UI: in the cost deep-dive sheet only (not the header), a single small gray line — MailSyncLine — reads { at, count } from the overrides metadata and shows something like Last mail sync: Sep 8, 4:12 PM · 3 receipts. No admin panel, no per-receipt list in the UI; that's intentionally out of scope. A matched receipt also becomes a real charge for that meter's windowed burn the moment it lands — see "Running costs (header + deep dive)" above.

Members module — selecting a band (or tapping "New", which mints the locked mockup's exact example band via buildNewMockBand: Grok Bot & the destroyers, with Rock Grok — role "Solo", one look already generated — and Stew Balls, no look yet) appends SkidmarksMembersModule: one shared pink-bordered box with the band name once at top, then a row per member (left: avatar — a real picked photo if one's set, else a look thumbnail if one's been generated, else a dashed placeholder ring; right: name + optional role, plus a trash/× remove control that deletes that member from the band outright). Tapping a member's row opens the generate/rename popup. The avatar itself has a tiny camera badge in its corner — same real file picker + readImageFileAsDataUrl downscale as the band cover, via setSkidmarksMemberAvatarImage — so a real photo can be attached directly, without going through the generate popup at all; a picked avatarImage always wins over a generated look. A "+ Add member" pill sits outside the box, under-right — never inside it — and appends a completely blank member (buildBlankMember: no name, no role, no emoji, no photo — nothing invented) up to MAX_MEMBERS_PER_BAND (3); a blank row shows dimmed "New member" placeholder text until it's named, given a real photo, or given a first generated look.

Generate popup (SkidmarksGeneratePopup) — a simple centered modal, no side chrome: that member's generated "looks" so far scroll horizontally across the top (three empty dashed slots before the first generate, real look swatches after), then a Name field (this is how a blank member gets a real name — commits on blur, Cancel/X/Escape, and right before Generate), a prompt textarea, a Photoreal 60–100% slider (default 80%), and Generate/Cancel. Generate has a short (700ms) fake "rendering" delay before buildMockLook mints a deterministic color-swatch stand-in (lookGradientClass, keyed off a random seed) — there is no real image model call here.

MP3 audio — appended right alongside the members module (same width/left-right alignment as its bordered box): SkidmarksMp3Card. Before anything's attached, a dashed "Attach MP3" box with a real native file picker (accept=".mp3,audio/mpeg" — no music is ever created here, only an existing file attached). Once a file's picked: a real <audio> element (via URL.createObjectURL, session- only — a File can't round-trip through localStorage, so a page reload loses playback, though the filename/duration/analysis-result metadata persists) drives a real play/pause, a real probed duration, and a real elapsed / duration readout (e.g. "1:23 / 4:16") that counts up live off the <audio> element's own timeupdate event and holds steady when paused mid-track, so Stuart can see where playback actually is instead of just the track's total length; the waveform itself is still decorative (waveformBars, deterministic off the filename, not derived from the real audio). Attaching also hands the raw File to useSkidmarksStudio, which kicks off the real analysis pass described next.

Real word-level transcription (lib/transcription.ts, transcribeAudio + segmentsFromWords, backed by app/api/skidmarks/transcribe/route.ts) — this is Stuart's actual ask after the live bug report: energy/formant heuristics + a long hysteresis hold glued "Talking To Concrete"'s intro + flute into one giant Vocal segment (0:07–4:02) when vocals really start ~0:32, so he wanted speech-to-text with word start times, like phone dictation, instead of another mid-band-energy guess. The moment a file's attached, the client POSTs it (as multipart/form-data) to /api/skidmarks/transcribe, which forwards it server-side to ElevenLabs Scribe (scribe_v2) — and only ElevenLabs Scribe, via Stuart's already-configured ELEVENLABS_API_KEY (or ELEVEN_LABS_API_KEY, checked as a fallback name). There is no OpenAI Whisper fallback — an earlier build of this feature carried one, but Stuart's explicit product call (after the fourth live bug report below) was to remove it entirely: it isn't reliable for sung tracks, and its "success" on a sparse word list was silently masking real ElevenLabs failures from him (see that report for the full story and the evidence behind it). If Scribe isn't configured, or its request fails for any reason, this route reports that failure honestly (see "Wiring up transcription" below for exactly which env var names are checked and what it costs) and the UI falls back straight to step 7's energy heuristic — and returns real per-word start/end timestamps, tagged with the provider that answered (provider, always "elevenlabs" on a real success today). segmentsFromWords then merges consecutive words into vocal runs (a gap over ~2s between words becomes an instrumental segment — see that function's doc comment), which is what lets a segment boundary land at an actual measured vocal onset instead of an energy-threshold guess. This is wired, not a stub — but it's also never claimed live without a key: if the key isn't set server-side, the route honestly returns 501/missing_api_key rather than pretending to have attempted anything, and the UI falls back to step 7 below without implying a failure. A genuine request failure (network error, bad audio, an actual upstream API error, or a request that doesn't get a response in time) is reported honestly too, distinct from "unconfigured". Word timings are kept on the attachment (words, alongside the merged segments) even though the UI only renders segments for now, so a later per-word lyric-emphasis pass (highlighting the current word during playback) can use them without re-transcribing.

Second live bug report, after the 413 fix above shipped: Stuart re-attached "Talking To Concrete" (~4:16, real singing from ~0:32) once transcription actually reached the server, and got green Lyrics + a clip list with one segment: Instrumental 0:00–4:16 — confusing and wrong, not a crash, which made it worse: the UI claimed real transcribed timing while showing the least useful possible answer. Root cause: the STT provider at the time (Whisper, a speech-first model) returned real, non-empty words for the whole song, but too sparse and scattered across the sung sections for segmentsFromWords's gap-based merge to find any vocal run worth keeping — and the code that turned a landed transcription result into segmentsSource: "transcription" / transcriptionStatus: "done" (the only path that turns Lyrics green) trusted "transcription returned some words" as "transcription produced a useful map", with no check in between. The fix has two independent parts, both load-bearing on their own:

1. Provider pivot — ElevenLabs Scribe (above) is now the primary path instead of Whisper; ElevenLabs markets Scribe explicitly for transcribing song lyrics, unlike Whisper, which is speech-oriented. This alone makes the specific failure less likely for a track like this one.
2. A real, provider-independent usefulness check — hasUsefulVocalCoverage (lib/transcription.ts) looks at the actual merged vocal/instrumental map segmentsFromWords produced and requires real singing coverage of at least min(8s, totalDurationSec * 50%) before applySkidmarksTranscriptionResult (lib/skidmarks.ts) trusts it as segmentsSource: "transcription". If a provider responds with real, non-empty words that still don't clear that bar, the attachment records the honest transcriptionStatus: "sparse" outcome instead — a provider really did answer, it just didn't yield a usable map for this track — and keeps showing whatever was already real (the energy heuristic once it lands, or the seed fallback until then) rather than the useless transcription result. Because this check runs on the real output regardless of which provider produced it, it's the actual backstop against this bug recurring even if a future provider swap or upstream model change reintroduces sparse output on some other track.

What a Concrete-like song should look like after this fix: the Lyrics chip only turns green once a provider's word timing maps to a real, substantial vocal onset/coverage (ideally landing right around the confirmed ~0:32 onset, with sung sections tracked through the track's flute/instrumental breaks); if a provider ever under-hears the singing again, the honest outcome is "sparse" — Lyrics stays amber, the timeline caption says so in plain language (naming which provider ran), and the real energy heuristic (or the seed fallback while that's still resolving) fills the clip list instead of a fake-confident wrong answer. Green Lyrics + a single Instrumental segment covering a whole sung song should no longer be reachable through this path.

Full-length songs are shrunk client-side before upload (lib/audioCompression.ts, compressAudioForTranscription) — this is the fix for a real production bug report: Stuart re-attached "Talking To Concrete" at its actual ~4:16 length and got a bare Transcription failed (Transcription request failed (HTTP 413).), even though a tiny test tone POST had already confirmed OPENAI_API_KEY was wired correctly. The cause wasn't OpenAI's 25MB limit (MAX_UPLOAD_BYTES in app/api/skidmarks/transcribe/ route.ts) — it was Vercel's own platform-level 4.5MB cap on a Function's request body (docs), which rejects an oversized multipart upload with 413 FUNCTION_PAYLOAD_TOO_LARGE before the route handler's code runs at all — a typical 128–192kbps 4+ minute MP3 is already 4–6MB, well over that cap on its own. There's no next.config/vercel.json setting that raises it. The fix: before every upload that isn't already comfortably small (DIRECT_UPLOAD_SAFE_BYTES, 4MiB), transcribeAudio now downmixes the attached file to mono, resamples it down, and re-encodes it to a duration-adaptive MP3 bitrate via @breezystack/lamejs (a pure-JS LAME encoder — no server binary, works unmodified on Vercel's Node.js runtime) entirely in the browser, picking the highest bitrate tier (MP3_BITRATE_TIERS_KBPS) whose predicted output still clears the cap. Compression is best-effort: if the browser can't decode/ resample/encode the file for any reason, the original file is still sent as-is (matching this feature's pre-fix behavior), and a genuine remaining 413 (or any other real failure) is reported in plain language — file size and the actual server limit, not a bare HTTP status code — rather than the cryptic message from before this fix.

Third live bug report, after the ElevenLabs pivot above shipped: Stuart added ELEVENLABS_API_KEY on Vercel and redeployed, then re-attached the same "Talking To Concrete" (~4:16 = 257s, real singing from ~0:32) — and got the same honest-but-wrong outcome as the pre-pivot Whisper report: amber Lyrics, caption reading "Transcription returned 28 words, but only 0.0s of that mapped to singing across a 257s track" (up from 20 words on the earlier Whisper attempt), timeline still one giant energy-heuristic Vocal block from 0:07. Stuart separately ran the exact same file through ElevenLabs' own hosted transcription tool and got back dense, correctly-timed sung lyrics with [singing] tags — direct proof Scribe can transcribe this track when it's given the original, uncompressed audio. Investigated, not assumed — five hypotheses, checked against the actual code and tests:

1. Real words landing in tiny islands mergeTinySegments folds away, throwing away good timing — checked directly: lib/transcription.test.ts's hasUsefulVocalCoverage/segmentsFromWords suite confirms 28 words spread ~7.5s apart (reproduced with the literal 28/257s/ 0.0s numbers) correctly collapses to ~0s of vocal coverage — each word is genuinely its own <1s island, further apart than the 2s gap threshold, so no merge-policy tweak recovers a real vocal run from a word list this sparse. Ruled out as the root cause; the merge math is doing its job.
2. ElevenLabs response shape mis-parsed (wrong field names/units) — checked the actual parsing in app/api/skidmarks/transcribe/route.ts's transcribeWithElevenLabs against ElevenLabs' own documented response shape (text/start/end/type, seconds, word/ spacing/audio_event) — it matches. Ruled out.
3. Coverage computed after a destructive merge instead of raw word spans — the merge (segmentsFromWords) is the thing that's supposed to turn word spans into runs in the first place; for a word list this sparse there's no "gentler merge" that manufactures 200+ seconds of coverage out of 28 real words. Ruled out as a standalone fix (see hypothesis 1).
4. Client compression before upload damages singing enough that the API returns sparse words — the one variable that actually differs between Stuart's two tests (dense results on ElevenLabs' hosted tool against the original file vs. 28 scattered words through this app's upload path). Before this fix, every full-length song was downmixed to mono, resampled to a fixed 16kHz (an 8kHz frequency ceiling — inherited from Whisper's own internal-resample rationale, never revisited once ElevenLabs Scribe became primary), and capped at a flat 64kbps ceiling regardless of available budget — for this exact 257s track, 64kbps used only ~2MB of the ~4MiB upload budget (lib/audioCompression.test.ts's regression test), real headroom left unused while handing both STT providers a fairly aggressively compressed signal to pick sung lyrics out of a full-band mix (drums, bass, flute) with. The strongest evidenced explanation, confirmed via the real lamejs encoder — not proven against the real ElevenLabs API, since this repo/environment has no live key to re-run the actual failing request end to end. Fixed: COMPRESSION_SAMPLE_RATE_HZ raised 16kHz → 22.05kHz (~11kHz ceiling instead of 8kHz) and MP3_BITRATE_TIERS_KBPS's top tier raised 64 → 128kbps — this exact track now lands at the new 128kbps top tier, ~3.9MB, still comfortably under Vercel's 4.5MB cap (verified via 's manual verification script and its own test suite), using the upload-budget headroom that was already going unused instead of a number sized for Whisper's speech-only assumptions.

