Repository navigation
feat(website): Made with Drama 3 sample rows and RUN DRAMA 3 CTAs on /fish-audio - #20970
Conversation
… prompts Restores the row list under the hero with the five Drama 3 clips Purz generated: one row per clip with a play button, clip length, the exact prompt and voice that produced it, a copy-prompt button, and a TRY NOW link to the matching Cloud template (directed for single-voice scenes, dialogue for the two-voice one). The clips are served from media.comfy.org/website/fish-audio/ as hushed-gossip, movie-trailer, sleepy-to-hyped, two-voices-one-take and noir-detective .mp3 and still need uploading to the bucket.
Each sample row now reads "Title (8s) · voice: Name" with only the title in warm white, and the one-line summary under it is gone. The length comes from the clip data so the line renders before the files are on the CDN, which also retires the runtime duration readout.
…h-audio The hero gets its primary CTA back, now pointing at the Cloud directed template, beside VIEW DOCS. Each sample row's prompt clamps to three lines with a Read more / Read less toggle that only appears when the prompt actually overflows.
…e with Drama 3 heading The sample rows use the standard yellow play button, so the muted variant goes away again. The steps section gets the same RUN DRAMA 3 primary button as the hero, and the sample list is headed Made with Drama 3.
|
Important Review skippedReview was skipped due to path filters ⛔ Files ignored due to path filters: 10
CodeRabbit blocks several paths by default. You can override this behavior by explicitly including those paths in the path filters. For example, including ⚙️ Run configuration
You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
🧩 Custom Nodes Ecosystem Matrix: 🟡 Withheld — could not measureExecutes the frontend JS of every registry pack that ships any, against this change, in both renderers. The eight 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 🌾 Fallow: ✅ PassedFallow 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 3557c088356200f3d70e4b2500cba416580b2926
Full docs: ADR-DEVEX-LINT-0015 🌐 Website E2ETip All tests passed.
🔗 Website PreviewWebsite Preview: https://comfy-website-preview-pr-20970.vercel.app Models is built and gated by PostHog (workshop-enabled), using staging Cloud. This commit: https://website-frontend-gj5uk7mpe-comfyui.vercel.app Last updated: 2026-10-14T02:18:00Z for |
Codecov Report❌ Patch coverage is
@@ Coverage Diff @@
## main #20970 +/- ##
==========================================
- Coverage 86.38% 86.27% -0.11%
==========================================
Files 3672 3569 -103
Lines 175241 172051 -3190
Branches 46278 45184 -1094
==========================================
- Hits 151381 148437 -2944
+ Misses 21818 21728 -90
+ Partials 2042 1886 -156
Flags with carried forward coverage won't be shown. Click here to find out more.
... and 106 files with indirect coverage changes 🚀 New features to boost your workflow:
|
Swarmhost agentic reviewThe detailed evaluation is available to employees in the internal Slack review thread. Evaluation failed to run ( Evaluation budget remaining for this pull request: 99 automatic and 99 manual. Updated by Swarmhost's agentic review process. |
🐑 Good news first: the five clips are live. Your description says they aren't.@june-bump the body reads "still need uploading to the bucket; until then the play buttons do nothing." Checked just now — all five are up:
So the caveat is stale and the play buttons will work. Worth editing the description — otherwise a reviewer reads "ships broken" and weighs it against you. Nothing in CI checks these URLs, so this is a one-off measurement rather than a guarantee, but it's the failure mode that would have mattered most. Blast radius is nil, which I verified rather than assumed. Only Where I actually am on thisScope is borderline and I'd rather say so than pretend it's clear. It's +494/−7 with 184 lines of new Vue — Against that: it's additive, one page uses it, rollback is one commit, and 143 of the added lines are tests. So I'm not refusing at intake, and I'm not promising to queue either. @webreviewer-bot is labelled in, because the real risks here are behavioural — clipboard permissions, audio policy, whether the Read-more toggle only appears when the prompt actually overflows — and those want measuring rather than reading. I'll make the call once its pass lands, which is how the procedure is meant to run: doing the work doesn't commit me to pressing the button. If it comes back clean and the scope still reads as a feature addition rather than a fix, I'll hand it to a human with the measurements attached rather than queue it myself. |
🐑 The CTAs are genuinely fixed — but I can't prove the templates exist, and I ran a control to show why.@june-bump the reason "RUN DRAMA 3" and "TRY NOW" were stripped on #20898 was that they pointed at Cloud home with no Drama 3 template. That's not what this does. Specific templates, with the dialogue one correctly reserved for the two-voice clip. That's the defect fixed rather than reintroduced. What I can't establish, and nobody should take my word that I did: whether those template IDs resolve to anything. I tried, and the measurement is worthless — because a deliberately fake ID behaves identically:
Cloud is an SPA; it answers 200 for anything. So "the links work" is unprovable from outside, and a button that lands on an empty Cloud session is the same user-facing failure as the one that got these removed. Someone with Cloud access needs to open both and confirm. One minute of a human's time, and it's the only thing that settles it. A separate, cheap one while you're in this file
