Skip to content

test(website): guard that models pages show content, not a loader, and don't shift - #19394

Closed
MaanilVerma wants to merge 7 commits into
maanil/fe-2942-model-page-ssrfrom
maanil/fe-2985-content-cls-guard
Closed

MaanilVerma wants to merge 7 commits into
maanil/fe-2942-model-page-ssrfrom
maanil/fe-2985-content-cls-guard

Conversation

@MaanilVerma

@MaanilVerma MaanilVerma commented Sep 29, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Adds an e2e guard that model pages and the /models/ catalogue don't leave a loader on screen and don't jump around (CLS < 0.1), with the workshop flag on and off.

Fixes FE-2985 (https://linear.app/comfyorg/issue/FE-2985/automated-guard-models-pages-show-real-content-not-a-spinner-and-dont)

Stacked on #18899.

Changes

  • New e2e/models-render-guard.spec.ts. Pages: the shared FLUX 2 Max page, the first video generation model in the published catalogue, the first catalogue model whose examples have empty values (today Bria Eraser, which has samples but no example values), and /models/.
  • JavaScript on, at 393px and 1280px from the shared VIEWPORTS, flag on and off: neither workshop-loading nor the post-mount models-loading is visible after 1 s, and CLS stays under 0.1. CLS comes from a layout-shift PerformanceObserver (same as mcp.spec.ts), read after the island hydrates and the Run area leaves resolving. No sleeps, no new dependency.
  • JavaScript off: /models/ should show the catalogue h1, no Grok showcase and a model link. It's test.fail() until feat(website): render the models catalogue heading and A to Z links as static HTML #18898, because the noscript ModelsShowcase renders there today.
  • h1, no-Grok and rollout-note checks on model pages are left to workshop-visibility.spec.ts and workshop-loading-flash.spec.ts.
  • New e2e/fixtures/workshopFlags.ts with stubWorkshopFlags(context, flags), now used by fixtures/modelsAccount.ts, workshop-loading-flash.spec.ts, workshop-visibility.spec.ts and this spec.
  • 17 tests, about 8 s locally.

Changed from the ticket:

Review focus

  • The 1 s loader grace: the worst local run was 774 ms on /models/ at 393px
  • Picking model pages from /models/catalogue.json plus workshop-display.json rather than fixed slugs

@MaanilVerma MaanilVerma added the cursor-review Trigger multi-model Cursor agent review on this PR label Sep 29, 2026
@MaanilVerma MaanilVerma self-assigned this Sep 29, 2026
@vercel

vercel Bot commented Sep 29, 2026 •

Copy link
Copy Markdown

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

2 Skipped Deployments
Project Deployment Actions Updated
billing-web Ignored Ignored Sep 29, 2026 8:07pm UTC
billing-web-test Ignored Ignored Sep 29, 2026 8:07pm UTC

Request Review

@coderabbitai

coderabbitai Bot commented Sep 29, 2026 •

Copy link
Copy Markdown
Contributor

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository: Comfy-Org/ComfyUI_frontend/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 491f6cf1-3c9a-4411-99c4-5eefca27eb2a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review

Comment @coderabbitai help to get the list of available commands.

@MaanilVerma

MaanilVerma commented Sep 29, 2026 •

Copy link
Copy Markdown
Collaborator Author

QA

Local production builds with the CI e2e env (WORKSHOP_IN_BUILD=1 PUBLIC_WORKSHOP_ROUTER_RUN=1 PUBLIC_WORKSHOP_CLOUD_ENV=test), served by astro preview, run with playwright test e2e/models-render-guard.spec.ts --project=desktop --retries=0.

