* fix(cli): non-English transcription reaches whisper, and whisper failures say why
transcribe --language <non-English> runs an installed whisper (Parakeet picks its own language); whisper models are pinned by size and sha256 and re-downloaded when damaged; a whisper-cli failure names its build, model file and last stderr lines without paths; the unused --dtw flag is dropped.
* docs(cli): --engine and --language say a non-English language runs an installed whisper
Also strips paths that follow a colon or a bracket from whisper-cli's error tail.
* fix(cli): timeline delete --start/--end removes only that range
timeline delete accepted --start and --end but never read them, so a range
delete removed the whole clip. It now cuts out only that range, the same way
split, split, delete would, and keeps the parts before and after.
Every edit verb shared one flag set, so a verb silently ignored flags it does
not use. Each verb now declares only the options it reads, from one table that
also checks apply-plan keys, so an unused flag or key is refused before
anything is written.
* refactor(cli): name the per-verb option table instead of commenting it
* perf(aws-lambda): run chunks through a Distributed Map to lift the 40-chunk cap
The inline Map runs at most 40 chunk renders at once. Both state machine
definitions (CDK and SAM) now use a Distributed Map with STANDARD child
executions, capped at the function's reserved concurrency so chunks past it
queue instead of being throttled. Progress reads each chunk's child
execution history so frames, cost and chunk errors still show up.
* fix(aws-lambda): retry Lambda throttles for minutes and stop on oversized chunks
A throttled task used to burn its four catch-all retries in about 30 s.
It now gets its own jittered backoff of several minutes, because another
render on the same function can still hold the reserved slots. Plan
treats CHUNK_EXCEEDS_WORKER_TIMEOUT as terminal in both definitions.
* chore(aws-lambda): trim stack comments to keep the comment share flat
* fix(aws-lambda): keep render --wait working at hundreds of chunks
Progress polls read the chunk Map Run's counts (one DescribeMapRun) instead of
every child execution's history, take exact chunk frames and cost from the Map
output once it exits, and read failed children only when the counts show a
failure. The SDK client retries throttled calls up to 8 times with backoff.
* feat(aws-lambda): size each render's concurrency to the account instead of a fixed 8
* fix(lambda): keep the concurrency SDK external in published modules
* fix(cli): keep an existing stack's reserved concurrency on redeploy
Redeploying without --concurrency sent ReservedConcurrency=-1, which
removed the cap of 8 that earlier CLI versions reserved by default and
stopped recording it in the state file. The override is now left out
when the flag is omitted, so SAM keeps the stack's current value (a new
stack still gets the template's unreserved default). The state file
records the deployed value, and deploy prints the kept reservation.
* fix(aws-lambda): keep progress working for callers without DescribeMapRun
Callers whose IAM policy came from an earlier `lambda policies user`
lack states:DescribeMapRun, so reading a running Distributed Map threw
AccessDeniedException out of `lambda progress` and `render --wait`
while the render itself went on to succeed. getRenderProgress now
catches only that error around the map-run read, returns the parent
snapshot (status, plan and assemble progress) and adds a warning to
re-apply the policy, which both commands print. Other errors still
throw.
* fix(cli): let the documented deploy policy drop a reserved concurrency
When a redeploy removes a function's reserved concurrency, CloudFormation
calls lambda:DeleteFunctionConcurrency. The policy printed by
`lambda policies role|user` granted only PutFunctionConcurrency, so the
update failed and rolled back. Grant it, and pin it with
states:DescribeMapRun in the required-actions test.
* fix(aws-lambda): read smoke chunk hashes from the Distributed Map output
With chunks in a Distributed Map, their task results live in child
executions, so the smoke gate's chunk-hash check found 0 chunks in the
parent history and failed every v1/v2 comparison. Read Chunks[] from the
RenderChunks / RenderChunksV2 MapStateExited output instead, as
getRenderProgress does. The stack this script deploys only has the
Distributed Map, so the inline-Map path is gone.
* fix(producer): cap a pinned chunk size at the worker concurrency, not 16
With --chunk-size or --target-chunk-frames and no --max-parallel-chunks,
the plan still capped the render at 16 chunks however many workers the
account allowed, so --chunk-size 30 on a 60 s render gave 16 chunks of
113 frames. The cap is now one wave across the worker concurrency, as
the --max-parallel-chunks docs say; 16 remains when the concurrency is
unknown.
* feat(cli): remove a stack's reserved concurrency with --concurrency=none
Since an omitted --concurrency now keeps the stack's reservation, there
was no CLI way back to the unreserved default. --concurrency=none sends
ReservedConcurrency=-1, the state file records the stack as unreserved
and deploy says the reservation was removed. Positive N and an omitted
flag behave as before; any other value is rejected instead of being
silently ignored.
* chore(aws-lambda): raise the package size budget for account-sized concurrency
Measured with pnpm@10.17.1 pack after rebasing on main with the S3 assembly change: 67,763 packed / 244,685 unpacked; the budget is the check's own 5% headroom over that.
* fix(core): color grading applies to canvas compositions
data-color-grading on a <canvas> was ignored by the runtime and refused by
lint, the Studio panel and the CLI writer. One source-tag list in the grading
contract now feeds all of them. A canvas is sized from its drawing buffer and
re-uploaded on every seek and transport tick.
* fix(core): keep the comment share of grading files at the base
* perf(cli): cloud render checks render status every 2 seconds instead of 10
A finished render was noticed up to 10 s late. Polling every 2 s cuts that
wait to at most 2 s and stays well under the status endpoint's rate limit.
`--poll-interval` still overrides it.
* perf(cli): cloud render downloads the finished video over parallel connections
The first request asks for the first 8 MB. When the server answers 206 with
the full size, the rest of the file is fetched in up to 7 more ranged
requests at once and each part is written at its offset into the staged
file. A server that ignores Range (200), a small file, or an empty file
is still fetched in one request. A failed part cancels the others and the
existing output is kept, as before.