Honest scope of this fix: hypotheses 1–3 were investigated and ruled out with direct evidence from the existing merge/coverage code and tests — they were not the cause here. Hypothesis 5 (the caption) is fully fixed and verified by test. Hypothesis 4 (audio compression) has the strongest circumstantial evidence of the five and is fixed in the way the bug report's own constraints asked for (more bitrate/sample-rate headroom when the byte budget allows it), verified against the real MP3 encoder's actual output — but not verified against a real ElevenLabs Scribe response on the real track, since no ELEVENLABS_API_KEY is available in this environment to run that request. If Stuart re-attaches this track after this fix ships and Scribe still returns a sparse word list, that's real signal that compression wasn't the (or the whole) story — worth a follow-up report with the new word count, since a provider that still can't hear the singing even at 128kbps/22kHz would point somewhere this PR couldn't reach (e.g. Scribe's own handling of this specific mix, or a genuinely quiet/buried vocal take) rather than back at this compression step.

Fourth live bug report, and Stuart's resulting product call — remove OpenAI Whisper entirely: after the ElevenLabs-primary pivot above shipped and started naming its provider in the caption, Stuart re-attached "Talking To Concrete" once more and the caption read "OpenAI Whisper returned 18 words…" — proof that ElevenLabs was tried first, failed, and the old Whisper fallback quietly took over and "succeeded" by the same low bar (any non-empty word list) the third bug report above was already about, without ever surfacing why ElevenLabs failed. Investigated directly against the live ElevenLabs API, using this route's exact request shape and a real (not necessarily Stuart's own) ElevenLabs key available in this sandbox: that call returned a real 401 — {"detail":{"type":"authentication_error","code": "unauthorized","status":"missing_permissions","message":"The API key you used is missing the permission speech_to_text to execute this operation."}} — i.e. a key that authenticates fine but is scoped without the speech_to_text permission. This is plausible context here specifically because Stuart's own account uses this key for voice generation (text-to-speech) elsewhere — a separate ElevenLabs permission from transcription (speech-to-text) that a key scoped for the former doesn't automatically carry. This is confirmed, reproducible behavior of the real ElevenLabs API against this route's real request — not confirmed to be Stuart's exact failure, since this sandbox's key isn't his Vercel project's key. Stuart's own product decision, independent of which exact cause this turns out to be: remove the Whisper fallback entirely — it isn't reliable for sung tracks, and a "success" on a sparse word list was actively hiding real ElevenLabs failures, which is a worse outcome than an honest failure caption. The fix, landed in this build:

• app/api/skidmarks/transcribe/route.ts no longer calls OpenAI Whisper at all — OPENAI_API_KEY isn't read by this route. ElevenLabs Scribe succeeds, or the route reports exactly why it didn't; lib/skidmarks.ts falls back to the energy heuristic either way, same as it always has when transcription isn't available.
• Every real ElevenLabs failure now reaches transcriptionError/the timeline caption verbatim — a permissions-scoped key, an expired/rotated key, a rate limit, a genuine outage, a timeout — instead of a bare "Transcription failed" or (the actual bug) a silent provider swap. The plain-language message from ElevenLabs' own error body (e.g. "…missing the permission speech_to_text…") is exactly what Stuart needs to act on: check the key's scopes in the ElevenLabs dashboard.
• Verified against Stuart's own separate Gemini troubleshooting notes as a second pass, point by point, against the actual code (not assumed): the endpoint/params/field names (file/model_id=scribe_v2/timestamps_granularity=word) match ElevenLabs' documented API; the multipart request never hand-sets Content-Type (letting fetch generate the real boundary) and always uploads a real File/Blob, never a path string; the xi-api-key auth header is correct and a 401/403 is now classified as its own "auth_error" code (classifyElevenLabsFailure in the route) rather than a generic upstream failure; ElevenLabs' documented error taxonomy ({ detail: { message, type, code, status } }) is parsed into specific codes — auth_error, rate_limited, payment_required, invalid_audio (a validation error specifically about the audio file itself — invalid_audio/ invalid_audio_format/audio_too_long/audio_too_short/ invalid_file_type), invalid_request (any other bad parameter), or upstream_error — each carrying ElevenLabs' own human-readable message verbatim.
• Encoding risk, considered and addressed with the smallest fix, not a new pipeline: Stuart's notes flagged that a highly compressed/VBR/corrupt MP3 could trip an invalid_audio rejection. For files this app already re-encodes for size (lib/audioCompression.ts's compressAudioForTranscription, anything over ~4MiB), that risk is already low — the pipeline fully decodes the original via the browser's own AudioContext and re-encodes fresh, clean CBR MP3 frames via , so the upload never carries the original file's own container quirks. The real remaining risk is a small file (under the direct-upload-safe size) sent — if upload comes back , 's now retries , forcing that same original through the same decode → re-encode pass () and re-uploading the result — reusing existing code as a normalization step, not standing up a second upload format. A WAV () upload path was considered and rejected as actually the smaller fix here: uncompressed 16-bit PCM runs 60–90x larger per second than this module's MP3 tiers, so a real song-length WAV upload would blow past Vercel's 4.5MB request body cap at any sample rate worth using for singing — infeasible as a general fallback for this app's actual use case (full songs), not just a fidelity trade-off. If the retry also fails, the real ElevenLabs error is reported rather than trying a third format. : the Whisper removal, the provider-swap-masking-failures root cause and the live 401 evidence behind it, the new error classification, and the re-encode retry are all landed and covered by and . What's verified: whether a permissions-scoped key is Stuart's real Vercel "deck" project failure (this sandbox has no way to inspect his key's scopes), and whether the re-encode retry actually resolves a real rejection against the live API on his exact track (no such failure was reproducible in this sandbox — the available key here fails Scribe call with / regardless of the audio, since the permission check happens before any file content is even read). If Stuart still sees an ElevenLabs failure after this ships, the message itself is the next diagnostic step, not a bare "Transcription failed".

Fifth report, a Vercel dashboard screenshot, not a new code symptom: Stuart confirmed ELEVENLABS_API_KEY exists on Production and Preview, added ~40 minutes before a live request still named a provider (OpenAI Whisper) this build no longer even calls — and both ELEVENLABS_API_KEY and OPENAI_API_KEY showed a yellow "Needs Attention" badge in that screenshot. Two things, both real, neither one requiring (or benefiting from) pasting any key material to diagnose:

• Redeploy after adding/changing an env var — this is documented Vercel behavior, not a guess: "Changes to environment variables are not applied to previous deployments, they only apply to new deployments. You must redeploy your project to update the value of any variables you change" ( Vercel docs). Adding a var on the dashboard does not restart or rebuild anything by itself — a Production Function that was already running, or was last built before the var existed, keeps not seeing it until a new deployment happens. How to tell which failure mode is live from the caption alone, without opening Vercel at all: if the caption still shows the missing_api_key text (see app/api/skidmarks/transcribe/ route.ts's POST, now updated to say so explicitly — "if you just added or changed it, Vercel only applies environment variable changes to new deployments — redeploy the project for this function to see it"), the deployed function genuinely doesn't see the key yet. If instead the caption names a real ElevenLabs error (an auth_error, rate_limited, etc. — see hypothesis 4 above), the key is visible to the function — the problem is on ElevenLabs' side, not deployment freshness.
• The "Needs Attention" badge is most likely a separate, security-focused flag, not a deployment-freshness one — per Vercel's Security Dashboard docs, the most likely match for "an env var that just got added and is now flagged" is the "Environment variables not marked Sensitive" check (medium risk: "Values that can be read back from the dashboard or API after they are written"). This is inferred from Vercel's own documented checks matching what the screenshot showed (one key's value still partially visible in its row, the other toggled hidden but still in a revealable, non-Sensitive format) — not confirmed by opening Stuart's actual Security Dashboard, which isn't reachable from this environment. Either way, marking a variable Sensitive is an orthogonal security-hardening step (remove and re-add it with Sensitive enabled, then rotate the previously-readable value — see that doc) — it is not the fix for a stale-deployment caption, and this PR does not touch either key's Sensitive setting. Whisper-specific asks from this report are already moot: this report's original framing (write in terms of "when EL fails and Whisper is used, surface the EL reason" / "don't let sparse Whisper hide an EL failure") predates, in this report's own words, a scenario this build can no longer produce — the fourth report above already removed Whisper from this route entirely, per Stuart's explicit product call. There is no Whisper path left for an ElevenLabs failure to hide behind; every ElevenLabs failure (including one caused by a stale deployment not seeing the key at all) now reaches the caption directly. What this report added that genuinely wasn't covered yet: the redeploy-required behavior, now both documented here and stated in the message itself ( covers that message's exact wording).

Real(ish) vocal/instrumental analysis, kept as a fallback (lib/audioAnalysis.ts, analyzeVocalActivity) — runs in parallel with step 6 above, unconditionally, the moment a file's attached, entirely client-side, no API key, no network call: it decodes the actual picked MP3 via the browser's AudioContext, frames the real PCM samples, runs a real FFT per frame, and scores each frame on (a) how loud it is (RMS), (b) how much of its energy sits in the ~300–3400Hz band human vocal formants live in, and (c) how concentrated that vocal-band energy is in a single FFT bin — a sustained near-pure tone (a lead flute is the motivating case, since it sits in the same band real vocal formants do) reads as "peaky", while a sung vowel's broader formant structure doesn't. Frames that are loud enough, vocal-band-dominant, and not too "peaky" are flagged "vocal"; an attack/release hysteresis pass (quick to call a frame "vocal", much slower to call it back "instrumental" — see the module doc comment in lib/audioAnalysis.ts) turns the frame-level flags into a handful of contiguous real-time segments without flickering mid-phrase every time a syllable, a beat, or a brief flute interlude gets in the way. It's a genuine signal-processing heuristic against the real file — not a mock — but it's also not speech-to-text and can't see song structure, so it only ever labels output Vocal/Instrumental (never verse/bridge — see below). It's tuned by ear against Stuart's reports on one track (Jack Ash – "Talking To Concrete"), not a trained/validated model, so expect it to still call loud instrumental sections, quiet vocals, or an unusually "vocal-shaped" instrumental (the flute cue is a heuristic, not a solved problem — see that constant's doc comment) wrong sometimes; that's why the UI calls it "real(ish)", not "real". This heuristic is never removed by transcription landing — it keeps running and stays the fallback signal whenever no transcription key is configured, a transcription request fails, or a provider responds but its result is too sparse to trust (transcriptionStatus === "sparse" — see step 6 above), per Stuart's explicit ask to keep it as a real fallback rather than a maybe-it-works stub. Real, useful transcription (step 6) always outranks it once it lands, though — see segmentsSource below.