That is not a regression from this PR and I'm not grading it as one. You're just well placed to close it in a one-line edit to the object you're already editing — the poster is uploaded and the precedent is I'll note I got this wrong earlier: when the hero shrank to 5.1 MB I told you dropping the guard was "correct rather than a regression". @webreviewer-bot measured the wire and phones still fetch ~2 MB in 7s. I reasoned from file size instead of measuring. Still holding the queue for @webreviewer-bot's pass, which is resolving every href and reporting final URLs after redirects — the thing I can only partially do from here. |
🐑 Ran the CTA check properly with a control. Signed out, a real template and a fake one are indistinguishable — so this genuinely needs someone with Cloud access.@webreviewer-bot suggested going past status codes to the rendered result with the bogus id as a control cell. Done, in a browser:
Identical behaviour. Cloud redirects to login before it ever fetches a template, so no request that could distinguish a real id from an invented one is made. I watched the network: So "do these templates exist?" is unanswerable from outside, and now that's measured rather than assumed. My earlier status-code table was worthless for the same reason; this at least establishes why it was worthless. Two things that are now settledThe CTAs are not the #20898 defect returning. They carry real template parameters, not bare Cloud home, and the template survives the login round-trip — it's preserved in Signed out, "RUN DRAMA 3" lands on the Cloud login page. Not broken, and consistent with what What's left for a humanSomeone with a Cloud account opening both links after signing in, and confirming they load the Drama 3 templates rather than an empty session:
A minute of work, and the only remaining way to know. @june-bump you're likely best placed — and if they do resolve, that closes the last thing I'd want checked before this is merge-ready on my side. Still holding the queue for @webreviewer-bot's pass. |
🐑 The registry doesn't know these templates. That turns "worth a click" into "please check before this ships".@june-bump following @webreviewer-bot's lead, I checked the nearest in-repo signal and verified it myself:
Three Fish Audio templates registered, and no This is evidence, not proof — So the human check moves from "nice to confirm" to "please do it before merge": sign in to Cloud, open both links, confirm they load Drama 3 rather than an empty session. If they don't resolve, five "TRY NOW" buttons and two "RUN DRAMA 3" buttons all land somewhere unhelpful — which is precisely the defect that got these CTAs removed on #20898. One thing that is confirmed good: the mapping. Correcting myself on the hero guardI told you
So this is the majority behaviour, not a fish-audio oversight. The one-line fix still works for your page, but #20932's suggestion of a test asserting every hero video has a fallback would fail seven pages, which is a project rather than a cleanup. I'll correct the issue. Credit to @webreviewer-bot for catching that my framing was wrong. Still holding the queue — its pass is still running, and it's flagged a likely |
webreviewer-bot
left a comment
There was a problem hiding this comment.
REQUEST_CHANGES — reviewed at dd2d1adf0b3d4cb7053a26deb05e9ede30308027
This PR merged at 02:47:06Z, while my pass was still running. My verdict is therefore advisory and gates nothing — I am not asking for a revert, and I won't re-post this. But one finding is a deterministic keyboard defect that is now live on /fish-audio, so the record belongs here and the fix belongs in a follow-up. Everything below was measured at the SHA above, in my own temp clone; I never read or built in a shared checkout.
First, the good news, because it is the main thing this PR had to prove: the list section genuinely renders this time. On #20898 these three files were deleted precisely because nothing rendered them — ModelLaunchPage.astro had a case 'list' branch and fishAudio.ts declared zero list sections. That is fixed. ModelLaunchPage.astro:119-124 reaches case 'list' under page.list &&, fishAudio.ts:50-51 declares both sectionOrder: ['list', …] and a populated list, and I confirmed it in the served HTML of a build:e2e build, not in a dev server: curl of /fish-audio returns all five titles, all five .mp3 filenames, <audio ×5, preload="metadata" ×5, line-clamp-3 ×5, TRY NOW ×10, Copy prompt ×10. Two independent build-and-serve passes agreed.
| Severity | Count |
|---|---|
| BLOCKER | 1 |
| SHOULD FIX | 5 |
| TRIVIAL | 3 |
BLOCKER
1. ModelLaunchListRow.vue:116 — "Read less" destroys its own button and drops keyboard focus to <body>.
measurePromptOverflow (:40-42) compares scrollHeight > clientHeight. While expanded is true the line-clamp-3 class is off (:102), so the paragraph is unclamped, the two are equal, and the ResizeObserver (:45) writes promptOverflows = false. Clicking "Read less" then sets expanded = false, so v-if="expanded || promptOverflows" evaluates false || false and Vue destroys the focused button; once the re-clamped paragraph measures overflow again, Vue mounts a different node.
Measured, keyboard-only, two independent harnesses over a real build:e2e build, 2/2 viewports (390x844 row 3; 1440x900 rows 4-5):
- expand (Enter on "Read more") →
activeElement=BUTTON "Read less",=== document.bodyfalse immediately and at +400 ms; MutationObserver logs no mutations;sameNodeAfterExpand: true - collapse (Enter on "Read less") →
activeElement=BODY,=== document.bodytrue immediately, at +400 ms, and at +1500 ms; MutationObserver logs{removed: BUTTON "Read less"}then{added: BUTTON "Read more"}4-6 ms later;sameNodeAsBefore: false
The asymmetry — expand reuses the node and keeps focus, collapse replaces it and loses focus — isolates the cause to the v-if, not to the click handler. Recovery costs one extra Tab (Chromium resumes sequential navigation from the removed node's position), so this is a lost focus ring plus a lost place, not permanent trapping. At 390x844 all five rows are affected; at 1440x900 rows 4 and 5.
I am not proposing a specific remedy — a suggested fix is a claim and I did not run one. The mechanism above is the whole of what I measured.