Lambda and Cloud Run renders refused 25 fps although it is an integer
rate like 24, 30 and 60, and PAL footage is 25 fps. The distributed fps
set is now {24, 25, 30, 60} in the producer's config validation, the
plan v2 manifest, the render request mapping, the Lambda and Cloud Run
event types, the CLI's lambda and cloudrun commands and the regression
harnesses.
Co-authored-by: Cursor <cursoragent@cursor.com>
skills update (bare or named) and bare skills now exit 0 with a one-line
notice instead of installing or pruning when HYPERFRAMES_SKIP_SKILLS=1, so a
pinned skills install survives the router's trigger-time update. Closes#5695.
* feat(cli): judge text behind a transparent video by its decoded alpha in check
The layout audit treated every <video> as opaque over its whole box, so a
caption between footage and a remove-background cutout always failed
text_occluded and needed data-layout-allow-occlusion, which also hid a word
the subject's head fully covered.
check now draws the video's current (seeked) frame into one reused,
downscaled scratch canvas, once per sample and only for a video a text probe
lands under. A fully opaque frame keeps today's behaviour. A frame with
transparency is judged per rendered word against its alpha: partial cover is
an info finding carrying the most covered word's share, and a word 80% hidden
or two neighbouring words each half hidden is an error. A frame that cannot
be read falls back to opaque with a video_alpha_unread info note.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test(cli): serve the video-alpha fixtures from memory, not request paths
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): check finds words under narrow, mirrored or backed cutouts
Review follow-ups on #5657:
- A matte the probe walk passes over, even at clear-alpha points, now triggers
the per-word check, so a feature between two probe columns (an arm, a mic
stand) still fails the word it hides.
- Mirrored videos and images map viewport points through the flip; a rotated,
skewed or 3D-transformed video falls back to opaque with video_alpha_unread.
- A video with an opaque CSS background keeps its whole box occluding.
- cli.mdx lists every unread cause and why a fully transparent frame is one.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(cli): the api provider speaks the HeyGen audio analysis contract
The api rung now sends its key as x-api-key, a new Idempotency-Key on each
request, and 16 kHz mono 16-bit WAV: the channels are averaged and resampled
with a windowed sinc whose cutoff sits under 8 kHz, so nothing folds back in.
It reads the answer from a { "data": ... } envelope, or bare. Times stay in
seconds, and the answered times still snap onto beats measured on the signal
as decoded. The response type gains time_resolution_seconds.
* fix(cli): reject redirects from audio analysis requests
* fix(cli): preserve audio when stereo downmix cancels
* feat(cli): audio answers beats, downbeats, bpm, sections and describe through a provider ladder
Beats and downbeats try api, apple-mu, beat-this, then signal; sections try api, apple-mu,
then signal; describe tries api, then apple-sound. bpm is read off whichever provider placed
the beats. Each provider says whether it can run here and why not, and the ladder warns
when it passes one over.
The signal rung is a TypeScript tracker: log-band spectral flux, autocorrelation tempo with
a log-normal prior, Ellis dynamic-programming beats with a windowed local tempo, a low-band
bar phase, and Foote novelty sections snapped to beats. It reports no steady beat when the
onsets do not repeat.
The api rung posts the audio to HYPERFRAMES_AUDIO_API_URL with HYPERFRAMES_AUDIO_API_KEY and
moves each answered time onto the nearest signal beat within a quarter beat, reporting what
moved and what did not. beat-this runs Beat This! from a configured Python and local weights
and never downloads. apple-sound runs macOS's sound classifier through a Swift helper.
apple-mu reports that Music Understanding needs macOS 27.
words are no longer accepted: they come from hyperframes transcribe.
* docs(cli): audio's beats, downbeats, bpm, sections, describe and their provider ladder
* feat(cli): beats tracks the music with the TypeScript beat tracker, without Chrome
hyperframes beats and Studio's beat detection now share analyzeMusicFromSignal: band energies,
then the spectral-flux tracker in audioRhythm.ts, then each beat's loudness for its strength.
The CLI decodes with the same ffmpeg path as hyperframes audio, so beats no longer launches
headless Chrome. Studio still decodes in the browser and passes every channel.
The old analyzer averaged bpm-detective's tempo with an energy-peak tempo it had only compared
after folding both into one octave, and phased a rigid grid from noisy peaks. On 22 real tracks
with exact script grids it found the tempo 18% of the time with beat F 0.437; this finds it
91% of the time with F 0.887, and refuses the other two rather than guess.
Below a tracker confidence of 0.2 no beats are reported, so ambient or rubato music gets no
invented grid and the CLI still writes nothing. bpm-detective, the energy-peak grid, the
headless analyzer and its prebuilt bundle are gone.
* fix(core): tempo reads 96 BPM off beats on 10 ms frames, not 95.2
The tracker places beats on 10 ms frames, so a 625 ms beat lands 620 or 630 ms apart and the
median interval reported 95.2 or 96.8 BPM. Tempo now averages the intervals within 10% of the
median: as robust to a skipped or doubled beat, without the frame-sized step. Both
hyperframes audio and hyperframes beats read it.
* fix(cli): remove redundant audio context comment
* fix(core): remove redundant tempo interval comment
* fix(ci): record beat analyzer removals and browser evidence
* fix(core): keep flat audio novelty in one section
* chore: keep PR capture evidence out of the source tree
* fix(cli): honor BPM provider pins before analysis
* feat(cli): audio answers beats, downbeats, bpm, sections and describe through a provider ladder
Beats and downbeats try api, apple-mu, beat-this, then signal; sections try api, apple-mu,
then signal; describe tries api, then apple-sound. bpm is read off whichever provider placed
the beats. Each provider says whether it can run here and why not, and the ladder warns
when it passes one over.
The signal rung is a TypeScript tracker: log-band spectral flux, autocorrelation tempo with
a log-normal prior, Ellis dynamic-programming beats with a windowed local tempo, a low-band
bar phase, and Foote novelty sections snapped to beats. It reports no steady beat when the
onsets do not repeat.