Lyrics popup (SkidmarksLyricsPopup, replacing the old green Lyrics/Timing/Ready checklist chip row, removed 2026-09-14) — a small "Lyrics" pill sits inside the MP3 card itself (next to the filename), opening a compact sheet with whatever real word-level transcription (mp3.words, lib/transcription.ts's SkidmarksTranscribedWord[]) has landed for this attach, plus a copy-to-clipboard button (navigator.clipboard.writeText, with a real "Copied" confirmation and an honest inline error if the copy itself fails). Stuart's stated reasons for the removal: "Timing is redundant, the waveform already shows time," and the old three- chip row's only real payload worth surfacing was whether useful lyric text actually existed — so that's now reachable directly instead of behind a chip's color. The old chip logic (skidmarksChecklistState in lib/skidmarks.ts, four real states: grey pending, blue spinning analyzing, green done, amber stub — never staged timers) still exists as internal state (skidmarksGlance still reads its ready field for the graph card's own terse one-line status), it just isn't rendered as a standalone chip row under the MP3 card anymore. Honest empty state: if no words have landed yet, the popup names why — transcribing, not configured here, came back too sparse to trust, or failed outright — rather than showing a blank sheet or a fake placeholder. No lyrics-editing, no paste-lyrics box, no manual vocal-start pin — this is read-only, same scope the old chip row always had.

Clip / segment timeline (SkidmarksClipTimeline) — appended right under the MP3 card, as soon as an MP3 is attached (not gated on any checklist reaching "Ready" — there's no checklist row anymore, see step 8): a collapsible Clip / segment list section. Per Stuart's explicit product-lock ask (now that ElevenLabs Scribe reliably works), the long paragraph-length "honesty caption" that used to sit above the rows — naming file paths (lib/transcription.ts, lib/audioAnalysis.ts), the energy-heuristic mechanics, and the exact provider on every state change — is gone from this screen once real, useful transcription lands (segmentsSource === "transcription", the case that used to turn the old Lyrics chip green): no caption, no note, nothing between the MP3 card and the first segment row. Short of that — transcription unconfigured, too sparse to trust, or failed outright (timelineNote in SkidmarksClipTimeline.tsx) — a single plain-language line still shows (e.g. "Lyrics timing failed — showing placeholder timing below." or "…showing estimated timing below." when the energy- heuristic fallback actually produced something), with no file paths and no explanation of how either fallback works; it just says which timing is showing so a fallback is never presented as if it were real. Segments prefer real, useful transcription (segmentsSource: "transcription") whenever it lands; short of that, the real energy heuristic (segmentsSource: "analysis"); short of that, the seed cadence (buildDemoSegments in lib/skidmarks.ts — 7 segments: intro instrumental → verse → instrumental break → verse → bridge → lead → verse, labeled Verse/Bridge/ Lead/Instrumental) as a fallback, flagged by the short timelineNote line above whenever it's showing — see that function in SkidmarksClipTimeline.tsx for the exact priority logic. That priority order stops applying the instant Stuart has actually tagged real content onto the current timeline — lib/skidmarks.ts's hasSkidmarksUserContent (checked by both applySkidmarksAnalysisResult and applySkidmarksTranscriptionResult before they rebuild segments) — a real live bug: the energy heuristic resolves client-side-fast while real transcription is a genuine network round-trip against a whole song, so there was nothing stopping Stuart from already tagging a door -> keyhole -> Jack plate strip on the fast heuristic's timeline before the slower "more correct" transcription result landed and silently discarded all of it — every plate, every shot prompt — with a brand-new, blank re-segmentation and no warning (reported live as "all Vocal plates disappeared except the first one" / "clip 1 resolved back to an older clip"). Once any segment has a filled plate, a shot prompt, an extra "+" plate slot, a manual plate selection, or a manually-picked model, a later resolve still records analysisStatus/transcriptionStatus/ transcriptionProvider honestly (the chips never lie about whether a real signal landed), it just leaves segments/segmentsSource alone instead of overwriting them. Separately, every one of these async resolve callbacks (apply*Result, mark*Failed, mark*Unconfigured, setSkidmarksMp3AudioUrl) now takes the attach's own attachId (SkidmarksMp3Attachment.attachId, minted fresh in createMp3Attachment), and the store re-checks it against the live session.mp3.attachId before applying anything — a durable, store-level guard, not just useSkidmarksStudio's analysisTokenRef (a useRef on the SkidmarksDetailSheet component instance, orphaned the moment that sheet unmounts — closing the sheet while a slow real transcription call is still in flight, then reopening it, used to leave that stale promise free to land squarely on whatever's live afterward). Each clip is its own collapsible row: collapsed shows just the time range (e.g. "0:15–0:45") and its label pill — no model glance, no other chrome; expanded appends that clip's SkidmarksClipStub panel, which is ruthlessly minimal by design after a round of live QA on an earlier pass (SkidmarksPlatesAndCamera, since deleted) that shipped a horizontal row of five named location-plate cards (Neon Stage, Rainy Alley, Desert Highway, Warehouse, Crowd Pit — deterministic gradient swatches, no real plate photos), a Camera Angles block, and a compact Model pill row, with every plate card mirroring the same one shotPrompt as its caption. Stuart rejected all of it: a location picker he never asked for, model chrome he doesn't want to operate, and a prompt that visually "duplicated" into every box as he typed (even though it was really only one value underneath). SkidmarksClipStub replaces that whole panel with exactly two things — and, in a later pass than the chrome-lock rewrite itself, the first of the two now holds a real plate still, not just an empty stand-in:

• A horizontal strip of plate slots (originally just one — see "Multiple plates per clip" below for why it's now a strip) — each slot a dashed-border box (same visual language as the other empty stubs already in Skidmarks, e.g. SkidmarksGeneratePopup's EmptySlot and SkidmarksMembersModule's dashed avatar ring) while empty, no location, no gradient, no duration label (Stuart doesn't know a clip's actual rendered length until it's actually generated, so a number on an empty placeholder would just be invented). Plate stills: upload or generate, real either way — tapping an empty slot opens a tiny two-option popover (Upload / Generate, no other chrome):
  • Upload opens a real native accept="image/*" file picker; the picked photo goes through the same readImageFileAsDataUrl downscale-to-JPEG pass the band cover/member avatar pickers already use, and is stored as that slot's still (SkidmarksClipPlateSlot.still, a data: URL — round-trips through localStorage same as a picked avatar).
  • Generate calls a real backend — xAI's Grok Imagine API (lib/plateGeneration.ts's buildPlateGenerationRequest + generatePlateStill, via app/api/skidmarks/generate-still/ route.ts) — requiring only a non-empty shot prompt; everything else is automatic (see "Model routing, vocalist auto-include, and Jack Ash's character lock" and "Continuity: 'Use last plate'" below). A spinner overlays the box while a request is in flight (over the existing still, on a regenerate). This is a real, wired image call, verified live against xAI's API in this sandbox (a text-to-image call and an image-edit call with a real reference photo both returned real 200s — see lib/plateGeneration.ts's module doc comment for exactly what that does and doesn't prove) — an automatic, whole-song render past this one still stays out of scope, and there's no longer even a stub button pretending otherwise (removed 2026-09-14 — see "Wiring up the real per-clip video render" below); a real plate still is cheap, a real video render is the expensive part, which is exactly why each plate's own render stays individually opt-in with its own explicit confirm rather than firing for the whole song at once.
  • Once a still exists, Stuart's ask was gestures on the plate itself, not a button row: tapping the still reopens the same Upload/Generate popover (replace/regenerate — editing the shot prompt first, then regenerating, is how an edited prompt actually updates the still; editing the prompt alone never auto-invalidates an existing one), a tiny always-visible "×" in its corner clears it back to the empty placeholder outright, and a press-and-hold on the plate does the same clear (redundant with the ×, not instead of it — both ship). Verified manually in this build: upload, generate (with a real xAI call and a real rendered image), tap-to-replace, the "×" clear, and press-and-hold clear all worked in a real browser session; see 's doc comment for why the plain needed / for the press-and-hold gesture to survive a real mouse-down (a native browser image-drag was intercepting it before that fix).

Multiple plates per clip (the 40s-door problem): Stuart's original report on this exact live track — Concrete's intro maps as one long 0:00–0:40 Instrumental clip, and a single generated door still for the whole 40 seconds is creatively useless; he needs door → keyhole → Jack seated, each its own plate, without pulling out Resolve for a fine cut just to add two more cut points. The fix stays as small as the chrome lock allows — not a second timeline row per beat, just a horizontal strip of plate slots on the same one clip:

• A fresh clip still starts with exactly one dashed empty plate — first-use/empty state is unchanged from the single-plate build above.
• A small "+" at the end of the strip appends one more empty slot (addSkidmarksClipPlate in lib/skidmarks.ts), capped at MAX_PLATES_PER_CLIP (6) so the strip (and localStorage) can't grow unbounded. Each slot keeps every gesture the single-plate box already had — tap-to-upload/generate, tap-to-replace once filled, the corner "×", and press-and-hold to clear — completely unchanged, just scoped to one slot instead of the whole clip. An empty slot beyond the first can also be removed outright (a small "×" in its other corner, removeSkidmarksClipPlate) to undo an accidental "+" — but only while it's still empty; a slot already holding a real still has to be cleared first, so one tap can never discard a generated/uploaded image by accident.
• One shared shot prompt for the whole clip, not one per plate — the smallest control that still lets each plate show different content, per Stuart's explicit preference: a generated still already bakes in whatever the prompt said at generation time, so editing the shared prompt before tapping Generate on each slot in turn ("a door creaking open" → tap Generate on slot 1 → edit to "a keyhole with a sliver of light" → tap Generate on slot 2 → edit to "Jack Ash seated, backlit" → tap Generate on slot 3) gets door/keyhole/Jack without a second prompt field repeating the same idea three times over. See lib/skidmarks.ts's SkidmarksClipSegment doc comment for the same reasoning in the data layer.
• Continuity still makes sense across the strip: the "Use last plate" checkbox in a slot's Generate popover now offers the still-slot immediately before it in that same clip's strip (so keyhole can continue from door, and Jack can continue from keyhole) — falling back to the previous clip's last plate for the strip's very first slot, exactly like the single-plate build's cross-clip continuity already worked.
• Data shape: SkidmarksClipSegment.still (a single optional SkidmarksPlateStill) is replaced by SkidmarksClipSegment.plates: SkidmarksClipPlateSlot[] (never empty — id + optional still per slot). A session saved before this shipped had at most one top-level still; normalizeSkidmarksSegment migrates that into a one-slot plates array on load rather than dropping it, same honest "never lose real data across a shape change" pattern this file already uses for a legacy model/shotPrompt.

Clip start/end edit (added 2026-09-13, revised same day): ElevenLabs Scribe's real word-level timing lands "mostly right but sometimes 3-4 seconds off," and Stuart wants to slip a clip's cut earlier/later to keep it with the music — a lightweight trim/ slip, not a heavy NLE; the real fine-cut still happens in DaVinci Resolve. First shipped as a compact −1s/+1s button stepper (SkidmarksClipTimingNudge) opening a clip's expanded panel; replaced the same day by SkidmarksClipTimingHeaderEdit on Stuart's direct follow-up ("I hate seeing big buttons like this and wasting great real estate... I'd rather be able to double tap into the head[er] with the timing and change it") — the clip row's own always-visible "0:00–0:32" header text is the control now, no expanded-panel real estate spent on it at all. Double-tap either number (a single tap there is deliberately swallowed rather than also toggling the row expanded/collapsed, which a plain single tap on that same row already does) to turn it into a small inline m:ss text field; Enter or tapping away commits it — lib/skidmarks.ts's parseSkidmarksTimeInput parses what was typed into a delta and sends it through the exact same nudgeSkidmarksSegmentStart/nudgeSkidmarksSegmentEnd setters the old stepper called, so an out-of-range typed value just clamps the same way mashing the old buttons past their limit used to; an unparseable one cancels the edit instead of committing anything.

