[Test] stabilize the Firefox blog listing E2E visit / 稳定 Firefox 博客列表 E2E 访问

[edwingao28]

Summary

Stabilize the recurring Firefox E2E failure Blog › Blog listing page › "before all" hook for "renders the blog page with heading" (Timed out after waiting 60000ms for your remote page to load).

Evidence. Across the last 30 failed tests-e2e runs, 16 failed on exactly this hook, all on E2E (firefox, shard 1) and never on Chrome (examples: 37076830984, 36959125845, 36952749489, 36622355367). In passing runs the same listing visit plus first assertion completes in ~200 ms in both browsers, so this is a lost load event, not a slow page. The hang survived the earlier Substack-only 1×1 stub (#1166), so a remote-only stub is not sufficient.

Why the hook fails wholesale. The listing visit lived in a before (before-all) hook. Cypress does not retry before-all hooks, so one lost load event failed the hook and skipped the other three listing tests. The visit's load event depended on every eager card thumbnail (featured card + first grid row) passing through /_next/image, one of which is a remote Substack source; Firefox routes all of that through the Cypress HTTP/1.1 proxy.

Change (packages/app/cypress/e2e/blog.cy.ts only).

• Stub every /_next/image request on the listing with the existing 1x1.png fixture, so the visit never waits on image resizing or a remote thumbnail host.
• Visit per test in beforeEach with a 20 s bound, so the configured retries.runMode: 2 covers a lost load event at bounded cost.
• Add serves local card thumbnails through the image optimizer, which cy.requests the first local thumbnail's real optimizer URL (200 + image/*), keeping optimizer coverage the stub would otherwise remove.

No production code changes. Other before hooks in the spec have not flaked and are left as they are.

AI model disclosure

• Model/version: Claude Fable 5.1 (identifier as exposed by the Perplexity Computer runtime).
• Role: CI log analysis, local reproduction attempts, fix, verification and PR preparation.
• No delegated agents contributed.
• I have completed the AI model disclosure and kept it current

Validation

Local production build (E2E_FIXTURES=1 bun run build && bun run start), Cypress 15.21.1, Firefox 157 and Electron.

• Repro attempts of the raw flake: 20+ cold-profile Firefox runs of blog.cy.ts and one full Firefox shard-1 replay (11 specs in CI order) all passed, so the raw race could not be forced locally. The failure mechanism was simulated instead:
• Before (master spec) with a 70 s stalled eager local thumbnail, Firefox: "before all" hook … Timed out, Skipped: 3, no retry (same signature as CI).
• After with the same stall, Firefox and Electron: 16/16 passing, 0 skipped.
• After with the first /blog document load stalled once: (Attempt 1 of 3) then 16/16 passing, 0 skipped. Confirms the retry path and bounded cost.
• Unstubbed blog.cy.ts on the new spec: 16/16 passing in Firefox.
• bunx oxlint, bunx oxfmt --check on the changed file, bun run typecheck, bun run check:typography: pass.

Not verified: the raw CI race itself (not reproducible locally); CI results will be reported after this PR's checks finish.

中文说明

修复 Firefox E2E 中反复出现的 Blog › Blog listing page before-all 钩子超时（60 秒内页面未触发 load 事件）。最近 30 次失败的 E2E 运行中有 16 次正是该钩子，且全部发生在 E2E (firefox, shard 1)，Chrome 从未出现；通过的运行里同一访问加首个断言只需约 200 毫秒，说明是 load 事件丢失而非页面变慢。之前仅针对 Substack 缩略图的 1×1 桩（#1166）未能消除该问题。

原因：列表页访问放在 before-all 钩子中，而 Cypress 不会重试 before-all，一次 load 丢失就导致钩子失败并跳过其余三个用例；该 load 事件依赖首屏所有缩略图经 /_next/image 加载（其中一张来自远端 Substack），Firefox 下这些请求全部经由 Cypress 的 HTTP/1.1 代理。

