fix(e2e): green main — retarget stale specs + fix update-preferences flake

[eliotlim]

Problem

main CI had been red for 8+ consecutive runs. The Playwright e2e job is a non-required check, so recent PRs (the AI-usage batch #111–#115, the island-hydrated viewer rearch, the identity work) merged over it and the failures accumulated. Latest red run: https://github.com/eliotlim/OpenBook/actions/runs/28804748447

Recon verdict: every consistent failure is a stale/mis-scoped test trailing an intentional, correct product change — not a product regression. One was a pre-existing flake. No product code needs to change.

Board: Epic: Green main — fix failing e2e (2026-07) → T1 (stale-locator retargets) + T2 (update-preferences flake).

Solution

Test-only fixes across 6 e2e specs + one shared fixture. Why test-side, not product-side: each failure is a test asserting old behavior after the product deliberately changed (verified against source, commit-attributed):

• account-switcher — the account email now legitimately renders in more chrome (ProfileMenu + SettingsPanel ProfileChip, since 406d174), so an unscoped getByText hit 3 nodes. Scoped the two email assertions to their [data-account-id] switcher rows (strictly stronger).
• ai — the server now redacts provider API keys unconditionally (fix(ai): server-only provider API key + write-only settings entry #111); GET /api/ai/status returns apiKeySet:true, not the raw key. Poll asserts the presence flag.
• block-editor / kit — block-doc "Interactive HTML" export now mounts the vendored OpenBookViewer (.obe-* DOM) instead of the retired bespoke reactive runtime (e177f38). Retargeted the offline-compute assertions to the real hydrated-viewer DOM; all numeric proofs preserved (6→18, 70/30, chart rect count 3, tick 20→5).
• export (flake hardening) — restore now proven by the name-pinned … (imported) twin instead of a raw count delta; poll windows widened.
• update-preferences (flake) — the real 100%-repro failure was the General-tab click racing the async shell re-init after page.reload() (not the originally-reported popover intercept). Hardened openGeneralSettings (expect(...).toPass(), reopens on transient remount) and made chooseValue/chooseLabel wait for popover teardown.

Key files: packages/web/e2e/{account-switcher,ai,block-editor,kit,export,update-preferences}.spec.ts, packages/web/e2e/fixtures.ts. New deps: none. No product/runtime code touched.

Before / After

No visual change (test-only PR) — behavioral state deltas below.

Behavior	Before	After
account-switcher:52	strict-mode violation: email matched 3 nodes → fail	email asserted in the correct switcher row → 3 passed
ai:46	polled redacted apiKey → never equals test key → fail	asserts apiKeySet:true → 8 passed
block-editor:377	asserted retired .reactive/[data-val] DOM → element not found	asserts hydrated .obe-* viewer DOM, 6→18 → 1 passed
kit:236	asserted retired bespoke selectors → element not found	asserts viewer DOM (formula/status/chart/number), 70/30 → 1 passed
update-preferences:21	General-tab click raced post-reload shell settle → 30s timeout	resilient open + popover-teardown waits → 20/20 under --repeat-each=20
export:48	restore proven by fragile raw count → flaky	proven by name-pinned imported twin → passed

Test procedure

Per-spec, on the web e2e harness (pnpm --filter @book.dev/web exec playwright test <spec> --project=chromium); the flaky ones under load:

playwright test e2e/account-switcher.spec.ts   # 3 passed
playwright test e2e/ai.spec.ts                 # 8 passed
playwright test e2e/block-editor.spec.ts:377   # 1 passed
playwright test e2e/kit.spec.ts:236            # 1 passed
playwright test e2e/export.spec.ts:48          # passed
playwright test e2e/update-preferences.spec.ts --repeat-each=5   # 20 passed

The full sharded suite runs in this PR's CI — that is the real proof main goes green (and that the shared fixtures.ts change has no blast radius across other specs).

Operational notes

• No deploy steps, no env vars, no migrations. Reversible (test-only).
• The shared chooseValue/chooseLabel fixture change touches every spec that opens a Select; reviewed for blast radius (the added wait is a no-op when no options are open) and proven by the full CI e2e run.
• Follow-up (owner): once this is green, make the Playwright e2e check required on main so PRs can't merge over red e2e again — the root cause of the 8-run pileup.

---

Verify: T1 — typecheck/lint/unit green (sdk 171 / ui 948 / server 645), 5 specs green individually. T2 — full pnpm verify exit 0, spec 20/20. Merge is the clean union of both (7 files, 60/30), no conflicts.
Reviews: code (Quinn) — T1 cleared; T2 in review. No security/design/product gates (test-only, no trust boundary, no user-visible change).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
app.book.pub	Ready	Preview, Comment	Jul 9, 2026 6:59am