The api rung posts the audio to HYPERFRAMES_AUDIO_API_URL with HYPERFRAMES_AUDIO_API_KEY and
moves each answered time onto the nearest signal beat within a quarter beat, reporting what
moved and what did not. beat-this runs Beat This! from a configured Python and local weights
and never downloads. apple-sound runs macOS's sound classifier through a Swift helper.
apple-mu reports that Music Understanding needs macOS 27.
words are no longer accepted: they come from hyperframes transcribe.
* docs(cli): audio's beats, downbeats, bpm, sections, describe and their provider ladder
* fix(cli): remove redundant audio context comment
* fix(core): keep flat audio novelty in one section
* fix(cli): honor BPM provider pins before analysis
* feat(cli): hyperframes audio measures where sounds land, in film time
Adds @hyperframes/core/audio-analysis: an energy envelope, spectral-flux
and energy onsets, the moments a sound lands (onset, steepest rise, peak,
tail end), BS.1770 loudness with true peak, silence spans and a first
guess at sfx/voice/music, all pure functions of PCM. Each capability runs
through a provider ladder so model or service providers can join later.
The radix-2 FFT moves out of audioCarve so both share it.
`hyperframes audio` decodes with FFmpeg and analyzes files, or the film's
own audio and video clips by element, second, range or boundary, reading
only the part of each file that plays. Timeline rows now carry
`mediaStart`, moved by an in-point host's cut, so file time maps to film
time.
* feat(cli): audio placements carry the element's own data-start, local to its composition
* feat(cli): audio placements carry the element's own data-start, local to its composition
* feat(cli): audio places a file with --place-at, drops --boundary
--start becomes --place-at and --media-start becomes --skip-file-seconds,
which only goes with --place-at, refuses to skip the whole file, and warns
when the skipped part held the file's loudest moment. --boundary is gone:
--at and --range cover the sounds a cut touches.
* fix(cli): pass a sound file to ffprobe as a file, on every platform
ffprobe treated a name starting with a dash as an option because the argument
list never ended the flags. The path check compared a Windows drive path to a
POSIX string, and a new timeline comment raised that file's comment share.
* fix(cli): hear a reversed stereo channel, and warn when its peak was skipped
Averaging the channels cancelled a reversed one, so a real hit was reported
as silence. The skip warning also compared against the decode margin, so a
peak just before the kept window produced no warning.
* feat(cli): snapshot --keep redoes one board still and leaves the others
* fix(skills): the board recipe waits for a released snapshot --keep
* fix(cli): snapshot --keep refuses two times with one name and replaces linked stills
* feat(cli): init --keep-original copies a clip as is, without transcoding
`hyperframes init --keep-original` copies the --video file under its own
name and wires it into the project the way a web-safe clip is wired today:
no transcode and no prompt, interactive or not. Default behaviour is unchanged.
Theora leaves init's web-safe codecs and joins the preview proxy's
browser-hostile codecs, so init converts it and the preview proxy serves an
H.264 copy instead of a blank picture.
* fix(studio-server): keep the codec table's comment share at its base level
* feat(cli): save check, snapshot and render artifacts per revision
* fix(cli): keep clean --snapshots inside the project
A symlinked .hyperframes or artifacts folder was followed, so the command
deleted snapshot and check files outside the project.
* test(cli): keep check, snapshot, and render successful when recording fails
A held history or a corrupt output registry warns and leaves the command result alone.
* fix(cli): drop snapshot records when a replaced run is not recorded
A reused --output rewrites the files even when history is busy.
Those records are removed so the old revision does not keep the new pixels.
The preview checkpoint stops waiting after five seconds.
A bad PNG header no longer fails the snapshot.
* fix(studio-server): keep the video-state comment share from rising
The forget helper's doc comment pushed the file above its base share.
* fix(cli): post the preview checkpoint and keep cleanup inside artifacts
A preview checkpoint with no body was sent as GET, so Studio never ran it.
Cleanup now requires the real path to stay under the artifacts folder.
A failed follow-up drop does not warn twice.
The studio-server unpacked budget covers the record-dropping helper.
* docs(cli): shorten the snapshot skill notes
The command examples already show range and composition.
The prose only keeps the output path and the flags that are easy to misuse.
* chore: leave package headroom for artifact cleanup
* test(cli): compare artifact paths portably
* style(cli): format portable artifact assertion
* feat(cli): guard history turns with --exclusive, --who and --if-open
* fix(cli): make exclusive history turns atomic
Re-read the open turn only after acquiring history ownership so
concurrent exclusive begins cannot both claim the same project.
Cover the race with a deterministic two-process regression test.
* fix(cli): refuse a second exclusive turn while a preview is open
The preview's begin window now rejects another agent's live turn, and an
idle turn no longer blocks exclusive begin. The two-process test waits on
pids collected from claim drafts instead of seeing both files at once.
* feat(cli): scope check to a time range or a set of components
* fix(cli): keep ranged check gates truthful
Refuse ranged caption and frame gates when none of their seek times
falls inside the requested range. Make ranged motion reporting,
overlap timestamps, filtering tests, and scope documentation match.
Add the four commands that `hyperframes --help` lists but the CLI reference
page left out, with their options taken from each command's args and
examples checked against a sample project.
Closes#5386
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): media-treatment applies one treatment to every matched image or video
* fix(cli): media-treatment reports its action from every target and documents the per-element merge
* fix(cli): media-treatment patches each match by its stable id so one write cannot redirect another
* fix(cli): media-treatment sets every match in one parse instead of re-finding elements
* fix(cli): media-treatment refuses media that is not an HTML element, as before
* chore(skills): refresh the media-use manifest hash for the treatment reference edit
* fix(cli): check passes a still title card whose timeline follows the seek
* fix(cli): check warns on a still frame and fails only when an animation is stuck
* fix(cli): count a seek clock once per sample and point the error at the registered timeline
* fix(cli): judge a finished animation by its own completion state, not its end time
* test(cli): pin a Web Animation paused at its end as finished
* fix(cli): judge a reversed animation's end in its own direction, and read plain timelines
* test(cli): pin a plain timeline held at its end as finished
* test(cli): keep an unfinished plain timeline beside the finished one
* fix(cli): skip a timeline whose time cannot be read instead of calling it stuck
* fix(cli): skip an animation whose time is not a plain number, such as a scroll-driven one
* fix(cli): a partial timeline with plain values no longer stops the browser checks
* feat(cli): let transcribe --to write srt or vtt straight from a media file
* fix(cli): fail a missing transcribe --output folder before decoding
* fix(cli): reject a transcribe --output that names a folder
* test(cli): check CJK caption grouping on the parsed cues
* fix(cli): keep transcribe --output from overwriting the media or its transcript
- Refuse an --output that resolves to the input file or the transcript.json the run writes,
before the engine starts.