SHOULD FIX
2. ModelLaunchListRow.vue:118 and :130 — two controls are under the 24x24 minimum at 390x844. Measured getBoundingClientRect(), identical on all five rows and reproduced by both passes: play 48x48 (pass), copy prompt 44x44 (pass), "Read more" 81.45 x 22.75, "TRY NOW" 76.19 x 18. Both fall below WCAG 2.2 SC 2.5.8's 24x24 CSS-px minimum, and neither is inline in a sentence, so the inline exception does not apply — each is the sole content of its own block. At 1440x900 "TRY NOW" measures 76.19 x 24 (at the line) and "Read more" stays 22.75.
3. ModelLaunchListRow.vue:75, :111, :132/:144 — three families of five controls each share one accessible name. Dumped from the a11y tree: the five play buttons are all exactly "Play", the five copy buttons all "Copy prompt", the five visible CTAs all "TRY NOW", and the "Read more" buttons all "Read more". Only the title links are distinguished ("Hushed gossip", "Movie trailer", …). The <audio> elements do carry per-sample aria-labels, but they compute display: none with no controls, so they expose nothing. A screen-reader user navigating by button or link rotor — a standard mode, not an edge case — gets five identical "Play" and five identical "TRY NOW" with no way to tell which sample. This is the largest a11y surface the page has gained and the row already knows its own title.
4. fishAudio.ts — mobileFallbackImageSrc is still absent, so #20932 stays live, and it is systemic rather than one page's oversight. Not a regression from this PR and not a reason I'd have withheld approval; raising it because this PR rewrites the exact object the fix belongs in. I verified the mechanism at this ref rather than trusting the issue: ModelLaunchHeroSection.vue:35-40 computes hasMobileMedia = Boolean(hero.mobileVideoSrc || hero.mobileFallbackImageSrc) and showVideo = !hasMobileMedia || (isMounted && isDesktopViewport) — with neither set, showVideo is unconditionally true, so the <video> is server-rendered to phones. Measured at 390x844 with a fresh browser process: fish-audio-hero2.mp4 is fetched, 352,780 B / 478,996 B / 539,040 B / 2,507,833 B / 3,439,400 B / 4,275,096 B across six runs of a 5,375,238 B asset — progressive range streaming, so the figure is time- and bandwidth-dependent and effectively unbounded.
Two corrections to the framing in #20932, both measured: fish-audio is not an outlier. Of the 13 launch pages whose hero sets videoSrc, 7 set no mobile fallback at all — fishAudio, flux3, geminiOmni, ltx, minimaxMusic3, seedance, wanAnimate2 — and 6 do. And the test that looks like it should cover this cannot: modelLaunchPages.test.ts expectHeroStills() guards every assertion with !== undefined, so absence passes. So the issue's suggestion of a test pinning "every hero video has a fallback" would fail 7 pages today, not pin one — a larger change than one line, and where it lands is not my call.
5. ModelLaunchListRow.vue:64 — five eager <audio> elements pull 275 KB-1.50 MB of audio on load at 390x844 with no interaction, and preload is asserted by nothing. The <audio> tags are in the SSR HTML, so they fire at parse time rather than on hydration. All five .mp3s are requested (206) on initial load without scrolling; measured transferred bytes across six fresh-process runs ranged 274,885 B to 1,502,681 B, with audio.buffered.end(0) at 14.25-17.89 s and readyState: 4 on all five — seconds of decodable audio, not a metadata header.
Attribution, because it matters for where the fix goes: I ran a control — a bare page with the same five <audio preload="metadata"> tags and no Vue, same browser — and it over-buffers too (96,593 / 93,730 / 68,065 / 65,536 / 90,296 B). So the over-fetch is Chromium's loose interpretation of preload="metadata", not component logic. The component's contribution is choosing five eager per-row <audio> elements at all. Total wire at 390x844 came to 5.18-8.08 MB across runs (vs 5.03-7.06 MB at 1440x900 — mobile is not cheaper). Mutation M6 confirms nothing would notice if preload were dropped entirely.
6. ModelLaunchListSection.test.ts:119 — the new suite structurally cannot observe finding 1, and asserts a class the repo's own doctrine rules out. mockPromptHeight(120, 60) pins scrollHeight/clientHeight for the whole test, so promptOverflows is true in every state — including while expanded, which is the one state where a real browser measures false and the only reason the button gets destroyed. The collapse the test exercises is therefore the only collapse where the button survives. Lines :119-126 then assert on className containing line-clamp-3; docs/guidance/testing-principles.md:45-47 at this ref says:
"Assert on an attribute or CSS class only when that class is the documented state contract… Do not use a class as a stand-in for the behavior it represents."
The rendered result — the paragraph's clamped vs. full height, and document.activeElement after collapse — is what would have caught it. Relatedly, mutation M7 shows the play button's accessible name is unasserted: with the aria-label removed the button has no accessible name at all and 127/127 tests still pass. And /fish-audio is the only model-launch page with no e2e spec, which this PR is what makes interactive.
TRIVIAL
7. types.ts:141 — @mrseek-reviews' P3 from #20898 recurs unaddressed. They noted that types.ts wrote out the same shape as ModelLaunchAudioCard.audioSources rather than referencing it, so a change to one source shape would not reach the other. At this head :141 still reads readonly { src: string; type: string }[], verbatim identical to :121. Structural typing means both work today; the coupling is the point. Credit to them for raising it first — grading it as they did.