# Check Proof Result
1 Green on this branch (base #18899 a6024e4bf2) --repeat-each=3: 60 passed (44.7s). Model pages, all 12 JS-on cases: loader never visible (loaderLastSeenAt=0). CLS desktop flag on FLUX 0.0057, Seedance 0.0057, Bria 0.0080; flag off 0 / 0.0005 / 0.0023. 390px flag on 0.0036 / 0.0036 / 0.0078; flag off 0 / 0 / 0.0052. JS-off h1 = model name on all three pass
2 Red on main (888e079ea9, content still gated) 15 unexpected failures out of 20. JS off, all three model pages: Expected: "FLUX 2 Max Text-to-Image" Received: "Grok Imagine in ComfyUI" (same for Seedance and Bria). JS on, flag off, both viewports: same Grok h1. JS on, flag on: the guard's own checks passed on a fast local server (loader last seen 375 to 945 ms, CLS 0.0025 to 0.0055) and the test failed on the fixture's Unexpected external requests: https://testcloud.comfy.org/api/features pass (red)
3 Hub cases are expected failures here JS off: Expected: "ComfyUI models" Received: "Grok Imagine in ComfyUI". JS on: Received: "What will you make next?" pass
4 Hub JS-off case flips once #18898 lands Same spec against comfy-website-preview-pr-18898.vercel.app, --grep "/models/ ": JS-off case Expected to fail, but passed pass
5 Hub JS-on cases still red after #18898 Against the #18898 preview the loader was last visible at 1444 ms (desktop, flag on), 3286 / 2413 ms (390px). Locally on this base, with the hub h1 check pointed at today's heading: loader last visible 2.2 to 5.4 s in all 12 runs, CLS 0 to 0.006. So the reason names the catalogue.json loader, not #18898 pass
6 Gates vue-tsc -p apps/website/tsconfig.json clean, eslint + oxlint exit 0, oxfmt applied, pnpm knip clean, pre-commit typecheck:website passed pass
7 Nothing else changed git diff a6024e4bf2 --stat: 1 file, 165 insertions pass
8 CI Not watched; the spec runs in the existing ci-website-e2e.yaml shards pending

Review round 1 (Ben's review, d64cddc740)

One build with WORKSHOP_IN_BUILD=1 PUBLIC_WORKSHOP_ROUTER_RUN=1 WEBSITE_INDEXABLE=1 PUBLIC_WORKSHOP_CLOUD_ENV=test, astro preview, --project=desktop --workers=4.

# Check Proof Result
9 Watcher now sees the post-mount loader /models/ loaderLastSeenAt: 393px flag on 774 ms, off 524 ms; 1280px on 559 ms, off 362 ms (was 0 with the old selector). Per-id probe: models-loading is the last loader visible (393px 419 ms after workshop-loading 373 ms; 1280px 273 after 240) pass
10 Model pages stay green All 12 model-page cases pass, loaderLastSeenAt=0, CLS max 0.037 (Bria, 393px, flag off). Pages resolved from /models/catalogue.json: bfl--flux-2-max, bfl--flux-3-text-to-video, bria--eraser pass
11 /models/ runtime reasons re-derived With the loader watched correctly all four runtime cases pass (CLS 0 to 0.0055), so their test.fail() is removed. JS-off hub still fails on Received: "Grok Imagine in ComfyUI" (noscript ModelsShowcase), reason updated. Row 5's 2.2 to 5.4 s loader was not reproduced on this build pass
12 Shared flag stub callers still pass workshop-loading-flash.spec.ts + workshop-visibility.spec.ts on the same build: 11 passed pass
13 Gates pnpm typecheck (website) 0 errors, pnpm exec eslint on the 5 e2e files exit 0, oxfmt --check apps/website/e2e/ clean, pre-commit hooks passed pass
14 1 s grace margin Worst loader case 774 ms vs 1000 ms budget, about 25% headroom locally under 4 workers watch in CI

Review round 2 (Ben's approval, 5c1f8e4c88, 3e736b006b)

Local runs were not reliable this round: the build lacked the CI e2e env, so model pages rendered as not runnable (no run-rollout-note with the flag off), and load average was about 37 against a 1 s loader grace. The pushed d64cddc740 failed the same 10 cases locally while CI passed it, so CI is the proof here.

# Check Proof Result
15 Explicit slugs, name-pinned h1, JS-off model cases restored, no-scroll expectIslandHydrated Website E2E run 36618294519 on 5c1f8e4c88: website-build and website-e2e-shard 1 to 4 all success. That covers this spec and the ~20 specs whose waitForIsland gained the count-1 check. Locally, the 3 JS-off model cases pass pass
16 Gates pnpm typecheck (website) 0 errors, pnpm exec eslint on the 2 e2e files exit 0, pre-commit hooks passed pass
17 Comment move (3e736b006b) Comment-only change to islands.ts, eslint exit 0; CI on it not watched pass

@github-actions

github-actions Bot commented Sep 29, 2026 •

Copy link
Copy Markdown
Contributor

🌐 Website E2E

Tip

All tests passed.

Status ✅ Passed
Report View Report

🧩 Custom Nodes Ecosystem Matrix: 🟡 Withheld — could not measure

Executes the frontend JS of every registry pack that ships any, against this change, in both renderers. The eight ecosystem-matrix shard checks are a harness self-test; the pack verdict is matrix-verdict.