- When the caption file cannot be written after transcribing, say the transcript was saved
and name the caption path and reason, instead of reporting a failed transcription.
- Docs: a one-run export groups cues for languages without spaces; exporting the saved
transcript later may split them differently.
* fix(cli): refuse caption paths that alias the media or transcript
- The overwrite check compares files, not path text: an existing target is matched by device
and inode, a new one by its real folder and its name ignoring letter case. Links, folder
aliases such as /var and /private/var, and other letter case are all refused.
- The check also covers the default caption path, so exporting transcript.srt to srt no
longer rewrites its own input.
- A trailing -o with no path fails before transcribing.
- One constant names transcript.json for the engines, the import and the check.
* fix(cli): refuse caption links to the transcript before it exists
- A caption path that is a link is followed to where the write would land, so a link to the
transcript.json the run is about to create is refused before transcribing.
- A test pins the folder letter-case check on a case-insensitive disk, and says which branch ran.
* fix(cli): check the caption file against the transcript right before writing it
- Right before the caption write, by when the transcript exists, a destination that is the
same file as the input or transcript.json is refused. The OS resolves every link, relative
link and alias, so the hand-written link following is gone.
- The run keeps the transcript and says the caption file could not be written and why.
An export whose caption write fails now says so too, instead of crashing.
- The check before transcribing stays as an early refusal for the common cases.
* fix(cli): compare caption paths by name when the disk reports no inode
- Some network shares report inode 0 for every file; such a pair is compared by folder and
name instead of being called the same file. The check before transcribing and the one at
write time now share one comparison.
- The link-through-a-file tests accept ENOENT as well as ENOTDIR, the code Windows reports.
* fix(cli): resolve both caption paths fully when the disk reports inode 0
- When either file reports inode 0 and both exist, compare their fully resolved real paths,
so a link to the media or the transcript is still refused. The folder and name compare
stays for a path that does not exist yet.
- Tests zero the inode on every path, on the caption path only, or on the protected file only.
* test(cli): say plainly why the caption-write test accepts two error codes
* feat(cli): models list shows which speech models transcribe can use
* refactor(cli): models list and transcribe share one Parakeet runner check
* fix(cli): models list skips whisper links transcribe would download again
* style(cli): keep models list comments under the package share
* fix(cli): enforce capture screenshot limits and report interruptions
* refactor(cli): split screenshot step out of capture animation phase
Keeps the screenshot limit outside the mirrored core-extraction argument
list and moves scroll screenshots into their own step, so both stay under
the audit's duplication and complexity gates with no behaviour change.
* fix(core): write only authored hue curves, so a single hue curve passes lint
* fix(cli): apply over a grading saved with empty hue curves instead of refusing it
* fix(core): read grading properties from computed style, and stop advising an inline start
* fix(cli): preserve validation during legacy hue repair
* fix(cli): preview grading writes without apply
* fix(auth): accept host-managed HeyGen access tokens
* fix(auth): share environment header validation
* docs(auth): condense credential resolver guidance
* fix(auth): report the account a browser login just signed in
Identify the account from the tokens the login issued, not a fresh credential
lookup that an environment token or API key outranks.
* docs(auth): list the host access token in the CLI credential order
* fix(cli): name a rejected host access token on publish
Publish told a host-token user to log in again, which cannot help while the token outranks the login.
* docs(auth): show the media credential order main resolves
* fix(cli): warn at logout that a host access token still signs commands
One map of environment credential variables, owned by the resolver, now feeds the logout warning and both publish messages.
* docs(auth): scope the media credential order to the audio scripts
* refactor(cli): read every env credential name from the resolver's map
* docs(skills): ask agents inside a host app to report its friction as feedback
* feat(cli): send missing features and host-app friction as ratingless feedback
* fix(cli): keep the rating error, refuse blank reports and document ratingless feedback
* docs(skills): leave names and figures out of missing-feature reports
* docs(skills): keep anything private out of missing-feature reports
* fix(check): see color, filter and media time under seek, and name data-no-timeline for stills
* fix(check): warn instead of fail when only audio advances, and judge text overlap on glyphs
* fix(check): judge glyph overlap by the glyph box, and see capitalized words and one-side borders
* fix(check): capitalize past leading punctuation, test outline color, trim a comment
* fix(check): capitalize words the way Chromium does, so digit-led words stay lowercase
* fix(cli): let parakeet-mlx detect the language and name the Parakeet install under auto
* docs(cli): say what --language does under each engine
* docs(cli): whisper transcribes as the --language, it does not filter
* chore(cli): drop a comment the install-hint condition already says
* feat(cli): feedback says whether a person or an agent wrote it, with the email an app attaches
* refactor(cli): split the .env loader so each part stays simple
* fix(cli): a broken .env loader fails loudly; only a missing .env is ignored
* fix(cli): distinguish a skipped browser check from a genuinely clean one
`hyperframes check`'s runtime/layout/motion/contrast JSON sections reported
the exact same ok:true/zero-findings shape whether the browser session
actually ran and found nothing, or never ran at all (a blocking lint error,
a lint-step crash, or an exception thrown before the audit could start all
took the same "empty result" fallback). Only the top-level ok field and an
undocumented layout.duration === 0 tell separated the two cases apart.
Add a skipped boolean to CheckBrowserResult, set at the only two places one
is constructed (false in the real audit result, true in the empty-result
fallback), and mirror it to a new top-level browserSkipped field on
CheckReport. The human-readable report now prints a warning when the
browser never ran; the JSON envelope carries the same signal for callers
that only check individual sections.
