the shards clear; #4084 queues behind the same gate.

---

Generated by Claude Code