8. ModelLaunchListRow.vue:58 — double separator in ja and zh-CN. metaLabel joins with ${voices} ${voiceNames}, but the ja and zh values of modelLaunch.voices already end in a full-width colon (ボイス: / 声音:) which carries its own trailing space. Rendered: (8 秒) · 声音: Friendly Women (en) — a visible double gap. The en value voice: uses a half-width colon and does need the space. Only zh-CN is a live route. One correction and then it's the author's call.
9. ModelLaunchListRow.vue:124 — the mobile CTA wrapper renders unconditionally. v-if="ctaLabel" sits on the inner <a> at :126, not on the <p>, so a consumer omitting the optional ModelLaunchList.ctaKey gets an empty paragraph carrying mt-4 under every prompt box below 768 px. fishAudio sets ctaKey, so this is latent, not live.
The CTAs: what is and is not knowable
The hrefs name specific templates rather than bare Cloud home, which is exactly the defect that got these buttons stripped on #20898. That is fixed, not reintroduced.
The mapping is correct — the one part of this anyone can settle from here, and a reversal would have been a real defect no URL check could see. fishAudio.ts:90 gives DIALOGUE_TEMPLATE_HREF to two-voices-one-take, the only item with two voices (Sarah (en) @Voice1, Adrian (en) @Voice2); the four single-voice clips get DIRECTED_TEMPLATE_HREF. Confirmed again in the live DOM: rows 1, 2, 3, 5 → …drama3_directed, row 4 → …drama3_dialogue, and all three links within a row share that row's href.
Template existence is not verifiable without Cloud credentials. I am deliberately not reporting status codes as evidence here: the template is a query parameter resolved client-side, so cloud.comfy.org answers 200 for any value and a status code proves only that Cloud is up. @webshepherd-bot ran the control that settles it — ?template=definitely_not_a_real_template_xyz redirects to cloud.comfy.org/cloud/login?previousFullPath=… identically to the real ids, preserving whichever id it was given without resolving it. Signed out, a real id and an invented one are indistinguishable. Credit to them for both that control and the registry check below. I did not authenticate.
And the in-repo signal turns out to be worth nothing — this is a premise I was asked to weigh and it does not hold. api_fishaudio_drama3_directed / _dialogue do appear nowhere in the repo but fishAudio.ts, and hubTemplates.json holds three Fish Audio entries (api_fishaudio_text_to_speech, api_fishaudio_voice_clone_tts, api_fishaudio_speech_to_text) and zero drama3. But hubTemplates.json is not the registry Cloud resolves ?template= against. It is a one-time committed snapshot of the hub browse listing — 610 entries including community uploads, with usage counts and username fields, consumed only by src/lib/hub/workflow-detail.ts:7-8 and components/hub/HubBrowse.vue:17 (neither mounted by any page at this ref) plus two generation scripts. One commit has ever touched it (1acdece6, 2026-09-11), its newest entry is dated 2026-09-01, and no script in the repo refreshes it.
The control that decides it: four already-shipped launch-page CTA ids are also missing from it — image_qwen_image_2_1_t2i, api_openai_gpt_image_25_flare_t2i, api_nano_banana_2_1_t2i, utility_nvidia_rtx_video_enhance. Absence from that file is a property shared by CTAs that presumably work, so it is not evidence these templates don't exist. The real registry is the external Comfy-Org/workflow_templates repo. Nothing in this repo validates a ?template= id against anything, so a typo in one of these two ids would be caught by no check we have — which is the actual, narrow thing worth a human's ten seconds.
i18n: parity is correct here, but ja is enforced by nothing
Key parity is clean. All six new leaf keys — fishAudio.hero.primaryCta, fishAudio.samples.heading, fishAudio.samples.cta, fishAudio.steps.primaryCta, modelLaunch.voices, modelLaunch.clipSeconds — exist in en, ja and zh-CN, as do all six pre-existing keys the new row depends on (modelLaunch.copyPrompt, ui.copied, ui.readMore, ui.readLess, player.play, player.pause). The single-form ja/zh voices values resolve correctly under t(key, {}, { plural: n }) rather than returning empty.
The mutation answer, which is the part worth more than the keys: nothing in this repository fails on a missing ja key. Deleted ja.fishAudio.samples.heading (guard git diff --quiet HEAD → exit 1), then ran every gate:
| gate | command | result |
|---|---|---|
| page config | vitest run src/templates/model-launch/modelLaunchPages.test.ts |
PASS, 120 collected |
| translations | vitest run src/i18n/translations.test.ts |
PASS, 13 collected |
| typecheck | astro check / vue-tsc-native --noEmit |
0 errors, exit 0 |
| locale check | tsx scripts/i18n/update-locales.ts --target website --check |
exit 0 (prints ja/main.json: 32 strings need translation, was 31) |
| lint | eslint on the changed files |
exit 0 |
Deleting the same key from zh-CN is caught: translations.test.ts fails 1 of 13 — main.json catalog > includes Chinese copy for every eligible English message at :129, expected [ 'fishAudio.samples.heading' ] to deeply equal []. Deleting it from en is caught by typecheck (ts(2820), not assignable to type 'TranslationKey'), not by lint.
So ja is not covered to the standard zh-CN is, and the mechanisms are all explicit rather than accidental: modelLaunchPages.test.ts:65 sets LOCALES = ['en', 'zh-CN'] and omits ja entirely; translations.test.ts:116-130 is hard-coded to the Chinese catalogue, while its sibling loop over all locales reads localized.get(key) ?? fallback at :143-144 and so renders English and passes; TranslationKey derives from typeof en alone, so no non-English catalogue is visible to tsc; update-locales.ts:403 returns non-zero only for protected-token violations, never for pending translations; and @intlify/vue-i18n/no-missing-keys has a localeDir listing only the en catalogue.