Co-Authored-By: Miguel Angel <miguel.sierra@heygen.com>
* docs(cli): document browserSkipped in the check --json envelope
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* refactor(cli): trim browserSkipped docs and cover the linter-crash skip path
* chore(skills): refresh the hyperframes-cli manifest hash
* refactor(cli): drop the browserSkipped type comments to keep the comment share flat
---------
Co-authored-by: Miguel Angel <miguel.sierra@heygen.com>
* fix(cli): fingerprint textual sweep state
* fix(cli): share one motion classifier between sweep guard and liveness
The frozen-sweep guard (sweep_static) and keepsMoving liveness (motion_frozen)
each owned their own definition of "what counts as motion": the sweep
fingerprint in layout-audit.browser.js read nine channels while
motion-sample.browser.js compared box + opacity only. They meet in one decision
in checkPipeline.ts, so a fixed-width attr() countdown was accepted by the
sweep guard and still reported frozen by liveness.
Move the classifier into motion-signature.browser.js, injected before both
samplers, and make it the single owner:
- Per-element channels (box, opacity, font-variation axes, clip-path, direct
text, form-control state, ::before/::after content, media pixels) as a
reader list so further channels (e.g. SVG dash) plug in.
- Liveness consumes the same signature with 2px / 0.08 bucketing; the sweep
guard consumes it exact. motion-sample throws at install if the shared
script is missing.
- Counter state is a composition-level channel over box-generating elements
and root ancestors. display:none subtrees generate no boxes and are
excluded, and only declarations naming a counter that some painted
generated content (::before/::after/::marker on a non-ignored, paintable
host) consumes are folded in, so a decoy hidden counter cannot make a
frozen composition read as live.
- clip-path is a channel, replacing the fingerprint's former reliance on the
clipped-away visibility probe, so a wipe over a static box still counts.
- data-layout-ignore layers are excluded from the signature but not from the
per-selector `visible` bit reported for explicit motion assertions.
- Root resolution for the "*" liveness scope now matches the sweep guard and
the root element is part of its own signature.
Tests: motion-signature.browser.test.ts (happy-dom, per-element fake computed
style so ancestor handling comes from the classifier) and
motion-signature.browser.chromium.test.ts (real Chromium via puppeteer-core;
skips without a system browser) pin every branch: attr() substitution,
display:none decoys, named-counter gating, ::marker, zero-box hosts, ancestor
counters, textarea/indeterminate, root text, clip-path, ignore layers, and
sweep/liveness agreement on the countdown fixture.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* fix(cli): prune unrendered content from the motion signature and measure asserted scopes
The shared motion classifier pruned only display:none subtrees, so text,
pseudo content, media pixels, and counter decoys inside skipped
content-visibility contents changed the frozen-sweep and liveness
signatures although nothing painted. Pruning is now platform-owned where
checkVisibility exists (display:none, skipped hidden/auto contents; a
display:contents host keeps its painting pseudo-elements), with the
computed-style fallback for happy-dom. A content-visibility:hidden host
signs its own box, opacity, clip-path, font axes, and widget state, but
not its text, pseudo content, control value, or media pixels; hosts where
size containment does not apply (non-atomic inline, display:contents,
table and internal table boxes, ruby) skip nothing. isVisibleElement
regains the checkVisibility floor layout-audit uses on its opacity-floor
path, so an asserted selector inside skipped contents reports hidden.
The layout-audit opt-out no longer applies to the element passed as the
signature root: a keepsMoving scope that is itself data-layout-ignore is
measured (an assertion naming an element outranks the opt-out) instead of
producing an empty signature that motionAudit reports as a missing
selector. Opt-outs inside the scope still apply.
motion-sample defers the missing-classifier error to the first call so
the driver's evaluate surfaces it instead of an opaque "is not a
function" page error; round() is shared.
The real-Chromium suite is opt-in on Windows (HYPERFRAMES_BROWSER_TESTS=1):
in the shared Windows package lane the runner's Chrome did not launch
within the package hookTimeout, and the suite already runs on Linux CI
where the system Chrome is found. The launch hook has its own 120s ceiling.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* fix(cli): prune display:contents in skipped contents; table-cell skips, caption paints
A display:contents child of a content-visibility:hidden host has no box
of its own, so the platform check could not tell it apart from an
ordinary display:contents owner and it stayed a counter consumer: a
change of the owner's counter-reset altered the sweep signature although
nothing paints. The parent's skipsContents verdict now prunes it; boxed
descendants keep being decided by the platform.
NOT_CONTAINABLE_DISPLAY had table-cell and table-caption inverted
versus Chromium 152, where a table-cell host skips its contents and a
table-caption host paints them.
Four Chromium pins: the display:contents-in-skipped-contents counter, a
table-cell host's own text and pseudo content (frozen), a table-caption
host's own text and pseudo content (motion), and a counter reachable
only through a skipped host's hidden ::marker / ::after (frozen; also
asserts the skipped child is not visible for motion-sample).