• Segments are always contiguous — every real source this store ever builds (the seed cadence, and the energy-heuristic/ transcription-derived timelines alike) walks a cursor forward with no gaps, so segments[i].endSec === segments[i + 1].startSec always holds going in. Nudging a clip's start/end is really nudging the shared cut point with its neighbor: lib/skidmarks.ts's nudgeSkidmarksSegmentBoundary moves both sides of that cut together — the previous clip's endSec on a start-nudge, or the next clip's startSec on an end-nudge, by the exact same amount — so the timeline can never end up with a gap or an overlap. This was the deliberate pick between the two options the original ask itself called out ("clamp so clips don't overlap" vs. "gently adjust the adjacent boundary") — a clamp-only design (only ever touching the one segment being nudged) either opens a silent gap or lets two clips overlap the moment the nudged edge crosses into the neighbor's own span; moving the shared cut point can't do either. Only the immediate neighbor at the moved cut ever changes — there's no "ripple the whole rest of the timeline" mode.
• Free edit within the song's own bounds, not a fixed ±few-seconds cap — nothing stops Stuart from tapping repeatedly past a "typical" 3-4s miss if a cut is further off than that. Clamped only by MIN_NUDGE_SEGMENT_SEC (1s — neither the nudged clip nor the neighbor it borrows from/lends to can be nudged shorter than this) and the song's own real bounds (0 at the very start; the mp3's own probed durationSec at the very end, unbounded above while that's still null — e.g. the brief window before the <audio> duration probe resolves). SkidmarksClipTimeline disables whichever of the four buttons would be a no-op (lib/skidmarks.ts's canNudgeSkidmarksSegmentBoundary, the exact same clamp math the real nudge commits with) rather than letting a tap silently do nothing.
• Never re-runs ElevenLabs Scribe or the energy heuristic, and never touches segmentsSource — a nudge only ever edits the already-resolved startSec/endSec values already sitting on session.mp3.segments, whichever real signal (or seed fallback) originally produced them; the internal checklist state and the timeline's honesty caption keep reporting exactly what they did before a nudge. It also never touches a clip's plates, shotPrompt, or model — same "one field, one job" discipline as setSkidmarksSegmentShotPrompt never touching model. Persisted the same way every other segment field already is, so a nudge survives a refresh.

Model routing, vocalist auto-include, and Jack Ash's character lock (lib/plateGeneration.ts) — all automatic, no picker/toggle UI added for any of this:

• Model routing steers prompt phrasing, not which API gets called — there is exactly one real image backend in this build (xAI, above); a clip's LTX/Grok/H3 tag (see the cost-lock section below) only changes this same call's prompt framing (routingFramingHint): Vocal clips tagged LTX Lip-sync get a tight, angled off-axis framing hint (¾/profile, mouth stays readable, eyes never toward the lens — reworded 2026-09-13 off a live-QA report that the old wording literally said "camera-facing," fighting Jack Ash's own no-stare lock below); H3-tagged clips ("simple stills" per the product ask) get a plainer single-subject hint; everything else (Grok, the instrumental/B-roll default, or a manual Seedance pick) gets a wider, dynamic establishing-shot hint.
• A single vocalist/lead auto-includes on Vocal clips (resolveVocalistForPrompt) — if a band has exactly one named member, it's them, unambiguously (both seed bands today: Jack Ash's sole member is Jack Ash; Solar Rebel's sole member is Nova); with more than one named member, whichever one's role reads like the actual vocalist/lead, falling back to the first named member. An Instrumental/B-roll clip (e.g. a door/keyhole shot) can still omit the artist entirely, even for a band with an obvious single vocalist — this only fires when the clip itself is tagged Vocal.
• Identity locks lead, the shot prompt trails (2026-09-19) — real reported bug, with a screenshot: a band member's own photo was attached and the render came back a blend, the shot prompt's described woman (afro, hoop earrings, olive tank top) fused with the photo's man (dreadlocks, purple sunglasses) into a person who was neither. No single string was wrong; it was the prompt's shape. buildPlateGenerationRequest led with the whole shot-prompt paragraph — by far the longest, most specific text in the request — and introduced the identity photo afterwards as a trailing "use this as the exact likeness reference" note, so the model averaged the two. The original Skidmarks repo (src/lib/plateCast.ts's buildPlatePrompt) never gives it that opening: it locks the person first ("same face identity, hair, age and body from image 2. Do not turn them into a different person"), forbids a second person, and puts the director's text last as Staging / tweak: — an adjustment to a locked subject, not the brief for a new one. Deck builds that order now, only when an identity reference is actually attached (a person-less B-roll plate keeps the original shape byte-for-byte). Three lock lines the old shape had none of are ported: do not turn them into a different person / do not change their gender / never merge two people into one face.
• Composite into a locked place, never paint a scene (2026-09-19) — the reorder above was not enough: The shape was right and the was still wrong. Asked for a finished scene, the model has to invent every pixel in it, the person included, and the director's paragraph is the most concrete thing in the request, so it decides who that invented person is. No lock wording fixes a request that asks for invention. Skidmarks never makes it: every plate call carries two locked images — image 1 the place, image 2 the person — with , and it refuses outright rather than proceed with either missing (, ). Deck does the same now (): a plate with a real artist to hold first generates one from that clip's own shot prompt — every person banned from the frame, including the one the shot prompt describes — and sends it as image 1 with the artist's photo as image 2. A continuity plate ("Use last plate") the locked place and takes that slot instead, never both. : a clip's plates share its shot prompt, so a door → keyhole → Jack strip generates the place once and every re-generate reuses it (cached in memory per page load, never ). A failed place still fails the whole plate on purpose — generating without it hands back exactly the drifting plate this exists to stop, while looking like a success. A person-less plate is untouched: nothing to drift, so no place still is bought and its prompt is unchanged. Auto-plate's Siray/master-still path still sends one image and is the known remaining gap.

Auto-assignment is still cost-locked in code (defaultSegmentModel in lib/skidmarks.ts) — Stuart: "be very wary of spend" — it just isn't shown as a picker anymore. It only ever picks two of the four models, and never rotates/cycles between them on its own:

• Vocal (verse/bridge, or either real path's Vocal) → LTX Lip-sync, unconditionally.
• Instrumental (lead/instrumental, any path) → Grok, unconditionally.
• H3 and Seedance are valid picks a clip's model can hold (setSkidmarksSegmentModel), but defaultSegmentModel will never return either, and — per Stuart's live-QA chrome lock — there is no model pill anywhere in this UI to tap either one from right now. The plan (not implemented) is a single tiny model badge once a clip has a real still; with only an empty placeholder today, a badge would just be more chrome around nothing.
• SIRAY and Kling are not in this SkidmarksModelId allowlist at all — SIRAY survives only as a narrow, not-wired- into-this-UI data-layer opt-in (uncensoredPlateStills, "uncensored plate stills only" — never read by model/Generate Clips); Kling has no equivalent carve-out and is removed outright (no subscription). This is unrelated to the real Siray integration Auto-plate uses (lib/sirayClient.ts, SIRAY_API_KEY, see this doc's "Auto-plate from a short brief" entry below) — that's a genuine API client picking a generation engine, orthogonal to this model still-prompt-phrasing tag entirely; uncensoredPlateStills stays exactly as unwired as described above.

The whole-song "Generate Clips" stub button that used to close this section is gone (removed 2026-09-14, Stuart's explicit ask — "It does nothing useful and confuses him"). It never called a real render for the whole song; it only showed a "Stub only — no Comfy MCP / LTX render kicked off" message, which was itself part of the problem — a pink, primary-looking button that visibly did nothing real reads as broken, not as an honest placeholder. This doesn't change the underlying cost lock at all — a whole-song, no-confirm render is still exactly the auto-fire- everything cost risk Stuart ruled out; removing the stub button is not the same ask as wiring one up, and nothing here builds that. A single plate's own render is real now — and, since the per-plate-select rework below, it's a plate's render, not a clip's: once a clip has at least one real plate still (generated or uploaded), each filled plate tile grows a small bottom-right corner control (SkidmarksPlateSelectControl, radio-style — one plate "selected" per clip at a time; tapping any filled plate's dot selects that one) that also doubles as a tick: a filled emerald check means that exact plate already has a saved render, an empty ring means it doesn't, independent of which plate is currently selected. If nothing's explicitly selected yet, lib/skidmarks.ts's resolveSelectedPlateId picks the first unrendered filled plate (or the first filled plate, if every one already has a render) — never nothing, as long as something is filled. The clip's one "Render plate" control below the strip (components/SkidmarksClipRender.tsx, inside SkidmarksClipStub) always animates just that resolved selection: its own still as the single xAI image-to-video reference, and its own stored camera- motion text (SkidmarksClipPlateSlot.motionPrompt — a small multi-line field, one stored value per plate now, not per clip and not local component state; capped at MAX_MOTION_PROMPT_LENGTH, lib/clipGeneration.ts). This replaced sending every plate on the strip as multi-reference continuity in one xAI call — Stuart's own framing: continuity across a clip's several plates (door → keyhole → Jack) now comes from rendering each plate separately with its own motion, then editing them together in Resolve, not from one bigger xAI call. Left blank, the motion field's original automatic push-in/zoom hint still applies; this is still a single free-text field, not a camera-angle picker. Duration is real and auto-computed now, not a flat 5s — lib/clipGeneration.ts's computePlateDurationSec splits the clip's own real length evenly across however many plates are on its strip (e.g. a 40s clip with 3 plates → 13s + 13s + 14s), clamped to seconds (Grok's documented ceiling) — still not a duration/resolution picker anywhere in the UI; Stuart sees the real computed number (and a real dollar estimate off it) in the confirm step, never a field to type a number into. Resolution stays fixed 480p either way. Always behind an explicit two-tap confirm, and only one plate across the whole timeline can render at a time ('s lock, keyed by ) — Stuart's "one render at a time or clear confirm" cost lock, enforced in code, not just by convention. — a breaking pathname-scheme change from the original per- layout (: , not the old ; any render already sitting under the old 2-level path is simply orphaned, not migrated). re-uploads a successful render to that stable, per-plate pathname (overwritten by each re-render of the plate — no accumulating history of past takes) and returns that durable URL; is the read side. Download uses a numeric, Resolve-friendly filename ( — e.g. for a single-plate clip, or // once a clip has more than one plate). If isn't configured, the render Stuart already paid for is still returned via xAI's own temporary URL, just honestly flagged as not saved. Seedance/Comfy MCP/LTX remain entirely unwired either way — this is xAI only, the same real backend plate stills already use, not a new provider. — see "Rendered-clips shelf" below for where it moved and why.

• Real reported bug, now fixed: "exactly one render per plate" wasn't actually enforced. The persisted filename bakes in clipIndex/startSec/endSec/a plate-count-dependent letter suffix — none of which are guaranteed stable between two renders of what Stuart still considers "the same plate" (the timeline reordering, a plate added to/removed from the same clip's strip in between) — so allowOverwrite: true alone could leave a second, differently-named blob under the same plate's own prefix instead of genuinely replacing the first: a ghost/ duplicate entry in the shelf, or "I paid for a re-render and the old clip is still there." Fixed with two layers: the route actively deletes every other blob under that plate's own prefix right after a successful save (pruneStaleRendersForPlate), and the read side defensively keeps only the most-recently-uploaded blob per plate (GET /api/skidmarks/clip-renders) in case a prune ever didn't run. A paid render's success is never silently invisible either — if xAI's call itself succeeds (money spent) but the save step afterward fails, SkidmarksClipRender shows a distinctly bordered/backgrounded alert stating plainly that the render finished, that Stuart was charged, and that it won't show up in the shelf or survive a refresh — never just a quiet "Rendered ✓" tick with nothing to show for it.