Context that de-escalates this rather than inflating it: ja/main.json is already missing 31 keys at this head with CI green, and there is no ja route for this page to begin with — locales.ts:57-59 sets PARTIAL_LOCALE_ROUTES = { ja: new Set(['/']) }, and src/pages/ja/ contains only index.astro. So this PR's ja strings are correct, complete, and currently unreachable. Nothing to fix in this diff; I am recording it because this is the first launch-page PR to touch ja and the next one may not be complete. Relatedly, LocalizedText is { en: string; 'zh-CN': string } & Partial<Record<Locale, string>>, so the five row title objects omitting ja is permitted by design — dropping 'zh-CN' instead is a hard ts(2322). A ja visitor would see English row titles, if a ja route existed.
Mutation results — do the new tests bite?
Own temp clone at dd2d1adf, node_modules/.vite cleared, every mutation guarded with git diff --quiet HEAD (exit 1 confirmed before each run), every revert re-guarded. Baseline: ModelLaunchListSection.test.ts 7 collected, modelLaunchPages.test.ts 120 collected, translations.test.ts 13 collected — no suite collected 0, and baseline re-verified green at the end. Two independent passes agreed on every cell.
| # | mutation | verdict | failing assertion |
|---|---|---|---|
| M1 | remove the v-for over list.items |
DIED (7/7) | links each row title and CTA… → no listitem role, :58; printed DOM is an empty <ul/> |
| M2 | all three :href="item.href" → a constant |
DIED | :64 expected Set{'https://example.com'} to deeply equal Set{'…?template=b'}. toHaveLength(3) at :63 still passed, so it pins href identity, not just count |
| M3 | delete the whole useEventListener(audioEl, 'play', …) block |
DIED | pauses every other row when one starts playing → :140 expected [] to include <audio>; pause never called |
| M3b | keep the block, drop only if (el !== audioEl.value) |
DIED | same test, the other assertion :141 — both halves bite |
| M4 | clamp class made unconditional | DIED | expands the prompt on read more… → :119 not.toContain('line-clamp-3'). Right reason, but see finding 6 — it reads a class, not geometry |
| M5a-d | meta separator; drop voiceNames; drop the @VoiceN alias; pin plural: 1 |
DIED (all four) | :88 Unable to find an element with the text: (44s) · voice: Narrator (en). The plural wiring is genuinely covered |
| M6 | remove preload="metadata" |
SURVIVED | vacuous — not unreachable (it renders), not redundantly guarded (no second mechanism sets preload); no test in apps/website asserts preload for this component |
| M7 | remove the play button's aria-label |
SURVIVED | vacuous — and verified not redundantly guarded: PlayPauseButton.vue has no text node and marks both SVGs aria-hidden, so under M7 the button's accessible name is empty, and 127/127 still pass |
Both survivors were diagnosed against all three explanations before being called gaps, not just the first.
Rendered checks — desktop 1440x900 and mobile 390x844
reducedMotion: 'reduce' on every context; build:e2e (cross-env WEBSITE_INDEXABLE=1 astro build, 1152 pages, exit 0); served bytes confirmed cmp-identical to dist/fish-audio/index.html in both passes, one of which avoided astro preview entirely in favour of a plain static server over dist.
- Clamp: collapsed
clientHeight68 px against a 22.75 px line-height = exactly 3 lines, on every clamped prompt, at both viewports. Verified by deletingline-clamp-3and re-measuring: at 1440x900 rows 1-3 hide 0 px, so their missing "Read more" is correct rather than a dropped control; rows 4-5 hide 46 and 69 px and do get one. At 390x844 all five overflow (114/160/319/455/569 px hidden) and all five get one. - Overflow: none anywhere.
documentElement.scrollWidth == clientWidth(390/390 and 1440/1440);ul, every row, and the prompt card all measure 0 overflow on both axes. - Expand reflow: clicking "Read more" on row 3 at 390x844 moved rows 1-3 by Δ0 and rows 4-5 by +318.5 px, matching row 3's growth exactly; document height +318,
scrollYunchanged, no overlaps. Rows above are never disturbed. - Idle-vs-idle control, taken before any pixel comparison: 0 diff pixels of 329,160 at 390x844 and 0 of 1,296,000 at 1440x900, two screenshots 3 s apart with zero interaction. Honest qualification: under
reducedMotion: 'reduce'the hero<video>is paused (paused: true,currentTime0.16 → 0.16 after 3 s), so this control establishes pixel stability with the video paused and does not certify a diff taken under active decode. I ran no before/after pixel diff, so no claim of mine rests on it. - CTA duplication: 10 CTA links in the DOM, 5 visible at both viewports — the
md:hidden/md:blockpair is correctly mutually exclusive, no row ever shows two. - Heading outline:
H1 "Fish Audio Drama 3 is now on Comfy"→H2 "Made with Drama 3"→ 5 ×H3. No level skip. - Audio: nothing autoplays (all five
paused: true,currentTime: 0after load). Exclusive playback works — starting row 1 gives[false,true,true,true,true]; then starting row 3 gives[true,true,false,true,true]with row 1 paused atcurrentTime 2.31rather than reset, and its button label correctly reverting to "Play". The<audio>elements computedisplay: none, carry nocontrolsand notabindex, and took focus 0 times in a 120-press Tab walk. - Keyboard: intra-row Tab order is Play → title link → Copy prompt → Read more → TRY NOW. Enter and Space both start playback on the play button (
paused: false,currentTime1.07 s and 1.51 s). Focus indicator isoutline-style: auto— so Chromium paints its own adaptive ring and the declaredoutline-color: rgb(16,16,16)is not the painted colour; I computed no contrast figure from it, which would have been a fake failure.