This candidate changes nothing the matrix measures, so every expensive stage was skipped.

Reproduce locally
# build one shard's fixture, then run it
python3 scripts/registry-census/build_matrix.py --shard 1/4
MATRIX_RENDERER=vue MATRIX_VUE=1 \
  npx vitest run --config vitest.matrix.config.mts

Full docs: scripts/registry-census/README.md

📦 Bundle Size

⏳ Size data collection in progress…

🌾 Fallow: ✅ Passed

Fallow audits the diff for duplication, complexity and dead code. It gates on new findings only, so when the audit completes, a red is something this PR introduced or moved — never inherited backlog. A red can also mean the audit itself did not finish; the details below say which happened.

No new findings in the changed files.

Reproduce locally
# the base the gate compared against
npx fallow audit --root . --quiet --format json \
  --changed-since 426f1563577c736496bd42adfe02bfca134f2176

Full docs: ADR-DEVEX-LINT-0015

🔗 Website Preview

Website Preview: https://comfy-website-preview-pr-19394.vercel.app

Models is built and gated by PostHog (workshop-enabled), using staging Cloud.

This commit: https://website-frontend-qfxwtwbwa-comfyui.vercel.app

Last updated: 2026-09-29T20:20:34Z for 3e736b0

@github-actions github-actions Bot added the risk:high PR risk grade (advisory shadow check; grader-owned) label Sep 29, 2026
@codecov

codecov Bot commented Sep 29, 2026 •

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.

@@                        Coverage Diff                        @@
##           maanil/fe-2942-model-page-ssr   #19394      +/-   ##
=================================================================
- Coverage                          84.14%   84.05%   -0.10%     
=================================================================
  Files                               3188     3158      -30     
  Lines                             155105   154272     -833     
  Branches                           42475    39945    -2530     
=================================================================
- Hits                              130509   129666     -843     
- Misses                             22923    22990      +67     
+ Partials                            1673     1616      -57     
Flag Coverage Δ
website-unit 78.88% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.
see 305 files with indirect coverage changes

🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
  • 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

@github-actions github-actions Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔍 Cursor Review — Consolidated panel

Triggered by @MaanilVerma.

Found 10 finding(s).

Severity Count
🟠 High 1
🟡 Medium 5
🟢 Low 3
⚪ Nit 1

Panel: 8/8 reviewers contributed findings.

@benjcooley benjcooley left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Review

The shape of this is good — one table of pages, no sleeps, no new dependency, CLS from a real layout-shift observer, and frames > 0 to prove the watcher actually ran. One finding is load-bearing enough to block.

Severity Count
BLOCKER 1
SHOULD FIX 5
TRIVIAL 2

Blocker: the loader guard watches the wrong test id

recordLoaderAndLayoutShift polls [data-testid="workshop-loading"]. That id only exists in the pre-hydration frame. Once ModelsPage mounts, its async-component loading state renders h(WorkshopLoading, { label, 'data-testid': 'models-loading' }) (ModelsPage.vue:59) — the fallthrough attribute replaces the root's own workshop-loading, because one element cannot carry two values for one attribute.

So the spinner that actually sits on screen while catalogue.json / page.json load — the exact 2-to-5-second loader named in the PR description — is invisible to this guard. models-catalogue-corrections.spec.ts:60 and ModelsPage.loading.test.ts:46 both already assert models-loading in that window, so this is settled behaviour, not a guess.

Consequences:

  1. loaderLastSeenAt is 0 on every page under test, so expect(loaderLastSeenAt).toBeLessThan(1000) passes without observing anything. The headline guarantee of the PR is currently vacuous.
  2. The /models/ runtimePending annotation blames the loader, but those cases fail on the CLS budget. The recorded reason will mislead whoever removes the test.fail.

Matching both ids fixes both, and does not put the three model-page cases at risk — router-backed slugs render ModelPage directly and never mount a WorkshopLoading at all.

On the review-focus questions

  • 1 s loader grace / 0.1 CLS budget — the budget is fine; the grace window is currently unexercised for the reason above. Worth re-checking the margin once the selector is fixed, since a cold-cache client:load hydration under CI parallelism can push the pre-mount frame past 1 s.
  • Settle signal for model pages — island hydrated plus run-button:not([data-gate="resolving"]) / run-rollout-note is correct and matches what ModelDetail.vue actually emits for signedOut and rollingOut.
  • Hub test.fail() reasons — the JS-off one is right as far as it goes, though /models/ with JS off also renders ModelsShowcase from <noscript>, so the Grok assertion fails too, not only the h1. The runtime reasons are wrong (above).