Wiring up transcription: no new key to add. Stuart confirmed he already has an ElevenLabs API key set on Vercel Production (he uses it there for voice generation in other productions), so this PR doesn't ask him to create or paste one anywhere. app/api/skidmarks/ transcribe/route.ts's resolveElevenLabsApiKey looks for it under ELEVENLABS_API_KEY first (the standard name ElevenLabs' own SDKs/docs use), then ELEVEN_LABS_API_KEY (a plausible manual-naming variant) — nothing in this repo or its sibling "Skidmarks"/"AIG!itch" project docs revealed an actual different existing name to reuse instead, so those two are the closest honest guess, not a discovered fact. If his real Vercel var is named something else entirely, the fix is a one-line alias (add a second Vercel env var under one of the two names above, set to the same value as his existing key) rather than a code change or a new key. Once found under either name, word-level transcription (step 6 above) goes live via ElevenLabs Scribe on the next deploy/restart. No OpenAI Whisper fallback exists to configure — OPENAI_API_KEY isn't read by this route at all (Stuart's explicit removal, see the fourth live bug report below); it's harmless to leave set for other features or unset, either way, but setting or unsetting it has zero effect on Skidmarks transcription. Leave ELEVENLABS_API_KEY/ ELEVEN_LABS_API_KEY both unset (or unfindable under a checked name) and the build runs exactly as before real transcription existed (energy heuristic + seed fallback), just with an honest "unconfigured" caption instead of a silent gap — that caption/error message names exactly which env var names were checked, so a naming mismatch is easy to spot and fix. Cost: ElevenLabs bills by audio duration (a few cents per hour of audio at current published rates, effectively pennies for a typical 3–5 minute song) — check ElevenLabs' current pricing before relying on this at any volume, since rates can change. This build makes exactly one transcription call per MP3 attach (plus, as of the fourth live bug report below, one further retry — same provider, a freshly re-encoded copy of the same file — specifically when Scribe reports the audio itself as invalid, never as a generic retry) — no second provider, no polling; re-attaching the same file re-transcribes it.

Wiring up plate-still generation: needs XAI_API_KEY — an xAI API key from console.x.ai — set server-side on Vercel. xAI is already one of Stuart's "four lanes" accounts (see "The four lanes" below). Stuart confirmed XAI_API_KEY is now added on the "deck" Vercel project — this PR doesn't ask him to create or add one. app/api/skidmarks/generate-still/route.ts's resolveXaiApiKey checks for it under this exact name only — there's no XAI_KEY/ X_AI_API_KEY fallback-name guessing the way the ElevenLabs route tries a second name, since XAI_API_KEY is the one name xAI's own docs/SDKs use. A key that's added on Vercel doesn't mean a currently-running Production function can see it yet — same documented Vercel behavior as ELEVENLABS_API_KEY above ( Vercel docs: env var changes only apply to new deployments) — so Production needs a redeploy after this PR merges for generate-still to actually see the key, unless merging itself already triggers a fresh deployment on this project's Vercel setup. If Generate still returns the honest "unconfigured" message after merge, that's the tell: redeploy, don't debug the code (missing_api_key, same shape as ELEVENLABS_API_KEY's own missing-key case). Leaving the key unset entirely means the same "unconfigured" message on tap instead of silently doing nothing; Upload still works either way (no key needed for that path — it's a local file read + resize, no network call). An optional XAI_IMAGE_MODEL override picks a different xAI image model without a code change (defaults to grok-imagine-image-2.0, the model this feature's live verification calls used — see lib/plateGeneration.ts's module doc comment) — a config knob, not a picker; there's still no per-request model choice in the UI. Cost: per xAI's own published pricing, a grok-imagine-image-2.0 still runs roughly one to a few cents (higher for 2K resolution or the higher-quality tier) — check xAI's current pricing before relying on this at volume, since rates can change; this matches Stuart's "plate stills are OK cost-wise" — a real video render (five to a few dozen times pricier per tap, see "Wiring up the real per-clip video render" below) is where the real spend risk lives, which is exactly why that stays behind its own explicit confirm and fixed cost cap rather than living on this same "just tap Generate" flow. This build makes one generation/edit call per tap of Generate — no polling, no automatic retry.

Wiring up the real per-clip video render (app/api/skidmarks/ generate-clip/route.ts, components/SkidmarksClipRender.tsx) — as of this pass, three real backends, routed automatically (never a persistent picker — see the cost-lock section above and this route's own module doc comment):

• Vocal/lip-sync clips → Comfy Cloud running the full LTX 2.3 IA2V graph (lib/comfyCloud.ts, workflow/LTX_2.3_IA2V_Cloud.json — copied byte-for-byte from the original Skidmarks repo, with 100+ real renders behind it; never edit it, and see lib/comfyCloud.ts for the five node inputs this app patches and nothing else), which needs COMFY_CLOUD_API_KEY (create at platform.comfy.org — an active Comfy Cloud subscription is required) and optionally COMFY_URL (leave unset/blank for Comfy Cloud's own hosted endpoint; set it only to point at a self-hosted/serverless ComfyUI instance instead). Driven by a real slice of the attached song's own vocal audio (lib/mp3Slice.ts), not XAI_API_KEY. This used to call Comfy's hosted LtxApi25AudioToVideo partner node, written off doc pages; it never survived a real call (its model input is a DynamicCombo needing a model.resolution sibling key, SaveVideo also wanted format/format.codec, and the completion reader looked for a video output key where SaveVideo actually emits images), and it hard-caps driving audio at 20s besides. The LTX 2.3 graph has none of those problems — duration is an ordinary graph input, so the 30s ceiling is real.
• Instrumental/B-roll clips → MiniMax H3 by default (Stuart's own 2026-09-13 "H3 please" ask), needing MINIMAX_API_KEY (pay-as-you-go, create at platform.minimax.io) and optionally MINIMAX_GROUP_ID (only some MiniMax accounts still ask for one) — request/response shapes mirrored from the original Skidmarks repo's own real H3 client, see lib/minimaxH3.ts's module doc comment. xAI Grok Imagine video (the same XAI_API_KEY plate stills already use, above) stays fully wired as the one-tap-away alternative — a small H3/Grok switch sits permanently beside SkidmarksClipRender's Render button (relocated 2026-09-14 off Stuart's own follow-up — it originally lived only inside the two-tap Render confirm step, which he reported as buried; "prefer visible beside the button"), and that same Render button is now about half width instead of edge-to-edge so the switch has room next to it. On a Vocal clip, the same spot shows a plain LTX label instead of a switch — there's only one real Vocal backend, so this never invents a fake second option just to mirror the Instrumental switch's shape. The choice persists per clip (SkidmarksClipSegment.instrumentalVideoModel, the existing session mirror — the Neon row as of #57, not localStorage, same as motionPrompt/selectedPlateId).
• Saving is compare-and-swap as of 2026-09-18 (app-wide, not just Skidmarks clips). The one Neon session row carries a revision; every save sends back the revision that page load actually read, and the server refuses the write (409) if the row has moved on. A page load that never managed to read the row refuses to push at all rather than writing over something it hasn't seen. Both cases show a banner saying so, and a conflict is never auto-retried — retrying is the overwrite. The reported failure this closes: a second device opened the app, showed an older copy of the project, and was one autosave away from pushing that older copy back over the good one. Edits on the refusing device stay on that device; nothing is discarded silently.
• Which copy wins on open is decided by revision, not clocks (2026-09-18). The mirror on each device records the server revision it was built from, so "this device is behind" is an exact comparison rather than a guess from two machines' clocks. The real workflow: episodes are built on the phone, then opened on a PC purely to download them for Resolve (the PC never edits) — and the PC was showing an older copy. A device holding genuinely unsent edits still keeps them even against a newer row, but that now shows as a conflict with a Load the latest saved version button rather than silently keeping the stale copy.
• None of the three keys blocks the other two paths — a Vocal render with no COMFY_CLOUD_API_KEY and an Instrumental render with no MINIMAX_API_KEY each fail with their own honest missing_api_key outcome; Grok stays reachable via the switch either way. Cost: xAI Grok Imagine video is $0.08/s at 480p (plus $0.01/reference image); MiniMax H3 is $0.08/s at 768P (its cheaper documented tier, first 5 reference images free); the Vocal/LTX confirm estimates $0.13/s (Lightricks' own published LTX direct-API rate) — but that one is a stand-in, not a verified bill: the 2.3 graph runs on Comfy Cloud's own GPUs, billed as Comfy Cloud compute/credits rather than as an LTX API call. Check each provider's own current pricing before relying on this at volume. Real, auto-computed per-plate duration (not a picker) is shown in the confirm step either way.

Wiring up render persistence: needs a Vercel Blob store connected to this project (Vercel dashboard → Project → Storage → create/connect a Blob store), which sets BLOB_READ_WRITE_TOKEN automatically — no key to create or paste in by hand, unlike XAI_API_KEY/ELEVENLABS_API_KEY above. app/api/skidmarks/ generate-clip/route.ts re-uploads a successful render to that store via @vercel/blob's put(), now under a per-plate pathname (skidmarks/clip-renders/{segmentId}/{plateId}/{filename} — a breaking scheme change from the earlier per-clip layout; see "Per-plate select rework" above); app/api/skidmarks/clip-renders/ route.ts reads it back via list(). Never blocks or fails a render if the store isn't connected yet: the render Stuart already paid for is still returned (xAI's own temporary URL) and playable, just with persisted: false and a plain-language reason shown in the UI instead of a saved state that didn't happen. Cost: negligible — Vercel Blob's free tier easily covers a handful of few-megabyte MP4s, and each plate only ever keeps its one latest render (overwritten by the next one, not accumulated), so this never grows into its own ongoing storage bill the way keeping every past take would.

Wiring up MP3 audio persistence & the finished-song archive: same BLOB_READ_WRITE_TOKEN/connected-Blob-store requirement as render persistence above — no separate key. Two client-side-direct upload paths share one token-issuing route (app/api/skidmarks/blob-upload/route.ts, handleUpload from @vercel/blob/client), rather than a normal POST body through this app's own function: lib/mp3Blob.ts (the attached MP3's own audio bytes, skidmarks/mp3-audio/) and lib/skidmarksArchive.ts (a finished song's full snapshot, skidmarks/archive/{id}/snapshot.json, plus the shared skidmarks/archive/index.json metadata list — written server-side via a normal put(), since the index itself is always small). The direct-upload split matters here specifically because both payloads can be genuinely large (a real song-length MP3; a snapshot with every plate's still embedded as a data: URL) — large enough to risk Vercel's ~4.5MB Function request-body cap the same way the transcription upload once did (see lib/audioCompression.ts's doc comment) — and upload()'s bytes go straight to Blob's own storage endpoint, never through this route. Never blocks or fails the thing it was going to save if unconfigured: MP3 playback still works from the local, in-session object URL either way (it just won't survive a refresh), and an unconfigured Archive tap fails with a real, honest message rather than silently discarding the song. Cost: same "negligible storage, not per-tap spend" reasoning as render persistence — an MP3 and a JSON snapshot are both small relative to Blob's free tier, and neither accumulates unbounded history (a re-attached MP3 or a re-archived song just gets a fresh pathname/index entry, not a growing pile). No Neon anywhere in this repo — this is Vercel Blob doing durable-storage duty until a real Neon migration lands (see the "Follow-up" note at the end of this section).