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* chore(cli): trim sweep motion comments to the comment ratchet
* chore(cli): shorten sweep motion comments to the merge-hook limit
---------
Co-authored-by: miga-heygen <miguel.sierra_miga@heygen.com>
* feat(cli): one owner for init media placeholders, audio stays on the video
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): init keeps a video's sound on the clip; silent files scaffold muted
Supersedes the patchVideoSrc changes in #3977 and #3942 and carries their tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: miga-heygen <miga-heygen@users.noreply.github.com>
Co-Authored-By: kapelame <kapelame@users.noreply.github.com>
* feat(registry): example templates keep the a-roll's sound on the video
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(lint): video audio rules recommend keeping sound on the video
Update video_missing_muted and video_audio_double_source message and
fixHint text to recommend the new default: keep audio on the video with
data-has-audio="true" (not muted). Tests verify the new wording is in place.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(parsers): a video is either audible (data-has-audio) or muted, never both
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio): follow the compiler's audio rule in the audio-track toggle
Confirm before background removal drops a video's sound.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(skills): hyperframes-core teaches sound on the video by default
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(skills): workflows and scaffold guidance keep a video's sound on the clip
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: footage with sound keeps it on the video clip
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): init from-file a-roll is timed, muted strip is attribute-exact
The from-file a-roll video gets data-start="0" so a scaffold lints clean, and
the orphaned audio-only slot comment is dropped. setVideoAudio now walks the
tag attribute by attribute, so a class value containing the word muted
survives. The ffprobe-missing warning says the video is assumed to have sound
and how to switch to muted, and DEFAULT_META notes the failure surfaces at
render until lint support lands.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio): keep has-audio in step with the muted toggle, test cutout confirm
Muting a video now clears has-audio and unmuting sets it to true, in both the
Design and flat media panels, matching the Has audio track toggle. Audio
elements are unchanged. Adds coverage for both panels plus the cutout
window.confirm guard: declined leaves src alone, accepted swaps src and mutes,
a silent video never prompts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(parsers): updateElementInHtml hasAudio only applies to videos, true or false
An audio element or an undefined hasAudio no longer picks up a muted attribute;
only an explicit false on a video writes muted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: video_missing_muted rule text, silent-footage notes, consistent audible examples
lint.mdx and cli.mdx describe the current video_missing_muted rule instead of
the retired unmuted-video wording. The editing recipes say how to adapt for
silent footage and require the outgoing clip's sound to end at a J-cut. The
Remotion media example shows the audible mapping consistently, data-attributes
states that has-audio is required unless muted, and the hairline frame note no
longer claims the video is muted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): encode init media filenames and probe --audio duration
Function replacers stop dollar-sign pattern expansion in filenames; src is percent-encoded.
--audio now passes the probed track length as the composition duration in both init paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(lint): audio-state-aware media rules and time-aware double-source check
video_missing_muted honors any explicit data-has-audio.
video_audio_double_source only considers audible videos and only fires
when the same-src clips overlap in time.
duplicate_audio_track now covers audible videos.
audio_file_without_element no longer claims silence when an audible video
carries the sound.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(skills): audio-stream check for remotion ports, followable J-cut recipe
Remotion Video maps to data-has-audio only when the file has an audio stream.
The J/L-cut recipe now shows both shots with muted pictures and separate audio.
embedded-captions removes the audio.mp3 scratch file it extracts.
skills-manifest.json regenerated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(lint): trim comments to satisfy the comment ratchet
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio): atomic muted edits, cutout keeps sound on a linked audio
Muted and Has audio track commit one batched write (one persist, one
undo entry) and remove data-has-audio instead of writing an empty
marker. Unmuting a video probes its src: audible -> true, silent ->
false, probe failed -> absent. Remove background on an audible video
inserts a linked sibling <audio> with the original sound in the same
write, replacing the drop-the-sound confirm.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(producer): skip hidden videos when auto-adding video audio
A video hidden in Studio is dropped by the mixer, so its src never reached
existingAudioSrcs and the render re-added it at full volume. The video parser
now records the shared hidden predicate and the auto-detect step skips hidden
videos.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio): keep useDomEditSession under the 600-line cap
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: miga-heygen <miga-heygen@users.noreply.github.com>
Co-authored-by: kapelame <kapelame@users.noreply.github.com>
* feat(cli): add hyperframes clean to find and remove what HyperFrames left on disk
`hyperframes clean` lists leftover render work dirs, debug renders, the history of projects
that no longer exist, idle video proxies and extracted video frames, with sizes, and removes
them. `--dry-run` removes nothing, `--json` is for agents, `--snapshots` also removes
snapshot folders (otherwise listed as reclaimable). It never touches outputs, sources or
anything a running render or preview still owns.
* docs(cli): say that clean does not search a project's subfolders
* fix(cli): clean test compares folder names on Windows and the skills manifest is regenerated
* fix(studio-server): the proxy cache sweep removes only files the transcoder named
It removed any idle .mp4 or .webm in .transcode-cache, so a user file placed there, or reached through
a symlinked cache folder, could go. Proxies are always <sha256>.<ext> and their temp files
.tmp-<uuid>-<proxy name>; nothing else is swept.
* style(producer): keep a renderDirOwner doc comment within 120 columns
* feat(cli): tag telemetry and feedback with the app that launched the CLI
* fix(cli): keep the client tag to a short slug and correct the telemetry docs
* test(cli): read the delivered batch through the shared helper
* refactor(cli): let the tag pattern alone bound the client tag's length
* docs(cli): describe the client tag generically as the launching app's tag
* style(cli): format the client tag test cases
* test(cli): use a neutral example value for the client tag
* chore(skills): regenerate the skills manifest for the telemetry reference
* docs(cli): render help says when page-side compositing is actually used
* docs(cli): page-side compositing help names webm and mov and the fast-capture flag
* docs(cli): say plainly that gif keeps page-side compositing
* fix(cli): commands exit without waiting on update checks or telemetry
The update and skills checks run in a detached child that refreshes the
caches for the next run; the parent only reads the caches. finalizeCli no
longer awaits the telemetry flush: the exit handler flushSync() delivers
what is queued. The events beacon, feedback and figma error paths stop
awaiting flush() too.
* fix(cli): retry a failed background check at most once an hour
Before starting the background worker, the CLI stamps a per-check attempt
time and skips a check it attempted within the hour, so a dead network no
longer starts a worker on every command. The stale-pin notice writes onto a
fresh read, detached spawns set windowsHide, and a test ties the worker path
to its build entry.
* fix(cli): a failed-check stamp dated in the future no longer silences update checks
* feat(cli): download Parakeet with models install parakeet and transcribe with it
`hyperframes models install parakeet` installs the sherpa-onnx runtime from
npm and the Parakeet TDT 0.6B v3 int8 model from Hugging Face, pinned to a
revision. Every file is checked against its pinned size and sha256 under a
temp name before it is renamed into place. Re-running it is a no-op. On
Linux with glibc older than 2.32 it refuses up front.
`transcribe --engine auto|parakeet` now prefers that engine when it is
installed, then an installed parakeet-mlx runner, then whisper. Audio is cut
into about 60 s windows at the quietest 100 ms near each cut, and decoding
runs in a child process. If the child cannot load the runtime or crashes,
transcribe falls back with one stderr line saying why. The JSON output and
transcript.json shape are unchanged.