Cleared
- All five clip URLs resolve —
https://media.comfy.org/website/fish-audio/<name>.mp3, HTTP 200:hushed-gossip96,593 B,movie-trailer218,846,sleepy-to-hyped255,522,two-voices-one-take408,182,noir-detective523,538. The PR body is stale on this: it says the clips "still need uploading to the bucket; until then the play buttons do nothing". They are live and the buttons work. A factual correction, not a finding. durationSecondsis accurate against the real clips (8.05 / 18.23 / 21.29 / 34.01 / 43.62 s vs. the declared 8 / 18 / 21 / 34 / 44).- The re-added code is improved, not replayed — the row is substantially rewritten (prompt box, 3-line clamp, voices/duration meta line, copy-prompt replacing copy-link, the
ArrowRightno-audio variant gone). One exception worth noting:ModelLaunchListSection.vuereturns byte-identical to the version deleted ate49f955. - Clipboard feedback is sound:
CopyTextButton.vue:44-50swaps the icon to a check, shows thecopiedLabeltext, and announces viasr-only aria-live="polite", with a 2 s window. The accessible name does say what it copies ("Copy prompt"), just not which row's — folded into finding 3. A rejectedwriteTextis discarded byvoid copy(value)and would leave no feedback, but that is pre-existing shared-component behaviour this PR neither introduces nor newly exposes, so I'm not filing it against this diff. - Exclusive playback copies the global
document.querySelectorAll('audio')pause verbatim fromAudioPlayer.vue— house pattern, not a new leak. DEFAULT_SECTION_ORDERgaining'list'changes nothing elsewhere: every render site is guarded bypage.list &&, and a pre-existing test already fails if a defined section is left out of the order.- The new
modelLaunchPages.test.tsadditions are real coverage: list item ids join the page-wide uniqueness check,item.hrefjoinsoutboundHrefsURL-format validation,audioSourcesare pattern-checked againstAUDIO_URL, andexpectListItemsCompletepins non-empty prompt,durationSeconds > 0and non-emptyvoices. - CI at this head is genuinely green — 146 check-runs: 66
success, 84skipped, 1cancelled, zero failures. No red X, so no cancelled-run artifact to disentangle.
Not verified, and why
- Whether the two Cloud templates exist. Not observable without credentials, established by @webshepherd-bot's bogus-id control rather than assumed, and the in-repo signal is worthless for the reason given above. I did not authenticate. This is the one item a human can settle in seconds: click both CTAs signed in.
- Pixel stability under active video decode — my idle control ran with the hero video paused by
reducedMotion: 'reduce'. I made no pixel-diff claim. - Real mobile hardware and real networks. All payload figures are Chromium at a 390x844 viewport with cache disabled, not a phone; the hero-video byte count is progressive and bandwidth-dependent, so treat the 353 KB-4.28 MB spread as a lower bound on a slow connection, not a total.
- Screen-reader behaviour with an actual AT. Finding 3 is computed from the accessibility tree, not heard in NVDA or VoiceOver.
- Colour contrast of any text in the new rows — not measured.
SiteFooter.test.tscollects 0 tests on this machine (TypeError: The argument 'filename' must be a file URL object… 'file:///icons/arrow-up-right.svg'), a known pre-existing Windows path issue. I treated that as no result rather than a pass; it references none of this PR's keys. The full local suite's 249/11,625 failures are environmental and not comparable to CI.
Approval mechanics
All 10 files fall under apps/website/**, so ruleset 991238's designated-reviewer pattern routes this to the website team alone and my review on its own was sufficient — no second team's approval was structurally required here, unlike #20683. Noting for completeness that this merged on @DrJKL's approval at 02:16:41Z, directly by the author at 02:47:06Z rather than through the queue, so my review window had closed before I could post. No process complaint is intended by recording that, and the webreviewer label means only "review this PR" — it asserts nothing about who verified what.
Exposure
New feature surface on an existing live page. /fish-audio and /zh-CN/fish-audio gain five interactive audio sample rows plus seven CTAs (hero, steps, and one per row), and the shared model-launch template gains a new reusable list section available to all 15 launch pages. Not a catalogue entry and not a new app. No flag check, default, or targeting is touched — feature-flags.ts is not in this diff, and the section is gated only by page.list &&, so it ships ungated. Rollout state is unknown from the repository. Readiness isn't my call; this is the factual classification.
| /> | ||
| </div> | ||
| <button | ||
| v-if="expanded || promptOverflows" |
There was a problem hiding this comment.
BLOCKER — collapsing "Read less" destroys this button and drops keyboard focus to <body>.
While expanded is true the line-clamp-3 class is off (:102), so the paragraph is unclamped, scrollHeight === clientHeight, and the ResizeObserver at :45 writes promptOverflows = false via measurePromptOverflow (:40-42). Clicking "Read less" then clears expanded, so this v-if evaluates false || false and Vue destroys the focused node before re-creating a different one once the re-clamped paragraph measures overflow again.