Data shape (lib/skidmarks.ts): SkidmarksBand (id, name, tagline, coverSeed, editIcon, members: SkidmarksMember[]); SkidmarksMember (id, name, optional role, emoji, looks: SkidmarksLook[]); SkidmarksLook (id, seed, prompt, photoreal, createdAt); SkidmarksMp3Attachment (fileName, durationSec, attachedAt, attachId — a stable id minted fresh in createMp3Attachment for this attach, never reused; every async resolve callback this attach kicks off (applySkidmarksAnalysisResult, applySkidmarksTranscriptionResult, setSkidmarksMp3AudioUrl, and their mark*Failed/mark*Unconfigured siblings) takes this id and the store no-ops unless it still matches the live session.mp3.attachId — the durable guard against a slow real API call landing on a different attach that's since replaced it (see the "Clip / segment timeline" step above for the exact live bug this fixes), segments: SkidmarksClipSegment[], segmentsSource: "transcription" | "analysis" | "seed-fallback", analysisStatus: "analyzing" | "done" | "failed", optional analysisError, transcriptionStatus: "checking" | "unconfigured" | "done" | "sparse" | "failed" ("sparse" — added in the same PR as this section — is the honest outcome when a provider responds with real, non-empty words that don't clear hasUsefulVocalCoverage's bar for this track: not a failure, but not trusted as segmentsSource: "transcription" or shown as green Lyrics either), optional transcriptionError (also used for the "sparse" case's plain-language coverage explanation), optional transcriptionProvider: "elevenlabs" | "openai" (which backend actually answered, once one does), optional words: { word, startSec, endSec }[] — real per-word timestamps once some provider responds (kept even for a "sparse" result), even though only the merged segments render today); SkidmarksClipSegment (id, startSec, endSec, label, model, shotPrompt — one shared value for every plate on the clip, see "Multiple plates per clip" above — uncensoredPlateStills, and plates: SkidmarksClipPlateSlot[], always non-empty — no plateId/cameraAngle; both were deleted outright along with their pickers); SkidmarksClipPlateSlot (id, optional still: SkidmarksPlateStill — the single-still field this replaced); SkidmarksPlateStill (dataUrl — always a data: URL, never a bare/temporary remote one, so it round-trips through localStorage; source: "upload" | "generated", informational only, not rendered as a badge anywhere per the chrome lock; createdAt); and SkidmarksState (bands, session: { projectKind, bandId, mp3 }, removedSeedBandIds — hand-seeded band ids Stuart has deleted, so normalizeState doesn't resurrect them). The internal checklist states (skidmarksChecklistState's lyrics/timing/ready, no longer rendered as a chip row — see step 8 above, kept only for skidmarksGlance's graph-card status line) aren't stored at all — it derives them on the fly from durationSec/segmentsSource/transcriptionStatus/analysisStatus, so there's nothing to keep in sync.

Persistence: localStorage (key the-tab:skidmarks-studio), mirroring the same in-memory-cache-plus-useSyncExternalStore shape as lib/control-plane.ts / lib/graphLayout.ts (see hooks/useSkidmarksStudio.ts, which also owns kicking off both transcribeAudio and analyzeVocalActivity against the attached file in parallel and applying whichever result lands, guarded by a generation-token ref so a slow result for a file the user has since removed/replaced can't land on top of what's current). Bands (seed + any "New" ones created this browser, capped at BAND_HISTORY_LIMIT), band/member deletions (removedSeedBandIds), wizard progress, the finished transcription/analysis results (segments + segmentsSource/ transcriptionStatus/analysisStatus, and words if transcription succeeded), and each clip's shot-prompt tag and, now, its real plate still (still, uploaded or generated) all persist; the attached audio File itself does not (see above) — which means a transcription or analysis request that's still in flight when the tab closes can never resume after a reload (no file left to re-send/ re-decode). normalizeState handles that honestly: a stuck "analyzing"/"checking" status is normalized to "failed" on load with an explanatory error, rather than leaving a chip stuck showing "in progress" forever (a pre-transcription-feature session with no transcriptionStatus at all normalizes to "unconfigured" instead — honest about "we don't know a real attempt happened here", not a fabricated interruption). A fresh browser (or private mode) always starts from the empty state; nothing here is shared across devices. This is a placeholder store, not the intended long-term one — see "Follow-up: real persistence" below.

Mock vs. real, at a glance: real — band/member identity (hand-seeded or user-created, no invented names), a picked cover/avatar photo (readImageFileAsDataUrl), deleting a band or member (removeSkidmarksBand/removeSkidmarksMember), the attached MP3 file and its real duration/playback, real word-level transcription when ELEVENLABS_API_KEY (or ELEVEN_LABS_API_KEY) is configured, the request succeeds, and the result maps to enough real singing to trust (hasUsefulVocalCoverage) (transcribeAudio/segmentsFromWords in lib/transcription.ts, via app/api/skidmarks/transcribe/route.ts), and the clip timeline's energy-heuristic vocal/instrumental segments whenever transcription isn't available or landed too sparse to trust (analyzeVocalActivity in lib/audioAnalysis.ts — a real FFT-based heuristic against the real file, see step 7 above for its honest ceiling), and — new in this pass — a clip's plate still: a real uploaded photo, or a real image xAI's Grok Imagine API generated (lib/plateGeneration.ts, app/api/skidmarks/generate-still/route.ts, keyed via XAI_API_KEY). Mock — generated "looks" (buildMockLook, a color swatch stand-in — distinct from a plate still, which is real), and the clip timeline's seed cadence (buildDemoSegments) whenever it's showing (a deterministic verse/bridge/lead/instrumental scaffold — while both real signals are still resolving, or as the honestly-labeled fallback if both failed/were too sparse to trust). See the module doc comment atop lib/skidmarks.ts for the same breakdown in code.

Follow-up: real persistence (Neon). Stuart wants Skidmarks' data (bands, members, looks, session, clip timeline) moved off localStorage onto real Neon Postgres persistence, so it survives across devices/browsers instead of being trapped in one browser's storage — matching how the rest of the app is meant to grow into "AIG!itch" backing services. That migration is a separate, larger change (a schema, a data-access layer swapping out lib/skidmarks.ts's localStorage read/write, and likely an API route) and is explicitly out of scope for this PR. The finished-song archive added in this pass (lib/skidmarksArchive.ts) is a deliberate interim step in that same direction — it's the one piece of Skidmarks state that's already durable (Vercel Blob JSON, not localStorage) ahead of a real Neon migration, precisely because "archive of record" was too load-bearing to leave in localStorage even temporarily. When Neon lands, it should absorb this archive index/snapshot shape too, not just the live session — Blob was never meant to be its permanent home, only its first durable one.

GraphView special-cases SKIDMARKS_NODE_ID (lib/constants.ts) to render SkidmarksNodeCard/SkidmarksDetailSheet instead of the generic GraphNodeCard/GraphNodeSheet, same pattern as Budju/Propfolio; the GraphBoard (≥768px) path gets the same swap, reading the studio state off the same useSkidmarksStudio store GraphView reads (one hook call, passed down as a prop, not a second independent subscription duplicating state).

Sunny Banks panel (pilot, 2026-09-15) — tapping the Sunnybank landing tile renders SkidmarksSunnyBanksPanel instead of the Music-video wizard's band picker; a sibling project kind, not a variant of Music video (no song/MP3 anywhere in it — an Australian outback adult-cartoon sitcom, 5–10 minute episodes). Two pieces:

• A Cast strip of the six locked series regulars (lib/ sunnyBanks.ts's SUNNY_BANKS_CAST — Shazza, Dazza, Nan, Nuggets, Ranger Bazza, Unit 4S; Hans is a one-episode guest and never shows in this always-on strip). Square thumbnails in one horizontally-scrollable row (2026-09-17, same shape as the "Choose a band" cover-tile strip above — not the round, wrapping chip row this shipped with initially), each showing that character's real hero still (public/skidmarks/sunnybanks/ *-hero.jpg — one cropped cell from the turnaround sheet; live-QA 2026-09-17: sending the full sheet as LTX's first frame animated a grid of Shazzas). The sheet (*-reference.jpg) stays the character bible; the cast strip uses resolveSunnyBanksStartImage. Unit 4S has no separate hero file (already a single figure). Name sits in a bottom scrim, and an honest "not ready" label on anyone missing a locked ElevenLabs voice id (Hans, today).
• A Script card — a textarea of six locked park plates used per queue row (SUNNY_BANKS_LOCATIONS, default Office Storefront until a [Location: id] tag). There is no act-level location dropdown above the script accordion. A swipeable Act I / Act II / Act III row sits at the top of that card (touch-pan-x, flex-nowrap, overflow-x), with the script tools (⇥ Format / ⤢ Full screen / ↩ Undo) on their own row underneath. Two separate rows is the fix (2026-09-18): acts and tools shared one strip, so adding the Full screen button pushed Act I/II/III clean off the left edge on a 390px phone and they read as missing. Wrapping the acts fixed that but cost two rows of vertical space for a one-line control, so they scroll again — in a strip that only ever holds acts and cannot be crowded out by a tool button. The scroll-to-end on Add Act stays, since a new act can land past the right edge. Buttons in this panel are rounded-md, not rounded-full (2026-09-18, Stuart: "square edge buttons only slightly rounded... them big round buttons, I don't like them") — the colour-key dots and the Idle/Done status chip stay round, being indicators rather than buttons. + Add Act appends the next roman in-memory buffer (IV, V, …, cap 20) with an empty script — not a Neon act table. Save and the episode zip walk every actIds entry so a typed Act IV is not dropped. The three EP02 seed acts still open first. The textarea and queued rows collapse behind Show Script Text & Queued Lines (default closed). The Script card uses touch-pan-y overscroll-y-contain; queue rows use touch-action: pan-x pan-y so iPhone Safari page scroll isn't trapped. Newlines become a dense one-row queue (parseSunnyBanksScriptBlock in components/SkidmarksSunnyBanksPanel.tsx: Name: / Name says:, empty dialogue = Hold, continuation lines keep the last speaker). God Script headers in that same paste (# EPISODE:, === ACT, [Location: id], [Action: text], [Character Name: look]) set the episode name, split act buffers, carry a locked park plate into startImageDataUrl for following rows, and append action / appearance text after the gold prompt. === LABEL === scene headers (e.g. === THE EPISODE TAG ===, === ACT III — CROWD CUTAWAY ===) stay in the in-memory parse array as sequence chunks, not queue rows; blank and tag-only lines also skip the spreadsheet. Empty Crowd: (any non-CAST Name:) is a location Hold cutaway — Idle row, park plate + [Action:] motion, no CAST overlay and no gold Hold (those strings assume one plated person). Live QA: pasting Act III on top of the 10 Done seed Speaks left Render on "Clips already loaded" because Crowd was skipped and [Action:] had nowhere to land. caravan_park_grounds is not one of the six locked plates, so that tag is ignored and the cutaway keeps the last resolved plate (dropdown still picks one). [Character Dazza wrapped in bandages] (no colon) stamps appearanceModifier on the next speaker. Render POST sends line / action / appearanceModifier / locationId (action after gold, never into TTS; gold look strings untouched). Queue rows do not scroll horizontally on a 390px phone. A Done row is static: character name (not a <select>), spoken line in a <details> disclosure under the name, status pill flex-shrink-0 on the right — no location pick, no 40px black preview cube (live QA: that cube ate the row and the spoken line vanished). Idle / failed rows keep character + location <select>s so a line that has not been billed can still be aimed, plus a spoken-line field (empty = Hold). A + on each queue row (40px hit target) inserts a silent Hold under that clip — same speaker, same plate — so a new shot can land between two already-Done rows without re-pasting the God Script; preserveRenderedRuntimes keeps the Done clips. A + above the list inserts at the start. − on an Idle/failed row drops that accidental shot. Render then reads "Render 1 line". Re-parsing keeps a Done clip when speaker + dialogue still match (preserveRenderedRuntimes); ↩ Undo beside the Act pills restores the previous script buffers. Tag-only [Character Name: override] lines never mint ghost Idle rows. Finished MP4s live in one default-open Clips shelf after the script (same card size and touch-pan-x as SkidmarksRenderedClipsShelf (w-44 / h-28, inline <video controls playsInline>), sectioned Act I / II / III with a divider between acts — not a lightbox). The panel opens on Crash Lab EP02 Drop Bears (mgen_20260827092841004_ea9: 46 already-rendered Speak clips, streamed from skidmarks.aiglitch.app — Act I caravan park, Act II BBQ shelter, Act III Unit 9 + announcement). Opens on that seed unless Neon already has a live Sunny Banks copy; Render skips lines that already have a clip so those 46 are not billed again. A named Save Project Workspace control under that strip writes the whole episode (every act) onto the existing Neon session row — one card per episode name, subtitle 3 acts · 46 clips not Act II. Live working copy persists beside the shelf so refresh does not empty it. A red ✕ drops that named card only; Download Episode Bundle (lib/sunnyBanksEpisodeBundle.ts) fetches each clip's MP4 and matching driving-audio MP3 (Crash Lab voiceFile for EP02 seed clips, or an explicit audioUrl) into video/NN.mp4 and audio/NN.mp3, with script.txt / gold prompts / URL records under data/. A dropped fetch is skipped so the zip still downloads. Not a Neon row. Speak/Hold still POST that location still as startImageDataUrl (the location canvas — Image 1) plus locationId/locationImage alongside characterName. The route (lib/sunnyBanksComposite.ts) then overlays the character hero as Image 2 via xAI edits (Studio plateCastIntoGen), and only the composed still goes to LTX node 269. The panel does not plate first — that was #120, and it would overlay twice once this route compositor is live. Not a second Comfy image node. Speak needs a locked ElevenLabs voice as well. Tap Render N lines to walk the queue sequentially via POST /api/skidmarks/sunnybank/generate-speak-beat: ElevenLabs text-to-speech (lib/elevenLabsSpeech.ts, keyed off the same ELEVENLABS_API_KEY Scribe transcription already uses — this is the first caller going the other direction, text to speech) synthesizes the line in that character's cloned voice, then feeds it into the exact same Comfy Cloud LTX 2.3 IA2V graph Skidmarks' own music-video Vocal render already calls (lib/comfyCloud.ts, COMFY_CLOUD_API_KEY) — same graph, same identity-holding LoRA, only the driving audio's source and the prompt text (Sunny Banks' own verbatim gold-doc look/style locks, not Skidmarks' 3D-noir vocabulary) are new. A Speak line that synthesizes under LTX's 2s LoadAudio floor (live QA, 2026-09-17: Shazza "You right?" → 0.8s) gets a silent MP3 tail (lib/silentMp3.ts's padMp3ToMinimumDurationSec) so node 276 actually has ≥2s; node 340:331 is that padded duration, not a fake number on short bytes. After LTX SaveVideo, lib/muxClipAudio.ts muxes that same driving MP3 onto the MP4 (-map 0:v:0 -map 1:a:0) — live QA: lips moved and the phone played silence because SaveVideo does not keep the LoadAudio track. A mux miss still returns the paid picture with audioMuxed: false. Gold Speak string unchanged; Dazza Speak/Hold prompts append SUNNY_BANKS_HELD_OBJECT_LOCK after the gold template so beers/coins/hair dryer cannot morph mid-clip. Appearance changes reach the picture, not just the motion prompt (2026-09-18) — a real reported bug: telling LTX a character is "holding two bottles" via the motion prompt alone morphed/ duplicated the bottles mid-clip, because the composed starting frame never showed them — LTX had to invent the prop from nothing. appearanceModifier (from [Character Name: description]) now also feeds compositeSunnyBanksCharacterOntoLocation's xAI edit prompt (buildSunnyBanksCompositePlatePrompt's new optional override param), so the starting frame already shows it before LTX ever runs; it's still appended to the motion prompt too, for consistency. No settle pause — the clip opens already in the new state (2026-09-18). PR #142 shipped SUNNY_BANKS_SETTLE_LEAD_SEC (1.5s) of silence prepended to a Speak beat carrying an appearanceModifier, plus a prompt sentence telling the character to hold the newly-staged pose before speaking. Stuart dropped it on sight the same day: now that the change is baked into the composed plate, frame 0 is the new action place, so there is nothing to settle into — the beat goes straight there and starts talking. A Speak beat's duration and driving audio are exactly what ElevenLabs returned; SUNNY_BANKS_SETTLE_LEAD_SEC and prependSilenceToMp3 are gone rather than left unused. The padMp3ToMinimumDurationSec tail is a different thing and stays — that's LTX's hard 2s audio-input minimum, not a deliberate pause. Don't reintroduce a lead-in without him asking. Script textarea text was invisible (2026-09-18) — SunnyBanksScriptHighlightOverlay is the only thing that draws the God Script textarea while idle (the real <textarea> is text-transparent so the colored tags can show through it), and it shipped with a text-white/0 base with only tag spans colored. On a real iPhone every ordinary dialogue line rendered black-on-black and only the bracket tags were visible. SUNNY_BANKS_HIGHLIGHT_CLASSES now carries a plain: "text-white" entry and the overlay colors every segment from that one table; a unit test asserts no entry is transparent. While focused / full-screen (2026-10-08, after PR #259 still drifted on a real iPhone) the overlay is unmounted and the textarea draws its own plain white 16px text (-webkit-text-size-adjust: 100%, autocorrect/autocapitalize/spellcheck off) so the caret sits on the glyphs you type; colours come back on blur / Done. Same lock on Music video Script Sequence, Skidmarks, and Shorts. Keep place after save (2026-10-08) — the inline box remembers scrollTop and the caret across save / re-parse / overlay remount / Done, and Full screen opens on that same place instead of autoFocus dumping you at the end; tap-back lands where you tapped, not at the bottom. Scene dividers (idle coloured view only) — a full-width rule and brighter heading sit on === ACT … — SCENE … === (Music video: each Part N ( line); none while typing. The Full screen editor has a Scenes jump strip when there are two or more. Shot list scene bars — the numbered queue repeats that heading when the scene changes, plus a short snippet of the [Action:] (or spoken line); collapsed finished shots keep the grouping. #N badges (idle only) — a small #13 in the script gutter on the [Action:] or speaker line of each shot, same number as the list row; tap the badge to scroll to the row, tap the row number to scroll the script. Display only, every genre, no parse/render/billing change. Chain from shot N does not inherit Cast (2026-10-08) — a silent House: after chaining from Arthur+Pip keeps House only (that row's own speaker + names in its own [Action:]/[Cast:]); the previous last frame is the start image, never their pictures, and that row's own [Location:] still beats a stale dropdown pick. Per-row length (2026-10-08) — silent Grok rows default to ~5s; optional [Duration: 10s] on its own line in that row's block (with [Location:]/[Action:]) sets that row only, clamped to the engine window in lib/clipGeneration.ts (Grok 5–15s). A 5/8/10/12/15s picker on the row writes/reads the same tag. Render this cost uses estimateRowVideoCostUsd at that length. Talking LTX rows keep the spoken audio unless the tag is longer, then the clip is padded. Same God Script panel for Sunnybank, Skidmarks and Shorts (Music video Script Sequence still sizes clips from Part start–end times). Full-screen God Script editor (2026-09-18, Stuart's ask: "open the God Script box to full screen and edit everything in there without it updating before we update") — ⤢ Full screen beside ⇥ Format opens SunnyBanksFullScreenScriptEditor, a fixed-inset sheet whose text is a local draft: nothing re-parses, re-derives the queue or autosaves until Done applies it through the same handleScriptChange, once. The draft buffer is the point, not the size — the inline textarea re-parses on every keystroke, so a half-typed line stops matching its rendered clip, preserveRenderedRuntimes can't rebind it, and the row flickers back to Idle mid-edit (nothing is billed by that: Render is still a deliberate tap). No colour overlay while it is open (same 2026-10-08 lock); its own Format button stays. Deliberate trade-off, stated in the sheet's own header: while it is open the draft is not saved — the one screen in this panel that opts out of keystroke autosave — and Cancel on a changed draft asks before discarding. A localStorage draft mirror with a "recover your draft?" prompt is the obvious next step if a real session ever loses work there; not built speculatively. Finished shots collapse (2026-09-18, Stuart's ask) — a default-closed N finished shots row above the queue hides every status === "done" row, so a re-rendered act doesn't bury the one line still being worked on. A failed row and the row currently rendering are never hidden: unfinished work and the thing being watched, not history. The prose hint that used to sit under the textarea is gone (one-speaker-per-line, + / −, Unit 4S stays barefoot) — every line of it is in the cheat sheet below, and two copies of the same rules is how they drift apart. The colour key stays: it is a legend for what the textarea is doing right now, not documentation. God Script Cheat Sheet (2026-09-18, Stuart's ask: "so we can always refer back to it when we're writing the next lot of scripts") — a default-closed 44px row under the script section (SunnyBanksGodScriptCheatSheet, content from lib/sunnyBanksGodScriptGuide.ts) covering the three real tags, the silent-hold shape (Name: with nothing after it), the "a [Character ...] / [Action: ...] tag is used up by the next row" rule, the speaking cast and the location ids — plus a Copy these rules as an AI prompt button (buildSunnyBanksGodScriptPrompt), since the scripts are drafted in an LLM chat and the rules would otherwise be retyped every session. Cast and location lists are derived from SUNNY_BANKS_CAST / SUNNY_BANKS_LOCATIONS rather than retyped, and lib/sunnyBanksGodScriptGuide.test.ts runs the guide's own worked example through the real parseSunnyBanksScriptBlock, so a parser change fails a test instead of leaving the cheat sheet teaching a shape that costs paid renders. Prompted by live QA: a drafted Act IV used [Outfit: ...] and [silence], neither of which is a real tag — both would have been spoken aloud by ElevenLabs in a billed clip — and arrived as one unbroken line, which the parser reads as a single clip of the whole block. Documentation only: not a validator and not a linter. Silent Hold (2026-09-17) is a script line with no dialogue after the speaker name: no ElevenLabs call. Same route with kind: "hold", a 5s silent MP3 (lib/silentMp3.ts, SUNNY_BANKS_HOLD_DURATION_SEC — LTX still needs a LoadAudio input; silence is the honest one for "No dialogue"), buildSunnyBanksHoldPrompt as the motion text. One clip at a time; a failed line stops the rest so they are not billed. A named Save Episode control at the bottom of the panel writes every act script, location id, and finished clip URL onto the existing Neon session row — not localStorage, not a new episode/beat table. Same episode name updates that one card. Live working copy persists beside the shelf so a refresh or ✕ on a named card does not wipe the episode. Download Episode (.zip) fetches MP4/MP3 bytes into video/ and audio/ (best-effort; CORS or a dead stream omits that file) and keeps script/prompts/URL records under data/. Sunny Banks has no song MP3; TTS is generated at render time. New Episode (2026-09-18) — live QA: "how do I create a new episode? I want to create new and then it clears everything, all the old stuff on the workspace should be saved." There was no such control at all — the panel could save, download, open and delete an episode but never begin one. + New Episode replaces the live working copy with a genuinely blank one (buildEmptySunnyBanksLive, deliberately not buildDefaultSunnyBanksLive, which is the Crash Lab EP02 demo seed). Saved cards are untouched. Two taps, because it discards the live copy: the first arms it and says whether what is on screen is saved yet (isSunnyBanksLiveSaved). A rename counts as unsaved — the saved card really does still carry the old name, and erring that way costs one extra tap on Save rather than costing the work. A blank episode still counts as real user content, so a refresh cannot reseed EP02 over the top of it. Clip proxy (2026-09-18) — live QA: the zip shipped 18 of 64 clips. The 18 were this app's own renders on Vercel Blob; the 46 that failed were the Crash Lab EP02 seed clips on skidmarks.aiglitch.app. A browser plays a cross-origin MP4 in a <video> fine, but refuses to let page JavaScript read its bytes without a CORS header, and that host doesn't send one — so fetch threw and the clip was silently dropped. GET /api/skidmarks/sunnybank/clip-proxy?url=… now streams a clip back from this origin (CORS is a browser policy and is never applied to a server-to-server request), and fetchSunnyBanksBinaryAsset retries through it only when the direct fetch fails — a Blob URL already works cross-origin, and proxying it too would double the bytes for nothing. The host allowlist in lib/sunnyBanksClipProxy.ts is the security boundary and is tested on its own: HTTPS only, exact host match or a dot-anchored .public.blob.vercel-storage.com suffix, no userinfo, no port, non-media content types refused, and the URL is validated before any fetch happens — the fetch itself is the vulnerability, so "fetched but not returned" would still be a hole. Save / Download / Open now show their results (2026-09-18) — live QA was "cannot download episodes", "I don't even know if the save button is working" and "how do we open it back up in an editor?". All three already worked; all three were silent or unlabelled. Save names what it saved. Download reports Getting clip N of M… per clip (buildSunnyBanksEpisodeBundle's onProgress) and then how many clips actually made it in — a 64-clip episode pulls 64 MP4s one at a time over a phone connection, which previously left the button silent for minutes and read as broken, and clipCount vs fetchedClipCount genuinely differ when a stream drops. Each saved episode is a full-width row with named Open in editor and Download (.zip) buttons, replacing the sideways-scrolling card whose only affordance was that the whole card happened to be an open target. Downloading a saved episode does not open it first (collectEpisodePrompts takes a source rather than closing over live state), so pulling an old episode never replaces what is currently open.
• Still not the episode wizard — Grok's own earlier scope was "render one speak beat, then stop." There is still no Neon episode/beat model, no lib/scriptSequenceRunner, no per-beat pathname/shelf (renders upload to a plain timestamped Blob path, not the per-(segmentId, plateId) invariant the music-video render path enforces). Last-frame chaining between beats is an optional Chain last→first toggle (default off, lib/chainLastFrame.ts) — never automatic, and never overwrites Make plate / Clip 1 upload / sleeve Keep. Each Idle/failed God Script row has Render this (one paid clip); full Render N lines stays the default. Sequential script Render is panel-local, explicit, and stops on failure — it is not music-video whole-song auto-render. Each of those already has real, working infrastructure elsewhere in this app (lib/clipRenderBlob.ts, lib/scriptSequenceRunner.ts, lib/serverVideoFrame.ts) that this will reuse once a real episode model exists to hang it off, not something to build speculatively ahead of that.
• Honesty note: not live-verified in this sandbox — no ELEVENLABS_API_KEY/COMFY_CLOUD_API_KEY configured here. The request/response shapes are correct per each provider's own documented API and this app's own already-proven Comfy Cloud LTX caller; only a real render on Stuart's own deploy proves the speak-beat pipeline end to end.