Overlap worth a look

The h1-text, no-Grok and run-rollout-note assertions here are already covered by workshop-visibility.spec.ts and workshop-loading-flash.spec.ts. The genuinely new coverage is CLS and loader timing. Trimming the duplicated assertions would keep this closer to the repo's parsimony guideline and cut the 20-test footprint against a 20-minute CI globalTimeout.

Note on the CLS numbers

blockExternalMedia fulfils all media with a 1x1 webp / tiny webm placeholder. Any shift that production pays for because a real image arrives late at its real aspect ratio will not be charged here. The budget is still a useful regression tripwire, just don't read it as a production CLS figure.

The render guard now matches the post-mount models-loading loader, reuses
waitForIsland and the shared VIEWPORTS, types its window hook locally, and
picks its model pages from the published catalogue. The /models/ runtime
cases pass once the right loader is watched, so their test.fail is gone.
Assertions already covered by the visibility and loading-flash specs are
trimmed. The cdn-cgi/trace and PostHog flag stub moves to one fixture used
by all four callers.

@benjcooley benjcooley left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-review of d64cddc

Blocker resolved, and resolved properly rather than papered over. Approving.

Severity Count
BLOCKER 0
SHOULD FIX (non-blocking) 3
TRIVIAL 1

The blocker is genuinely fixed

The watcher now matches workshop-loading, models-loading. What moves this from "plausible" to "verified" is the evidence you brought rather than the one-line selector change: per-id timings showing models-loading is the last loader visible (393px: 373 ms then 419 ms; 1280px: 240 ms then 273 ms) prove the guard is now observing the post-mount spinner and not just the pre-hydration frame. Removing the /models/ runtime test.fail on the back of that — instead of leaving it in place with a corrected excuse — is the right call, and it converts four muted cases into four enforcing ones.

All four website-e2e-shard jobs are green on this head (681 tests, the one flaky is an unrelated screenshot). That also settles the grace-window question you flagged for the owner: your 774 ms worst case was measured on a laptop, and the budget held on a slower Actions container. I'd leave LOADER_GRACE_MS at 1 s. If it ever does flake, the useful response is to find out why the loader got slower, which is exactly what this guard is for — not to raise the number.

Fixture extraction

stubWorkshopFlags collapses four copies into one and the three call sites read better for it. One quiet behaviour change worth knowing about: workshop-visibility.spec.ts's old disableWorkshopFlag did not stub cdn-cgi/trace, and now it does, so that test's geo resolves to US deterministically instead of whatever astro preview returned. That's an improvement, but it wasn't in the commit message.

What's left (all non-blocking)

One theme runs through three of the four notes: in trimming assertions and making page selection dynamic, the suite lost its grip on page identity. resolvePath picks a URL at runtime, the h1 checks are gone from the runtime path, and the JS-off block is now a single expected-failure. What survives is a strong assertion that a model page hydrated without a loader or a jump — which is most of the value — but nothing ties a run to the page it claims to be testing. One toHaveText(name) sourced from the catalogue entry you already fetch puts that back for about three lines.

The waitForIsland scroll is on me — I recommended reuse without checking that it scrolls, and scrolling before reading CLS widens what the budget is measuring. Worth a follow-up, not a revert.

None of these block. Good, evidence-driven turnaround.

@benjcooley benjcooley left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-review of 5c1f8e4

All four notes from the last round addressed, each one at the root rather than at the symptom. Approving.

Severity Count
BLOCKER 0
SHOULD FIX 0
TRIVIAL 1

What landed

  • Page identity restored. publishedModel returns the catalogue name alongside href, and the runtime cases pin it with toHaveText(name) after settle. The guard can no longer pass on the wrong page.
  • Scroll removed. expectIslandHydrated split out in islands.ts, with waitForIsland delegating to it and keeping the scroll for the specs that click. This spec stops moving the viewport it is measuring, and the two callers still share one definition of what "hydrated" means. Clean resolution of a problem I introduced.
  • Subject pinned. Back to explicit MODEL_SLUGS, with publishedModel throwing The catalogue no longer publishes <slug> when one is delisted — presence asserted without identity being derived. The whole readFileSync / workshop-display.json / example-values block is gone with it, which is a good subtraction: the spec no longer reads app content off disk at collection time.
  • JS-off coverage restored, and better than the original — the expected h1 now comes from the catalogue instead of a hardcoded string, so the assertion and the route share a source.
  • Test titles now name the slug via pageTitle, so a CI failure identifies the page without digging into annotations.