* fix(skills): report the real word count and engine from transcribe fallbacks
The media-use transcribe script read `words` from the CLI's transcript,
which is a flat array, so its whisper path always reported an undefined
word count. The embedded-captions script labelled every CLI transcript
whisper.cpp even when `--engine auto` ran Parakeet; it now uses the engine
the CLI reports.
* fix(cli): pin the sherpa-onnx native package to the runtime's exact version
* test(cli): type the sherpa install test doubles
* fix(cli): transcribe audio with no speech as 0 words instead of failing
Parakeet writes an empty word list for a music bed, a silent intro or a
very short clip, and loading it failed with "Unrecognized JSON transcript
format". An empty list is now read as a transcript with no words, as
whisper's empty output already is. A non-empty list still has to look like
words.
* fix(cli): re-decode a Parakeet window that skipped the start of its speech
At some window lengths the model silently drops the first clause: on a
looped speech clip, 4 of 167 lengths between 55 and 60 s, and a 61.91 s
window, lost their first 9 words. A window with a token gap of at least 1 s
that is as loud as speech is decoded again with 0.5 s of leading silence,
and the result with more tokens wins. Clean windows are unchanged.
Decoding with 1.5 s of overlap before each cut was measured too and was
worse on 573 s of audio (2.62% WER against 0.98%), so it is not used.
Empty audio no longer makes a zero-length window, which aborted the
decoder.
* fix(cli): stop Parakeet on Ctrl-C and name the repair when its install breaks
- Ctrl-C during a Parakeet transcribe now stops with exit 130 instead of
starting whisper, and a SIGTERM to the CLI also stops its decode child.
- Ctrl-C during models install parakeet aborts the download and the npm
install, keeps no partial file, and exits 130. A staging dir left by a
killed install is swept by the next run.
- When Parakeet fails under --engine auto, the warning leads with the
Parakeet error and names hyperframes models install parakeet. If the
fallback fails too, the Parakeet error is reported rather than
whisper_unavailable, so --optional no longer hides a broken install.
- An explicit --engine parakeet no longer falls back: it fails with the
Parakeet error and the repair command.
- The install refuses up front on systems with no native package
(naming the six supported ones) and on musl, instead of a raw npm error.
- 24-bit and float WAV go through ffmpeg, since sherpa-onnx reads only
16-bit PCM. The decode timeout scales with the audio length.
* docs(cli): point Parakeet setup at models install parakeet
The transcribe docs and the media-use references named parakeet-mlx as the
only way to get Parakeet. They now name hyperframes models install
parakeet, and the CLI reference gets a models section.
* test(cli): cover the npm cancel signal and decode timeout, and fix a doubled period
The Parakeet failure line printed "failed.. To repair" when the decoder's
own message ended in a period. Tests now also pin that Ctrl-C reaches the
npm install and that the decode timeout scales with the audio length.
* fix(cli): report unreadable audio as an input error, not a Parakeet failure
Prepare the WAV once before the Parakeet fallback and hand the same file to whisper,
so ffmpeg runs once. Ctrl-C while ffmpeg runs now ends as a cancel (exit 130)
instead of starting whisper.
* fix(cli): recover a clause Parakeet drops after a pause by decoding the gap alone
Each loud gap without tokens is decoded on its own with 0.5 s either side, and only
its new tokens are spliced in. The padded whole-window retry stays as the fallback.
A window with no tokens at all is no longer decoded twice.
* perf(cli): size the decode timeout from the WAV header instead of reading the file
* docs(cli): say a broken Parakeet install still fails transcribe --optional
* fix(cli): treat Ctrl-C during audio preparation as a cancel and drop the partial WAV
ffmpeg traps SIGINT and exits 255 rather than dying by the signal, so that status now
counts as a cancel unless the timeout fired. A failed or stopped ffmpeg removes the
temp WAV it was writing. The decode timeout now sizes a prepared WAV from its file
size, so a large header can no longer hide its length.
* fix(cli): splice a re-decoded gap by whole words so its neighbours are never doubled
The gap decode splits edge words into other tokens than the window decode did. Match
neighbours as whole words within 0.3 s of each gap edge, add only whole words that start
inside the gap, and decode the gap behind 0.5 s of silence like the padded retry.
* fix(cli): count a cancel only when ffmpeg says it received a signal
ffmpeg 7 and later also exit 255 when a file is not permitted (EPERM), so exit 255
alone read a permission error as a cancel. ffmpeg's stderr is now captured, a cancel
needs its received-signal line, and any other failure reports ffmpeg's last line.
The output path is passed to the shared runner explicitly.
* fix(cli): keep a spliced gap word that has no leading space
The filter dropped such a word; it is now kept and given its word boundary.
* refactor(cli): trim three long comments in the Parakeet decoder and installer
* fix(cli): repair a Parakeet runtime missing its native binary, and pick Parakeet by language
The runtime dir is per platform and arch. Installed now needs the runtime
manifest and the native sherpa-onnx.node; a half-installed dir is removed
before npm runs, since install() keeps any dir holding the runtime manifest.
The decode worker flattens a multi-line loader error onto its prefixed line
and stops between windows once orphaned. transcribe opens its cancellation
scope before audio prep, so Ctrl-C under --json cleans up. auto uses
Parakeet only for its 25 languages and only where the host can run it;
--engine parakeet with another language fails.
* fix(cli): sweep reused-pid staging, name failed model downloads, keep ffmpeg's cause
sweepStaleStaging also removes a staging dir older than 6 hours whose pid
is alive, since containers reuse pids. A failed model download names the
file, the host and a retry. runFfmpeg keeps ffmpeg's stderr on its error,
so transcribe shows "Output file does not contain any stream" for a video
without audio, and reads a cancel only from ffmpeg's own signal line.
* fix(skills): make media-use --engine whisper run whisper and report the engine that ran
The whisper path now passes --engine whisper when asked, or --engine auto,
and reads the engine from the CLI's --json line, so a CLI Parakeet run is
no longer labelled whisper. Docs state the language rule, the real repair
behaviour and the platforms Parakeet runs on.