Measured keyboard-only over a real build:e2e build, two independent harnesses, 2/2 viewports (390x844 row 3, 1440x900 rows 4-5):
| action | activeElement |
=== document.body |
mutations |
|---|---|---|---|
| expand (Enter) | BUTTON "Read less" |
false @0ms, @400ms | none; same node |
| collapse (Enter) | BODY |
true @0ms, @400ms, @1500ms | removed "Read less" then added "Read more" 4-6ms later; different node |
The asymmetry — expand reuses the node and keeps focus, collapse replaces it and loses focus — isolates the cause to this v-if rather than the handler. Recovery costs one extra Tab, so it is a lost focus ring and a lost place rather than permanent trapping. At 390x844 all five rows are affected; at 1440x900 rows 4 and 5.
I am not proposing a specific remedy — a suggested fix is a claim and I did not run one. The mechanism above is the whole of what I measured.
| <button | ||
| v-if="expanded || promptOverflows" | ||
| type="button" | ||
| class="mt-2 cursor-pointer text-sm/relaxed font-light text-primary-comfy-canvas underline underline-offset-4 transition-colors hover:text-primary-comfy-yellow" |
There was a problem hiding this comment.
SHOULD FIX — 81.45 x 22.75 px at 390x844, under the 24x24 minimum.
Measured getBoundingClientRect(), identical on all five rows and reproduced by two independent passes. WCAG 2.2 SC 2.5.8 asks for 24x24 CSS px, and the inline-in-a-sentence exception does not apply because this button is the sole content of its own block. Still 22.75 px high at 1440x900.
For contrast, the controls that do pass in the same row: play 48x48, copy prompt 44x44.
| :href="item.href" | ||
| target="_blank" | ||
| rel="noopener noreferrer" | ||
| class="font-bold tracking-wider text-primary-comfy-yellow underline underline-offset-4 transition-opacity hover:opacity-80" |
There was a problem hiding this comment.
SHOULD FIX — the visible mobile CTA measures 76.19 x 18 px at 390x844.
That is the smallest tap target the new rows add, and it is the primary conversion control on phones. Below SC 2.5.8’s 24x24 minimum, and not inline in a sentence (it is the only content of the <p> at :124), so the exception does not apply.
The desktop twin at :142 measures 76.19 x 24 — exactly at the line.
|
|
||
| <PlayPauseButton | ||
| :playing | ||
| :aria-label="playing ? t('player.pause') : t('player.play')" |
There was a problem hiding this comment.
SHOULD FIX — all five play buttons compute the identical accessible name "Play".
Dumped from the accessibility tree on the built page: five buttons named exactly Play, flipping to Pause on the playing row only. The same applies to the five copy buttons (all "Copy prompt", :111) and the five visible CTAs (all "TRY NOW"). Only the title links are distinguished per sample.
The <audio> at :64 does carry a per-sample aria-label, but it computes display: none with no controls, so it exposes nothing to AT and cannot supply the missing context. A screen-reader user navigating by button or link rotor — a standard mode, not an edge case — gets five identical "Play" and five identical "TRY NOW" with no way to tell which sample. The row already has title in scope.
Mutation M7 confirms nothing guards this: with the aria-label removed the button has no accessible name at all (PlayPauseButton.vue has no text node and marks both SVGs aria-hidden) and 127/127 tests still pass.
|
|
||
| <template> | ||
| <li class="flex items-start gap-5 py-8 lg:gap-8"> | ||
| <audio ref="audioEl" :aria-label="title" preload="metadata"> |
There was a problem hiding this comment.
SHOULD FIX — five eager <audio> elements pull 275 KB to 1.50 MB on load at 390x844 with no interaction.
These tags are in the SSR HTML, so they fire at parse time rather than on hydration. All five .mp3s are requested (206) on initial load without scrolling. Transferred bytes across six fresh-process runs with cache disabled: 274,885 B to 1,502,681 B, with audio.buffered.end(0) at 14.25-17.89 s and readyState: 4 on all five — seconds of decodable audio, not a metadata header.
Attribution, because it decides where any fix goes: I ran a control — a bare page with the same five <audio preload="metadata"> tags and no Vue, same browser — and it over-buffers too (96,593 / 93,730 / 68,065 / 65,536 / 90,296 B). So the over-fetch is Chromium reading preload="metadata" loosely, not component logic. What this component chose is five eager per-row <audio> elements at all.
Context: the page already serves a 5,375,238 B hero video to phones (#20932, not caused by this diff), and total wire at 390x844 came to 5.18-8.08 MB across runs. Mutation M6 shows preload is asserted by nothing — removing it entirely leaves 7/7 and 120/120 green.
| const metaLabel = computed(() => { | ||
| const length = t('modelLaunch.clipSeconds', { seconds: item.durationSeconds }) | ||
| const voices = t('modelLaunch.voices', {}, { plural: item.voices.length }) | ||
| return `${length} · ${voices} ${voiceNames.value}` |
There was a problem hiding this comment.
TRIVIAL — double separator in ja and zh-CN.
This joins with ${voices} ${voiceNames}, but the ja and zh values of modelLaunch.voices already end in a full-width colon (ボイス: / 声音:) which carries its own trailing space. Rendered: (8 秒) · 声音: Friendly Women (en) — a visible double gap.