Explicitly out of scope for this build: voice and in-app stitch (still true for Skidmarks' own music-video flow — the Sunny Banks pilot above is the one narrow exception, and only for one speak beat at a time, never a stitched episode); Comfy MCP and Seedance specifically, at all, ever (no key, no endpoint, no request shape anywhere in this repo). Real Comfy Cloud LTX and real ElevenLabs text-to-speech are no longer on this list — both now real (Comfy Cloud LTX 2.3 IA2V: Skidmarks' own music-video Vocal render and the Sunny Banks speak-beat pilot above, COMFY_CLOUD_API_KEY; ElevenLabs TTS: the Sunny Banks pilot above, same ELEVENLABS_API_KEY Scribe transcription already uses) — this bullet went stale after both shipped and is corrected here rather than left asserting the old claim; ElevenLabs voice cloning management (creating/editing a voice itself, as opposed to calling an already-cloned voice id) is still unwired. Also out of scope: any automatic/whole-song clip video render (the old "Generate Clips" stub button that used to sit at the bottom of the clip timeline is gone entirely as of 2026-09-14 — see step 9's note above — but the underlying "never auto-fire a batch of paid video calls" lock is unchanged); and real trained/validated singing detection (the energy heuristic is a real signal, not a trained model — see step 7 above; the seed cadence is still a pure fallback whenever neither real signal produces anything usable). Real plate-still image generation is no longer on this out-of-scope list — see step 9's "Plate stills: upload or generate, real either way" note below, and neither is a real, opt-in, one-plate-at-a-time video render — see the "Wiring up the real per-clip video render" note above and components/SkidmarksClipRender.tsx (lib/plateGeneration.ts, xAI's Grok Imagine API); it's the video render pass past that one still that remains a stub. A UI for per-word lyric emphasis during playback (words timing is stored, per Stuart's ask, but nothing renders it yet); creating/editing music (MP3 attach is existing-file-only); real plate photos or real camera coverage capture (plates/angles are seed tags, not renders); and replacing skidmarks.aiglitch.app's own Crash Lab. Also out of scope: the "Skidmarks" landing tile (rendered, inert — its own separate Crash Lab flow). Sunnybank is no longer inert — see "Sunny Banks panel (pilot)" above; it left this out-of-scope list 2026-09-15. Also still out of scope: editing a band's name or a member's name/role after creation. Real word-level speech-to-text is now wired (see step 6 above) — it's no longer on this out-of-scope list, though it's honestly inert without ELEVENLABS_API_KEY (or ELEVEN_LABS_API_KEY) configured, and honestly falls back to the energy heuristic/seed cadence (transcriptionStatus === "sparse") on a track where the configured provider's real output doesn't map to enough singing to trust.

Multi-cast shots (2026-10-03, every genre) — one shot can hold 1–4 Cast characters, each matched to their own Cast card picture, silent or talking. One shared helper, lib/shotCast.ts (resolveShotCast), decides who is in a shot for Sunnybank, Music video and Shorts alike: an explicit [Cast: A, B] tag (Shorts: the shot's own picks), then the speaker, then any [Character X: …] look for another Cast card, then whole-word Cast names in the shot's [Action: …] text, then the other speakers of the same scene. Matching ignores capitals and spacing (STUIE = Stuie); names in the spoken words never add anyone; deduped, capped at 4, main picture only. Anyone picked with no picture refuses the row before anything is billed (missing_cast_picture).

• Plate: one xAI /v1/images/edits call — the place first, then one picture per person (xAI allows up to five images per edit: docs.x.ai "Multi-image editing", checked 2026-10-03). The prompt says "Exactly N people", labels each picture, closes every mouth, and frames a lone talker nearer and larger. One-person shots send exactly the same request as before.
• Sunnybank: talking lines straight under one [Action:]/[Cast:] with no tag between them form a scene. The first line makes the shared plate (saved to Blob as …-beat-NN-stuie-bloom-plate), the rest reuse it (scenePlateUrl), and on LTX each line appends Stuart's speaker/listener text to the gold speak prompt. Silent rows append "everyone animates, mouths closed" to the gold hold prompt. Engines unchanged: one start image into Grok 720p / H3 / LTX.
• Music video: plates send the other members after the vocalist (extraCast), MAX_REFERENCE_IMAGES is 5, and Vocal clips get the listener text.
• Silent Cast characters (2026-10-04): a Cast card with no voice ID (an animal, an extra) still gets its own silent row in Sunnybank and Skidmarks: Sparrow: with nothing after the colon sends the Sparrow Cast card picture with the location. Only talking needs a voice (a clear red note / missing_voice, nothing billed). Crowd: still never sends Cast pictures.
• Shorts: the shot picker feeds the same helper; talking shots with several people use the same speaker/listener text.
• Locations: "People already in this picture — don't add Cast" (DeckLocation.peopleInPicture) uses a pre-made plate as it is.
• UI: small cast chips ("Stuie + Bloom", components/CastChips.tsx) on Sunnybank rows, Music video clips and Shorts tiles. The Cheat Sheet (lib/sunnyBanksGodScriptGuide.ts) teaches the tags.

Episode Extras (2026-10-04, every genre) — an "Extras" card row under each episode's acts/script (Sunnybank, Skidmarks, Shorts and the Music video song sheet all mount the one components/EpisodeExtrasRow.tsx with only their genre passed in). Tap the dotted + at the end of the row to pick an outside mp4 / mov / webm / mp3 / wav (a real file input, so the iPhone Safari picker and the Windows file dialog both work), type a readable name ("container drop") and where it goes ("Act III between 8 and 9"), and it uploads straight from the browser to Blob with the existing client-upload route (/api/skidmarks/blob-upload, multipart over 20 MB, up to 2 GB) to deck/<genre>/episodes/<episode>/extras/container-drop.mp4 (Music video: deck/music-video/songs/<song>/extras/…). A name clash becomes container-drop-2, -3, never a random suffix; the token refuses to overwrite. The name and placement are saved in the same Neon session row as the episode's clips (episodeExtras in the session state, a protected key), so they survive reloads on every device. Tap a card to play it (portal lightbox), the pencil to edit the name/placement (the file is never moved), the bin twice to remove it from the episode (the Blob file is kept, like deleted clips). Episode zips get an extras/ folder named extras/act-3-between-8-and-9 - container-drop.mp4. Extras are never added to acts, never rendered, and cost nothing. Limits: the Shorts zip skips any single file over 200 MB (listed in missing.txt), and the Music video zip is the Library archive zip of an archived song.

Each Skidmarks episode has its own Cast and Locations (2026-10-04) — Skidmarks episodes are standalone shows. Tapping "+ New" on the Skidmarks Episodes row gives empty Cast and Locations rows (dotted + at the right end), an empty script, no clips and no extras; reopening EP00 — Cornish Arsehole shows Dap, Sparrow, town_street and park and its 10 clips exactly as before. Everything made before this belongs to the pilot; nothing was moved or deleted (lib/skidmarksEpisodeCast.ts). New cards and places are saved under the episode (deck/skidmarks/episodes/<episode>/characters/… and …/locations/…). The "In this episode" tick row is gone: everyone in the episode's Cast is in the episode. New Skidmarks Cast cards are still semi-photoreal. Sunny Banks is unchanged (one shared Cast).

Each Shorts episode has its own Cast and Locations too (2026-10-04) — Shorts works exactly like Skidmarks now (one shared rule, lib/episodeCast.ts). "+ New" on the Shorts Episodes row gives empty Cast and Locations rows, an empty shot, no extras and nobody starring (Starring can stay empty). EP01 (BLONDE GIRL _1) and EP02 still show Skye and SKYLAR with their shots, plates and clips exactly as before: everything saved before 4 October belongs to them, worked out on read with nothing moved, rewritten or deleted (lib/shortsEpisodeCast.ts). New Shorts people and places are saved under the episode (deck/shorts/episodes/<episode>/characters/… and …/locations/…) and stay photoreal. The episode needs a name first. Sunny Banks (shared Cast) and Music video are unchanged.

• Two real reported bugs, both fixed: first, this lock only ever applied on a Vocal clip's auto-included vocalist, so an Instrumental clip that explicitly named Jack Ash in its own shot prompt (his "door → keyhole → Jack seated" case) silently dropped both the identity reference and the hallmark lock — fixed by adding signal (a) above. Second, once continuity ("Use last plate") was also wired as a trigger, it fired on any continuity reference regardless of what the source plate actually showed: the keyhole plate (empty of people, its own prompt never names him) continues from the door plate (also empty of people), and the cruder "any continuity image at all" signal wrongly injected his silhouette/neon lips into that person-less shot purely because the band happens to be Jack Ash. Fixed by tracking whether a still is itself already known to feature the locked character (SkidmarksPlateStill.featuresLockedCharacter in lib/skidmarks.ts, set from buildPlateGenerationRequest's own resolved featuresLockedCharacter whenever a still is generated — never guessed for an uploaded photo) and gating signal (b) on that fact, not on continuity alone. Door and keyhole opener plates stay empty of people either way.