* test(cli): pin that a long Parakeet decoder error is kept whole up to a bound
* fix(cli): report a broken Parakeet runtime with the repair instead of skipping it
* refactor(cli): name the staging age limit instead of commenting it
* fix(skills): let media-use --engine parakeet use the CLI Parakeet when parakeet-mlx is absent
* fix(cli): repair a Parakeet runtime that fails to load, and read Ctrl-C after whisper
models install parakeet now loads the runtime in a child process, the way
the decode worker does, and reinstalls it when that fails, so a missing
shared library, a corrupt binary or missing runtime JS is repaired. It
reports changed: false when the runtime loads. transcribe keeps its cheap
manifest check. After a failure, transcribe lets the event loop turn once
before reading the Ctrl-C scope, since a signal that landed during the
synchronous whisper fallback is only handled then.
* fix(cli): read Ctrl-C from the whisper child, and re-check the Parakeet runtime after a repair
A whisper fallback killed by SIGINT or SIGTERM is a cancel (exit 130), the
same rule decode and runFfmpeg use; the loop-turn wait is gone. The
runtime probe treats a child stopped by Ctrl-C as a cancel and keeps the
runtime; a reinstall is probed again and fails with the loader's error if
it still does not load. media-use keeps the CLI's last lines, stdout
included, when hyperframes transcribe fails.
* fix(cli): give the user-stop rule one owner, SIGHUP included, and end a cancelled install at 130
stoppedByCancelSignal in renderCancellation.ts uses the scope's own signal
list (SIGINT, SIGTERM, SIGHUP) and is false whenever Node stopped the child
(a code, killed, or a spawn error). decode, runFfmpeg, the runtime probe and
the whisper fallback all route through it, so a probe killed by SIGHUP keeps
the runtime. isWhisperTimeoutError now keys on ETIMEDOUT only, which Node
sets on execFileSync's own timeout, so a SIGTERM reaches the cancel rule.
A cancelled models install sets exit 130 and returns instead of failing, and
the probe says what went wrong (timed out, crashed, no output).
* refactor(cli): let the stop-rule helper name carry its meaning
* fix(cli): Studio's catalog install goes through add, so it records the item and honours paths
* fix(cli): Studio's catalog lists the project's registry and install paths survive symlinks
* refactor(cli): name the Studio catalog's registry option plainly
* feat(cli): registry_item_added says whether add or Studio installed the item
* fix(cli): dropping the same catalog block in Studio twice installs it again
* chore(cli): drop a comment that restates add's source option
* fix(cli): Studio's re-record finds lock entries by their resolved path
* test(cli): the symlinked-folder install test checks the resize happened
* fix(cli): Studio mounts the installed item's own file by the path add recorded
* fix(cli): Studio's install never hands back a file the project kept for its own edit
* docs(cli): explain how catalog search works and which model it downloads
* docs(cli): name the model catalog search was chosen against
* docs(cli): keep only live rules in the word-search header comment
* docs(cli): correct the hosted vector size in the local vector build comment
The catalog already loads every item's manifest, preview included, but the
--json rows dropped it, so an app listing the catalog had no way to show an
item before add downloads it. Rows now carry preview when the item has one.
* fix(cli): install onnxruntime-node and google-genai on first use
Both carry deprecated transitive dependencies (boolean, node-domexception)
that showed up as npm warnings on every npx hyperframes install. They move
to devDependencies (types and the Intel-binding check still need them at
build time; the bundler externalizes them) and load through a new
on-demand loader that installs into the CLI's own cache dir the first
time a feature needs them, with one stderr line, no prompt, and a
manual-install command named in the error if the network is unavailable.
hyperframes doctor lists each as installed (with version) or
not-installed-yet.
* fix(cli): sweep stale install staging and bound network retries
* fix(cli): optional-package install only sweeps staging dirs whose process is dead
Staging dirs now carry a per-call suffix after the pid, and the sweep skips any dir
whose pid is alive, so concurrent first runs no longer delete each other's install.
The install error suggests the network only for network failures, and the manifest
path has one owner. install() takes an injectable npm runner so tests drive the real code.
* fix(cli): optional-package sweep also clears old-format staging dirs and tests a live foreign dir
The concurrent-callers test shared one pid, so it never exercised the sweep rule; a live foreign
staging dir test now does. Old-format dirs of a dead pid are swept, and a failed cleanup no
longer aborts a viable install.
* fix(cli): optional-package sweep names a stale dir it could not remove
A locked leftover is still skipped so the install proceeds, but the warning makes it visible.
* docs(cli): shorten the install docblock to the comment limit
Comment-only change.
- HANDOFF section 4a: the single-worker fallback applies on default routing
(non-DE router and the manual opt-in); a DE-router plan keeps its
precomputed fallback. Adds the review's verdicts: three confirmed defects
fixed, one latent hardened, one split accepted, one low left open (a worker
wedged before its first capture is bounded only by player-ready and CDP
timeouts and is not retry-eligible).
- cli.mdx: the parallel-stream paragraph no longer promises a single-worker
retry for every case.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- tests/long-form/README: the two-clip in-flow fixture, the per-frame luma
check over clip B, the 2026-09-18 matrix (branch -w 1 / -w 2 disk / -w 2
streaming / absolute CSS / published 0.8.44 all 0 of 100 black; 0.8.35 and
0.8.36 99 of 100 as the positive control; fixed in 0.8.37 by #3893), and
why the multi-worker trace's captureMode was not the engine's mode.
- HANDOFF: R3 row now reads "hardening built"; new section 4a lists the
guards, the 694 gate result and what the flip still needs; counts refreshed.
- cli.mdx: the parallel-stream paragraph says the opt-in path now retries a
stalled pool or a mid-stream encoder exit on a fresh single-worker session.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Three workers produced a byte-identical file to the single-worker segmented
render at 1.6x the speed, which is the property the gate exists to check:
segments finish out of order, so a mismatch would mean the concat list had
picked up completion order instead of the plan.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Miguel Ángel