改动仅限 packages/app/cypress/e2e/blog.cy.ts：统一用 1x1.png 桩替换列表页全部 /_next/image 请求；改为在 beforeEach 中按用例访问并限制 20 秒超时，使已配置的重试生效；新增用例直接请求首个本地缩略图的真实优化器地址（200 且 image/*），保留对图片优化器的覆盖。未改动生产代码。

验证：本地生产构建 + Cypress 15.21.1（Firefox 157、Electron）。20 余次冷配置 Firefox 运行及完整 shard-1 顺序回放均未能复现原始竞态，因此改为模拟机制：旧用例在首屏缩略图停滞 70 秒时出现与 CI 相同的 before-all 超时并跳过 3 个用例；新用例在同样条件下 Firefox 与 Electron 均 16/16 通过；仅使首个 /blog 文档加载停滞时，新用例经一次重试后全部通过。lint、格式、类型检查与排版检查通过。原始 CI 竞态本身无法在本地复现，CI 结果将在检查完成后汇报。

AI 模型：Claude Fable 5.1（Perplexity Computer 运行环境提供的标识），负责 CI 日志分析、本地复现尝试、修复、验证与 PR 准备；未使用委派 agent。

---

Note

Low Risk
Changes are limited to Cypress E2E in blog.cy.ts; no production or runtime behavior is modified.

Overview
Stabilizes flaky Firefox E2E on the blog listing suite by changing how /blog is visited and how card images are stubbed.

The listing hook moves from before → beforeEach with a 20s visit timeout so Cypress run-mode retries can recover from a lost load event instead of failing the whole block. cy.intercept now stubs every GET /_next/image with the existing 1x1.png fixture (not only Substack URLs), so eager featured/grid thumbnails do not block the page load on resizing or remote hosts.

A new test serves local card thumbnails through the image optimizer uses cy.request against the first local grid thumbnail’s real /_next/image URL to assert 200 and an image/* content type, preserving optimizer coverage that the broad stub removes from the browser path.

^{Reviewed by Cursor Bugbot for commit aa3be4b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

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

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Oct 3, 2026 8:03am UTC

[github-actions]

Claude finished @edwingao28's task in 1m 44s —— View job

---

Review complete

• Read the changed spec in full context
• Analyze the stabilization approach and new test correctness
• Check CI status
• Post review

🟢 LGTM — no blocking issues found.

This is a well-reasoned, test-only change (packages/app/cypress/e2e/blog.cy.ts). I verified each assumption the diff relies on:

• beforeEach + per-test cy.visit('/blog', { timeout: 20_000 }) correctly leverages retries.runMode: 2 (confirmed in cypress.config.ts:13-16). Moving off the before-all hook is the right fix: Cypress doesn't retry before-all hooks, so a single lost Firefox load event previously failed the hook and skipped the whole suite. Re-registering the intercept in beforeEach is also necessary, not just stylistic — cy.intercept routes are cleared between tests even with testIsolation: false, so a stub left in the old before would not have applied to tests 2–4.
• Stubbing all /_next/image removes the remote-thumbnail / optimizer-resize work from the visit's load event, which matches the diagnosis (lost load, not a slow page).
• New serves local card thumbnails… test preserves the optimizer coverage the stub removes. The key correctness detail holds: cy.request is not intercepted by cy.intercept, so it hits the real optimizer rather than the fixture. The selector img[src^="/_next/image?url=%2Fimages%2F"] is valid — optimizedThumbnailSrc (src/lib/blog-thumbnail-src.ts:13) emits exactly /_next/image?url=%2Fimages%2F…&w=…&q=75 for local figure thumbnails, and data-testid="blog-post-grid" exists (blog-index-content.tsx:126).

One minor, non-blocking observation: the new test's .first() assumes at least one grid card carries a local /images/ thumbnail (posts without a figure render the text-free tile with no <img>). That holds today given the published posts with figures; if the listing ever showed only figure-less posts the .first() would fail with no match — acceptable for a test given current content.

No production code, no user-visible Chinese strings, and no new interactive elements, so coverage/analytics//zh requirements don't apply. The change adds coverage rather than removing it. CI: Lint & Format passed; Unit and E2E were still in progress at review time.
· branch fix/blog-listing-e2e-firefox-load