On the tightened schema

I checked catalogueSchema requiring name on every entry, since catalogue.json is [...workshopModels, ...workflowModels, ...appModels] and a single entry missing the field would take out all fifteen tests with a zod error. It's safe, and safe for a good reason: WorkshopPresentation in models-catalogue.ts declares slug, name and href as required, and all three variants extend it, with appModelsFor skipping entries lacking displayName and workflowPageFor throwing on them. The schema is pinned to a compile-time invariant rather than to incidental data, which is the right thing to assert against.

CI

typecheck, typecheck-aux and lint are green on this head. The website e2e run for 5c1f8e4c88 is still queued, so this approval rests on the previous head's full green (all four shards, 681 tests) plus the fact that this commit's behavioural delta is small and checkable by reading: removing scrollIntoViewIfNeeded is safe because both routes hydrate with client:load, so hydration never depended on visibility, and toHaveCount(1) already provides the existence wait the scroll was incidentally supplying. I'll flag it if the run comes back red.

One leftover, not worth another round

The catalogue's own h1 (ComfyUI models) is asserted only in the test.fail JS-off case, so no passing test checks it. Defensible — /models/ is a literal path, so there's no wrong-page risk to guard against, and h1 text isn't what this spec is for. Mentioning it so it's a known gap rather than an assumed one.

Four rounds, each one converging. Nice work.

Comment thread apps/website/e2e/fixtures/islands.ts Resolved

@benjcooley benjcooley left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-review of 3e736b0

Comment-only, and it lands the split correctly: the Astro ssr-attribute explanation stays with expectIslandHydrated (true of both functions), the click rationale moves to waitForIsland (true of only that one), and the reworded opener is self-contained now that it no longer follows the sentence it used to lean on. Verified the diff touches nothing but // lines. Approving.

Severity Count
BLOCKER 0
SHOULD FIX 0
TRIVIAL 0

Correcting my last review on CI

I owe you an amendment. Last round I approved with the caveat that the e2e run for 5c1f8e4c88 was "still queued" and said I'd flag it if it came back red. That was wrong, and wrong in a way worth naming: I read the run-level status from gh run list, which reports queued while the trailing website-e2e-report job waits on a runner, and concluded the tests hadn't executed. At job level all four shards had already finished:

website-e2e-shard (1, 4) :: completed/success   172 passed (1.7m)
website-e2e-shard (2, 4) :: completed/success   171 passed (1.6m)
website-e2e-shard (3, 4) :: completed/success   171 passed (2.5m)
website-e2e-shard (4, 4) :: completed/success   171 passed (2.6m)

685 passed, zero failed, zero flaky. So the three things I said needed runtime confirmation — removing scrollIntoViewIfNeeded, the toHaveText(name) identity pin, and the tightened catalogueSchema requiring name on every catalogue entry — are all CI-verified on real infrastructure, not just argued from the types.

The test count corroborates that the restored cases actually ran rather than being filtered out: d64cddc740 totalled 682 (681 passed + 1 flaky), 5c1f8e4c88 totalled 685. Exactly +3, which is precisely the three JS-off model-page tests that came back in that commit. Nothing silently skipped.

Same misreading applied to d64cddc740, which I reported as run-level cancelled — its website-e2e gate job is completed/success; only the cosmetic post-result-comment step was cancelled by the next push.

The run for this head is pending, but since the diff is three comment lines there is nothing for it to invalidate.

Where this ended up

Five rounds from a guard whose central assertion could not fail to one that is pinned to named pages, watches both loader ids, verifies page identity from the published catalogue, measures without perturbing the viewport it measures, and is green on CI. The expectIslandHydrated / waitForIsland split is a small piece of real cleanup that outlives this PR — the next spec that needs hydration without a scroll now has somewhere to reach for.

No remaining comments. Good to merge once #18898 and the base land.

pull Bot pushed a commit to Penguinjanator/ComfyUI_frontend that referenced this pull request Sep 30, 2026
## Summary

**Deploy gate: comfy-router#46 is merged and live (2026-09-30). Its
follow-up comfy-router#48 (serves `/hub/models.md` instead of 301ing it
to a /workflows 404) must be deployed before this merges. After the
merge, purge Cloudflare for `/hub/models*`, `/hub/workflows*` and the 30
workflow names in `hub-workflow-names.json`. The redirects ship as 307;
FE-3083 flips them to 308 later.**

This branch combines every Workshop /models SEO PR, so merging it to
main replaces merging them one by one. It also launches: Comfy-Org#19405 commits
`launchedModelPages = 'all'`, so all 148 model pages become indexable in
the same deploy as the URL move.