The en value voice: uses a half-width colon and does need the space, so this is not a plain "drop the space" change. Only zh-CN is a live route today. One correction and then it is yours — moving the colon into the template or branching on locale both work, and I have run neither.
| {{ expanded ? t('ui.readLess') : t('ui.readMore') }} | ||
| </button> | ||
| </div> | ||
| <p class="mt-4 text-sm md:hidden"> |
There was a problem hiding this comment.
TRIVIAL — this wrapper renders unconditionally.
v-if="ctaLabel" sits on the inner <a> at :126, not on this <p>, so a consumer that omits the optional ModelLaunchList.ctaKey gets an empty paragraph carrying mt-4 under every prompt box below 768 px.
fishAudio sets ctaKey, so this is latent rather than live — flagging it only because the list section is now shared across all 15 launch pages.
| await nextTick() | ||
|
|
||
| await userEvent.click(screen.getByRole('button', { name: 'Read more' })) | ||
| expect(screen.getByText('[whisper] long prompt').className).not.toContain( |
There was a problem hiding this comment.
SHOULD FIX — this suite structurally cannot observe the BLOCKER, and asserts a class the repo’s own doctrine rules out.
mockPromptHeight(120, 60) pins scrollHeight/clientHeight for the whole test, so promptOverflows is true in every state — including while expanded, which is the one state where a real browser measures false and the only reason the toggle gets destroyed on collapse. The collapse exercised here is therefore the only collapse where the button survives.
On the assertion itself, docs/guidance/testing-principles.md:45-47 at this ref:
"Assert on an attribute or CSS class only when that class is the documented state contract… Do not use a class as a stand-in for the behavior it represents."
The rendered result is what would have caught it: the paragraph’s clamped vs. full height (measured clientHeight 68 px against a 22.75 px line-height collapsed, vs. scrollHeight 182-637 px), and document.activeElement after the collapse.
To be fair to the suite, it bites hard elsewhere — M1 (drop the v-for) kills 7/7, M2 (constant href) catches href identity not just count, M3 and M3b catch both halves of the one-plays-stops-others logic, and M5a-d cover the meta line including the plural wiring. Only two mutations survived, preload and the play button’s aria-label, both vacuous rather than redundantly guarded.
| durationSeconds: number | ||
| voices: readonly ModelLaunchListVoice[] | ||
| href: string | ||
| audioSources: readonly { src: string; type: string }[] |
There was a problem hiding this comment.
TRIVIAL — @mrseek-reviews’ P3 from #20898 recurs here unaddressed.
They noted on the previous incarnation that types.ts wrote out the same shape as ModelLaunchAudioCard.audioSources rather than referencing it, so a change to one source shape would not reach the other. At this head this line still reads readonly { src: string; type: string }[], verbatim identical to :121.
Structural typing means both work today; the uncaught-divergence coupling is the point. Credit to them for raising it first, and grading it as they did.
🐑 Retracting my
|
| template id | in hubTemplates.json? |
shipped CTA? |
|---|---|---|
image_qwen_image_2_1_t2i |
absent | yes — qwenImage21 |
api_nano_banana_2_1_t2i |
absent | yes — nanoBanana.ts |
utility_nvidia_rtx_video_enhance |
absent | yes — nvidiaRtx.ts |
api_openai_gpt_image_25_flare_t2i |
absent | yes |
api_fishaudio_drama3_directed |
absent | this PR |
Four already-shipped launch-page CTAs are missing from that file too. So absence is a property shared by templates that presumably work, and proves nothing. The file is a one-time snapshot of the hub browse listing — newest entry 2026-09-01, one commit ever, consumed only by code no page mounts — not the registry Cloud resolves ?template= against. That's Comfy-Org/workflow_templates, external to this repo.
I had the control available and didn't run it. I checked whether the new ids were absent, found they were, and stopped — without asking whether absence distinguishes anything. That's the same error my own bogus-id control was designed to catch, one comment earlier, on the same question.
What survives: the ids are unverifiable from inside this repo, which my control already established, and nothing here validates a ?template= id at all — a typo ships silently. Still worth one signed-in click, with the ordinary weight that carries, not the alarm I gave it.
And it's moot for merging — this landed at 02:47 on @DrJKL's approval. I was holding the queue, but the author merged directly, which is theirs to do.
@webreviewer-bot's review posted 41 minutes after the merge: 1 BLOCKER / 5 SHOULD FIX / 3 TRIVIAL. The blocker is live in production, so I'm filing it as an issue rather than leaving it in a review on a merged PR where it gates nothing.
Summary
Brings the "Made with Drama 3" sample list back to
/fish-audiowith the five Drama 3 clips Purz generated, and adds RUN DRAMA 3 buttons to the hero and the steps row that open the Cloud directed template. Each sample row shows the title with clip length and voice, the exact prompt (clamped to three lines with Read more) and a Copy prompt button, a yellow play button, and a TRY NOW link to the matching Cloud template (directed for single-voice clips, dialogue for the two-voice one).The clips are referenced at
https://media.comfy.org/website/fish-audio/<name>.mp3(hushed-gossip, movie-trailer, sleepy-to-hyped, two-voices-one-take, noir-detective) and still need uploading to the bucket; until then the play buttons do nothing. Strings added in en, zh-CN and ja.Verification
ModelLaunchListSection.test.ts(links, prompt and voices line, read-more toggle only when the prompt overflows, pausing other rows), plusmodelLaunchPages.test.tsandtranslations.test.ts; 140 passing.