Part of FE-2912, FE-2942 (PR 2 is FE-3086), FE-2974 (/minimax 308 is
FE-3087)
Fixes FE-2978, FE-2989, FE-2991, FE-2990, FE-2959, FE-2981, FE-2977,
FE-2979, FE-3014, FE-2936, FE-2980, FE-2975, FE-3085, FE-2972, FE-2935,
FE-2950, FE-2940, FE-2930, FE-2985, FE-2957, FE-2961, FE-2986, FE-2921,
FE-2922, FE-2947

## Changes

Router follow-up: comfy-router#48 (`/hub/models.md` routing, 324/324
router tests), deploy before this merges.


The included PRs, in merge order. Every head was checked as contained at
the final tip.

| PR | Ticket | What | Ben |
|---|---|---|---|
| Comfy-Org#18887 | FE-2978 | Pricing Product markup and full video `uploadDate`
| approved |
| Comfy-Org#18889 | FE-2989 | A unique meta description on every model page |
approved |
| Comfy-Org#18890 | FE-2991 | Model page JSON-LD, emitted once noindex lifts |
approved |
| Comfy-Org#18892 | FE-2990 | Cleaner tag pills, guard against duplicate
summaries | approved |
| Comfy-Org#18893 | FE-2959, 2981, 2977, 2979 | Social tags, hub title,
Organization data. The FE-2977 title was confirmed by the owner |
approved |
| Comfy-Org#19362 | FE-3014 | Every non-production build is noindex | approved |
| Comfy-Org#19366 | FE-2936 | Model titles lead with "API & Playground" |
approved |
| Comfy-Org#19402 | FE-2980 | Every internal link ends with a slash, plus a CI
check | approved |
| Comfy-Org#19419 | FE-2975 | Footer links the catalogue on every page and locale
| approved |
| Comfy-Org#19476 | FE-3085 | `NOINDEX_ROUTES` decides noindex for
serverless-animation | approved; re-approval pending on a one-line
follow-up (CodeRabbit approved) |
| Comfy-Org#19359 | FE-2974 | One typed redirect list that generates
`vercel.json` | approved |
| Comfy-Org#19369 | FE-2972 A | The old to new model URL map | approved |
| Comfy-Org#18891 | FE-2935 | Model content shows whatever the flag says |
approved |
| Comfy-Org#18899 | FE-2942 | Model pages render in the static HTML | approved |
| Comfy-Org#18898 | FE-2950 | Catalogue h1 and A to Z links in the static HTML |
approved |
| Comfy-Org#19375 | FE-2940 | Related models leave no page unlinked | approved |
| Comfy-Org#19385 | FE-2930 | Examples show the prompt that made them | approved
|
| Comfy-Org#19394 | FE-2985 | Render guard e2e: no loader, no layout shift |
approved |
| Comfy-Org#19357 | FE-2957 | robots.txt stops blocking our scripts and styles |
approved |
| Comfy-Org#19360 | FE-2961 | Turning Workshop off keeps every Models page |
approved (a CodeRabbit changes-requested review is still open) |
| Comfy-Org#19386 | FE-2986 | One registry of every Models address | approved |
| Comfy-Org#19421 | FE-2972 B1 | Model pages served at `/hub/models` | approved |
| Comfy-Org#19422 | FE-2972 B2 | Every old model URL redirects (307) | approved |
| Comfy-Org#19424 | FE-2972 B3 | Workflow pages at `/hub/workflows` with
redirects | approved |
| Comfy-Org#19399 | FE-2921 | One switch decides model robots and sitemap listing
| approved |
| Comfy-Org#19404 | FE-2922 | Clean markdown twins for model pages | approved |
| Comfy-Org#19405 | FE-2947 | Lifts noindex on every canonical model page |
approved |
| comfy-router#46 | FE-3075 | Serves `/hub/models*` from the website and
routes `/hub/workflows` by manifest. Merged and live | merged |

These fixes exist only on this branch, because they only break once two
PRs meet:

- `scripts/check-models-html.ts` reads each page at `model.href` and
skips models without a page (f1df1e4).
- `e2e/fixtures/modelsCatalogue.ts` reads `dist/hub/models`
(f87bb4c).
- `routerIdsByModelPage` accepts an optional `href`, since Comfy-Org#19421 made
it optional and Comfy-Org#19405 no longer typechecked (9783819).
- `ModelStatus` links a successor only when it has a page, for the same
reason on Comfy-Org#18899 (c4761d8).
- The model breadcrumb in `modelJsonLd.ts` points at `/hub/models/`
instead of the `/models/` redirect (eb985b6). Its label still says
"Models" while the nav says "Hub".
- Twin, meta-description and entry tests use the hub pages and the new
tag set (86ac3b0, d4ffc0f).
- The render guard finds the catalogue through `getRoutes().workshop`,
and its `test.fail` is gone now that Comfy-Org#18898 renders the h1 (831499f).
- A new test renders a real model page and checks its markdown twin has
the model's h1 (9a6c5cb).
- Main's newer catalogue and workflow e2e cases use the hub addresses
(2aeda77).
- `resetWorkshopAccountSource` resets `decision`; a clean auto-merge of
main's Comfy-Org#19454 into Comfy-Org#18891 left it assigning a variable that no longer
exists (c28c7a7).

The conflict resolutions, briefly: `ALTERNATE_TWIN_SOURCES` stays
deleted as Comfy-Org#18898 wants, so `/hub/models.md` comes from the real
catalogue. The hub keeps Comfy-Org#18898's single static h1 instead of the
`WorkshopHero` h1 that main's Comfy-Org#19320 added. `ExamplesTab` keeps Comfy-Org#19385's
figure and caption with main's video poster. The Comfy-Org#19399/Comfy-Org#19405 indexing
tests were rewritten for hub paths. The Comfy-Org#19359 redirect test no longer
builds `/enterprise//` from slash-ended routes. The Comfy-Org#19369 URL-map test
imports `PROVIDER_NAMES` from its new home. The full table of 47
resolutions is in the integration report.

## Review Focus

- Launch rides along with the URL move. If launch should stay a separate
step, set `launchedModelPages` in
`apps/website/src/config/model-page-launch.ts` to `new Set()` before
merging.
- Main's Comfy-Org#19320 gives the catalogue island its own h1 through
`WorkshopHero`. This branch keeps Comfy-Org#18898's static h1 and renders only
the subtitle in `ModelsCatalogue.vue`, so the hub has exactly one h1.

- Comfy-Org#19476 overlaps the open community PR Comfy-Org#18591 (same sitemap exclusion
for serverless-animation). One of them should be closed once this lands.

## QA

- [x] Gates: `pnpm typecheck` and website typecheck clean, `pnpm lint` 0
errors, eslint on `apps/website` clean, `pnpm knip` clean
- [x] Website unit: 547 files, 10347 passed, 4 skipped
- [x] Build A (`build:e2e`, CI env): 1193 pages; release gate,
`check:hreflang`, `validate:llms-txt-links` (225 links),
`validate:jsonld` (947 pages), `check:link-slashes`, `check:models-html`
(148 pages) all pass
- [x] Build B (`WORKSHOP_IN_BUILD=0`): same 182 Models pages as the on
build, no `dist/workshop`, `/` is noindex; with the flag stubbed on, no
Run, no Hub nav and no Firebase requests
- [x] Build C (preview shape): 945 of 945 pages noindex, no canonicals
except redirect stubs, no hreflang in the sitemap
- [x] E2E on Build A: 644 of 645 passed; the one mobile failure passed
on rerun. Earlier load flakes (navigation, demos, fdct, wan-3.0, network
isolation, render guard) all passed when rerun alone
- [x] Launch: 148 model pages indexable with canonicals, 149 hub entries
in the sitemap, 148 non-empty twins with the model h1, 148 in
`llms-full.txt`, 0 titles over 60, 30 workflow pages still noindex
- [x] Redirects: 853 rules, 782 to /hub all 307, the MiniMax rows 307,
every /hub destination built
- [x] comfy-router#46 under `wrangler dev` in front of this build: 312
router tests pass; 148 old model URLs (both slash forms) 307 then 200;
30 workflow names served by the website and their old URLs 307; a
non-manifest workflow 302 with no-store; `manifest.json` passes through;
`/workflows/` and bare `/hub` unchanged
- [x] After the router deploys and this merges: purge Cloudflare, then
curl `https://comfy.org/hub/models/` for a 200 and three old
`/models/<slug>/` URLs for one 307 each
- [x] Rich Results Test on `/pricing/` and
`/learning/animations/character-sheet/`, then "Validate fix" on Merchant
listings and Videos
- [x] Rich Results Test on two live model pages: breadcrumbs valid and
the breadcrumb URL returns 200 with no redirect
- [x] GPTBot and Googlebot curls on the hub and one model page, then a
GSC live test on both
- [x] Resubmit the sitemap in Search Console and URL-inspect about 20
model pages on day 0 and day 14
- Not covered: Rich Results and GSC need production; the redirect matrix
used a local shim for `vercel.json`, since `astro preview` ignores it

---------

Co-authored-by: GitHub Action <action@github.com>
Co-authored-by: benjcooley <benjcooley@gmail.com>
Co-authored-by: Deep Mehta <42841935+deepme987@users.noreply.github.com>
@MaanilVerma

Copy link
Copy Markdown
Collaborator Author

Superseded by the integration branch #19494, which combined this PR with every other Workshop SEO slice and has been merged to main. Closing this one.

Penguinjanator pushed a commit to Penguinjanator/ComfyUI_frontend that referenced this pull request Oct 15, 2026
…ion (Comfy-Org#21047)

<!-- ccr-slack-attribution -->
_Requested by **Alexander Brown** · [Slack
thread](https://comfy-organization.slack.com/archives/C0B61J7QCJW/p1792015617100449)_

## Summary

`models-render-guard.spec.ts` timed the `/hub/models/` loader from
navigation start, so its 1 s grace also had to cover booting the site's
JavaScript on the runner. That boot alone takes most of the budget on
CI, so the catalogue cases fail on timing, not on the loader lingering.
This PR counts the grace from the moment the page's island hydrates.
That is when the component takes over the screen, and when the
post-mount `models-loading` loader the guard was written for starts.

**Before:** `loaderLastSeenAt` was `performance.now()` at the last frame
a loader was visible, compared with 1000 ms. The catalogue renders
`WorkshopLoading` on the server (the island is not mounted yet), so that
number covers the whole HTML, JS, hydration and `catalogue.json` chain.
When Comfy-Org#19394 added the guard, the worst local run was 774 ms, and its
review focus already flagged the 1 s budget as the risk. On CI the
catalogue cases now land at 1.07 to 1.22 s:

- main push run 38649631281 (shard 3, job 116145534816): 393px flag on
was flaky (passed on retry).
- merge-queue run 38650370738 (shard 3, attempt 1): 393px and 1280px
flag on failed all three tries, 1280px flag off was flaky.
- merge-queue run 38652042112 (shard 3, job 116153522370): 393px flag on
failed at 1139, 1224 and 1068 ms, and 1280px flag on was flaky.

None of these come from a change to the catalogue: the same tree passed
on rerun. In a local trace about 660 to 940 ms of the 1.0 to 1.3 s went
to booting and hydrating the page. `ModelsCatalogue` and
`/models/catalogue.json` are only requested once the island mounts.

**After:** the init script records when each `<astro-island>` hydrates.
Astro dispatches a non-bubbling `astro:hydrate` on it, so the script
listens in the capture phase. `waitUntilSettled` reads that time for the
island holding the page content and waits a full grace window after it.
The assertion becomes "ms after the page island hydrated that the loader
was last visible", still under 1000 ms. The CLS budget, the pages, the
flag matrix and the readiness waits are unchanged. A loader the server
renders into a model page is still caught by the existing JavaScript-off
cases, which assert `workshop-loading` is hidden.

## Verification

- Local run against a `build:e2e` of current `main` with a local
Chromium, launched through `executablePath` from a scratch config that
is not committed.
- Before, `--repeat-each 5 --workers 2` on the catalogue cases: 16 of 20
failed, at 1044 to 1323 ms.
- After, the whole spec with `--repeat-each 10 --workers 3`: 200 of 200
passed. The loader was last visible at most 456 ms after hydration.
Under the old metric, 14 of those 40 catalogue runs would have failed,
peaking at 1466 ms.
- The guard still catches a loader that lingers. Delaying
`/models/catalogue.json` by 1.5 s with `context.route` failed all four
catalogue cases, with the loader last visible about 1.5 s after
hydration.
- The pre-commit hooks (oxfmt, oxlint, eslint, server-facts, website
typecheck) and the pre-push hooks passed.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_018GfgJuhjWFFm4LbKojTDxV

Co-authored-by: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

cursor-review Trigger multi-model Cursor agent review on this PR risk:high PR risk grade (advisory shadow check; grader-owned)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants