* fix(studio): releasing a clip drag outside the window cancels it
Pointer capture delivered the release to the timeline even outside the window,
so the move or trim committed wherever its preview was last drawn. A release
outside the window now cancels the gesture, the same as Escape.
* test(studio): pin a release exactly on the window's edge as outside
* fix(studio): releasing a fade-handle drag outside the window cancels it
Fade handles run their own gesture, so they now use the same outside-window
check as clip moves and trims, moved into one shared module.
* fix(runtime): a scene swap refuses when an animation outside the scene moves its elements
* test(runtime): pin the scene swap volume re-probe, host CSS variables and diagnostics-free swap
* fix(runtime): a scene swap keeps a video its scene script animates instead of reloading it
* fix(runtime): a scene swap stops every animation the old scene script started
* fix(runtime): record scene script animations only on pages a scene swap can act on
* docs(core): reword the scene animation recording comment
* fix(runtime): watch media as the page parses only on pages a scene swap can act on
* perf(runtime): a scene swap checks for outside tweens in one pass over the film's tweens
* test(runtime): pin the single revert and the cleared animation record of a scene swap
* fix(runtime): a scene swap rebuilds media when the edit changes the scene script
* fix(runtime): a scene swap undoes the sets a scene script made
* fix(runtime): a scene swap rechecks outside tweens and runs new scripts after every host
* test(runtime): pin that a tween started during the caption wait leaves the page untouched
* fix(runtime): a scene swap refuses when its script set a value on an element the swap keeps
* docs(core): say the scene recording drops a finished tween that the timeline would keep when paused
* fix(runtime): a scene swap refuses when two of its animations write to an element it keeps
* fix(runtime): a scene swap resets kept media and refuses animations outside the scene
* fix(runtime): a scene swap keeps a moved video in place and swaps tweens on plain objects
* chore(core): regenerate the inlined position-edits script for the kept-video move reapply
* fix(runtime): a scene swap checks the timeline registry as it is after the caption wait
* fix(runtime): a scene swap refuses when a scene changed during the wait or a revert re-registers
* fix(runtime): a scene swap stops every animation a revert chain registers and rechecks its hosts
* fix(runtime): a scene swap refuses a moved host or a cancelled swap and stops at a shortened end
* fix(runtime): a scene swap refuses when a scene leaves its place or the film during the wait
* fix(runtime): a scene swap checks only its host's place, so a later body append does not refuse it
* fix(runtime): a scene swap refuses when the film's scene layout changed during the wait
* fix(runtime): a scene swap refuses when the order around a scene changed during the wait
* test(runtime): pin that a scene swap checks the film root order when every scene is wrapped
* fix(runtime): a scene swap gives kept media the mute, rate and loop a fresh load gives it
* fix(runtime): a scene swap rebuilds media a script wrote attributes on, keeping the rest
* fix(runtime): a scene swap keeps media carrying the runtime classes or reordered attributes
* fix(runtime): a scene swap resets a kept video's volume as a load does, not through the volume probe
* perf(core): a bundled film inlines each font once instead of once per composition
* fix(core): font dedupe keeps the last copy, so the rule the browser uses never changes
* fix(core): font dedupe merges only exact copies, from styles that always apply
A style postcss cannot parse ships as authored instead of failing the bundle.
The key keeps !important and whitespace inside quoted family names.
Styles behind a media query, a non-CSS type, <noscript> or <svg> neither lose
a copy nor keep one.
* perf(core): font dedupe runs before inlining, so bundling is no slower
Deduping after inlining made postcss parse every inlined font: a three-scene
fixture went from about 40 ms to about 450 ms. The same url text inlines to the
same bytes, so deduping first gives the same output.
* fix(core): a scene-part bundle keeps each scene's own font copy, so a scene still swaps alone
* chore(core): shorten the font dedupe comment
* chore(core): trim the font dedupe comments to keep the file's comment share
* feat(studio): export snapping, the marquee and picker describe for a host's own preview
* fix(studio): a host pick type must bring its own marquee resolver
* fix(engine): forced-SDR renders show tone-mapped HDR footage as Studio does
Forced-SDR extraction tone-maps HDR frames with zscale, but the frames kept the
BT.709 transfer tag, so Chrome re-curved them (a tone-mapped 388d96 painted as
2d8790), and JPEG frames kept the BT.709 matrix every JPEG reader decodes as
BT.601. Tone-mapped frames now get the same sRGB tag and BT.601 JPEG conversion
as other SDR frames.
macOS used VideoToolbox, which does not tone-map: its frames stay PQ and Chrome
tone-maps them itself, darker than the hable curve Linux and Studio use. macOS
now takes the zscale path whenever ffmpeg has zscale, found through the shared
ffmpeg filter probe; without it the old path stays and the render warns once.
Co-authored-by: Kenneth Shortrede <95665480+sontiac@users.noreply.github.com>
* fix(engine): forced-SDR renders tone-map HDR footage that lacks matrix or primaries tags
zscale finds no path between colourspaces without them, so the extraction
failed. Untagged HDR footage is BT.2020 in practice, and now gets those tags.
Co-authored-by: Kenneth Shortrede <95665480+sontiac@users.noreply.github.com>
* fix(engine): forced-SDR tone map also tags the transfer ffprobe read from the container
From ffmpeg 8.1 ffprobe reports a container-only PQ or HLG tag; the frames
still lack it, so zscale failed where macOS VideoToolbox used to render.
Co-authored-by: Kenneth Shortrede <95665480+sontiac@users.noreply.github.com>
* test(engine): gate the container-only transfer case on the ffprobe that reads it
* docs(engine): shorten the filter-probe comment to the comment limit
* fix(engine): keep valid per-frame colour tags when tone-mapping HDR footage
* test(engine): build the mixed HLG and PQ fixture without B-frames so its timestamps are valid
* test(engine): build the trimmed mixed fixture without B-frames too
---------
Co-authored-by: Kenneth Shortrede <95665480+sontiac@users.noreply.github.com>
* fix(studio): dragging an audio clip's edge trims it instead of moving it
The clip picture layer (waveform, filmstrip, stills) paints above the trim
handles and swallowed the press, so an edge drag started a move. The
picture layer now takes no input.
* docs(studio): say clip content takes no pointer input
* fix(cli): an update never replaces the CLI's files while a hyperframes command is running
* fix(cli): a killed command's leftover pid file cannot hold updates back
* fix(cli): judge a stale pid file by the installer's own clock, so sleep cannot drop a live command
* fix(cli): after a sleep the installer watches every command's file afresh
* docs(cli): shorten the installer script comment
* docs(cli): keep the new files within their comment share
* refactor(cli): split the upgrade eligibility check out of the installer scheduler
* docs(cli): keep the scheduler's doc on the scheduler
* fix(studio-server): remove the history of a project that no longer exists
* fix(studio-server): keep a project's history while its drive is disconnected
* fix(studio-server): keep an older history whose drive may only be unmounted
* fix(producer): gif renders survive rgb frames among transparent ones
* test(producer): run the gif mixed-frame encode test in the ffmpeg lane
* fix(producer): gif encode test passes the argv, reachability and comment checks
* fix(producer): transparent gifs clear every frame so an opaque frame never lingers
* fix(producer): only gifs that turn translucent after an opaque frame get whole cleared frames
* fix(core): plain text keeps its own gif control block; gif encode failures share one shape
A canvas drag stores the element position as a global gsap.set above the
timeline declaration. A resize that followed it inserted its tl.set right
after that line, so the saved script referenced the timeline before it was
declared and every reload and render threw a ReferenceError.
Both GSAP writers now bound the "after the last statement" anchor by the
timeline declaration, whatever the timeline variable is named. A drag or a
resize on its own writes the same bytes as before.
* fix(lint): a missing image, audio or mask file names the composition that references it
* test(cli): a missing file is reported once per composition that references it
* fix(lint): Studio's lint panel and the CLI label a finding by the composition it names
* refactor(lint): drop a comment and unused first-entry lines from the grouping change
* test(producer): run the late-host pip and dogs-captions fixtures in CI
The pip fixture wrote its video start in root time without the global
basis marker that PR 3732 made the contract, so it rendered the empty
host. Mark it global and schedule both fixtures that never ran in CI.
* ci(producer): split regression fixtures across nine shards
dogs-captions runs about eight minutes on its own; a ninth shard keeps
the other shards near their previous wall time.
* ci(producer): run dogs-captions in a nightly regression job instead of on every PR
* ci(producer): base the dogs-captions timing and nightly reason on two CI runs
* feat(cli): hyperframes projects finds the HyperFrames projects on this machine
* refactor(core): keep the project rule's comment share under its package's
* fix(cli): projects keeps Spotlight hits under a symlinked root and refuses a missing root
* style(cli): format the projects walker and its tests
* fix(cli): projects tests pass under a symlinked temp folder and a bad root errors on stderr
* test(cli): projects tests hold on Windows and cover an unreadable root
* style(cli): format the projects command and tests
* refactor(cli): split the projects Spotlight check and regenerate the skills manifest
* refactor(core): findProjects moves to @hyperframes/core/projects so Desktop can import it
* chore(core): publish the projects subpath
* fix(core): findProjects fails on a root it cannot read instead of finding nothing
* docs(core): name the project marker writers without naming one app
* refactor(core): ship findProjects as a core export only, with no CLI command
* feat(core): findProjects stops on an abort signal and defaults to the home folder
* fix(core): findProjects checks an aborted signal first and stops when onProject throws
* fix(studio): read and write a clip's in-point the way playback does
* fix(studio): a group head trim keeps the minimum clip length off the grid
* fix(studio): a group head trim stops where a single-clip trim does and never goes below zero
* fix(studio): group members on the grid move together when a trim stops at its limit
* fix(studio): level-out measures and places its lift on the clip clock at any speed
* fix(studio): round level-out lane times and tighten the half-speed test
* fix(studio-server): ffmpeg and ffprobe run from the configured binary; unchecked uploads say so
* fix(studio-server): the upload result type lists unchecked files
* fix(media-use): spawn the configured ffmpeg and ffprobe, and name a missing configured ffprobe
* style: format the ffprobe path test, doctor spawns and upload check
* fix(studio): show the not-checked upload notice as info, not an error
* fix: a configured ffmpeg or ffprobe path that cannot run fails loudly
* fix: refuse a folder as the ffprobe path, and report broken paths through --json
* test: reuse the file manager test setup and split the upload stream check
* fix(media-use): a configured ffmpeg or ffprobe must run -version, and a relative one runs itself
* fix(media-use): a configured ffprobe path must name ffprobe, not another tool
* fix(studio): playback keys do nothing while a dialog is open
With a modal dialog open over Studio (the lint results, for example),
Space, the arrow keys, J/K/L, M, I/O and A/E still reached the player,
so the film behind the dialog played and seeked.
The playback shortcut filter now ignores keys while a visible modal
dialog is in the page: an aria-modal dialog or a native dialog opened
with showModal, wherever focus sits. Dialogs left mounted but hidden,
display:none or inert do not block keys.
* fix(studio): a dialog hidden behind fullscreen preview leaves playback keys working
* test(studio): the no-checkVisibility case runs without an earlier stub
* perf(core): playback housekeeping runs on a timer, and rebinds stop re-sampling silent media
While playing, the transport re-posted the clip manifest every 20 frames and
re-bound media every 30, frame-counted, so a 120 Hz display did both twice as
often for a film that was not changing. Playing, those jobs now run when the
composition changes (rate-limited as before) and on a one-second poll for
what nothing can push.
A rebind re-probed every media element's volume by seeking the whole timeline
60 times per second of clip, about 1,500 seeks and 50-140 ms in one frame.
The probe seeks with events suppressed, so only a tween on the element that
names volume can move it; an element no tween touches is now answered from
the tweens without seeking.
* fix(core): a data-width or data-height change applies while playing and parked
postTimeline applies the root's sizing, and it now runs when the composition
revision moves. The revision only watched timing attributes, so a size edit
waited for a pause while playing, and on main it was missed while parked
too. The observer now also watches data-width and data-height. The tick
comment is trimmed to three lines, the tick reuses its playing flag, and the
cadence test is renamed to what it checks now.
* test(core): share one 120 Hz frame step in the transport park tests
The two playing tests each defined the same frame helper; one frame120Hz
now sits beside settle and all four loops use it. No source change.
* refactor(core): drop the playing poll doc comments
The comment ratchet does not let a file's comment share rise; the reason
lives in the PR body.
* fix(core): a playing clip's track, label and z-index reach Studio again
The playing path had dropped the periodic manifest post, and the
composition observer does not see data-track-index, data-timeline-label or
an inline z-index on an existing clip (watching style would fire on every
GSAP frame). On the playing poll the runtime now builds the manifest and
posts it only when it differs from the last one posted, so an edit lands
within a second and a steady film still posts nothing.
* fix(core): keep sampling volume when a tween drives a non-DOM object
A tween on a plain-object proxy whose setter writes the element's volume
names neither the element nor volume, so the fast skip dropped it and
WebAudio-routed preview played the static data-volume (0 where the film
asks for 0.375). The skip now applies only when every tween targets DOM
nodes; any other target falls back to the full volume probe, as on main.
* fix(core): skip the volume probe again for spacers, calls and counters
The non-DOM fallback fired on any plain-object or function target, so a
tl.to({}, {duration}) spacer, a tl.call or a tweened counter brought back the
full probe on ordinary films. It now falls back only when a tweened property
resolves to a setter on the target or its prototype chain, which is what a
gain proxy or a class with a volume setter uses; data properties skip.
* fix(core): post a clip edit made while playing when playback stops
A label, track or inline z-index edited while playing and then paused before
the next one-second poll was never posted: once parked, nothing watches it.
The tick where playback stops now runs the same manifest diff once and posts
only if the manifest changed.
* 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
* fix(studio): the live clip list says whether each clip is heard and refreshes when that changes
* fix(studio): a muted clip reads as muted and a timed wrapper does not take its video's sound
* test(studio): a timed wrapper ignores the data-has-audio of the video inside it
* refactor(studio): keep the clip change check and audio reader under the complexity gate
* fix(studio): a fade-only change refreshes the live clip list
* perf(studio): a film with no manual edits skips the edit reapply on each seek
The reapply runs after every timeline seek, so on every playback frame, and
ran six document queries plus the motion rebuild even when Studio had never
edited the film. One combined query now answers "any edits?" first.
* test(studio): the seek skip still clears a motion timeline after an undo
An undo removes a motion edit's marks while its timeline stays registered,
and the reapply is what kills it. A test now fails if the skip stops
treating that timeline as an edit. The moved comments are cut to three
lines and lose an all-caps word.
* perf(studio): the per-seek edit check no longer walks large films or adds a lookup with edits
The combined comma selector took Chrome off its lone-attribute fast path: about 150 us per
seek on a 1.5k-element film, against 6 us for the whole reapply on main. The check now uses
one selector per mark and caches its answer per document, rescanning only after a
MutationObserver sees a mark attribute or the tree change.
* test(studio): the seek edit cache sees inserted edits and ignores style writes
Covers the two observer settings the cache depends on: a node inserted with its edit mark
after a seek is reapplied on the next seek, and a style write between seeks causes no rescan.
* 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
* perf(core): skip hidden clips' content so the browser stops decoding their images
Hidden timed clips get content-visibility:hidden while their inline visibility is hidden,
from script evaluation on. A clip due within the look-ahead (2 s by default) keeps its
content live, so its images load before it shows.
* perf(studio-server): load preview images lazily so hidden scenes fetch nothing
* fix(core): stop the readiness scan waiting on and decoding later scenes' images
* perf(studio-server): keep first-frame images eager so the preview opens as fast as before
Only images inside a clip with a numeric start after 0 go lazy; a start the route
cannot read stays eager.
* perf(core): keep clips that start at 0 laid out before the first pass so the first frame is on time
* perf(core): open the preview on time by hiding only lazy images before the first pass
GSAP reads the computed style of the elements it animates while timelines build. Inside a
content-skipped clip every read forces a style update, which cost about 65 ms at open on the
bench composition. Before the first pass the rule now takes only lazy images out of layout, so
hidden later scenes still do not load theirs; after it, the skip rule keeps their content skipped.
* fix(core): hold a paused jump on the previous picture until the next scene's images decode
* fix(studio): swap in a reloaded preview only after its restore seek has painted
* fix(core): apply a held paused jump after 1 s when an image never decodes
* fix(core): keep hidden-scene image skipping and held jumps to Studio's preview, never renders
* fix(studio): reapply manual edits once a held paused jump lands
* fix(studio): copy a clip without the preview's lazy-loading marks
* fix(studio-server): mark preview images lazy without re-encoding the document's entities
* fix(core): share one preview-mark constant between the runtime and the preview route
* fix(studio): reapply manual edits a second time only after a seek that was held
* refactor(core): share the hidden-ancestor walk between clip visibility and image holds
* fix(core): import the studio preview mark from its own subpath
* fix(studio-server): find preview images with the linear html scanner
* chore(core): drop two doc comments the scanner's names already carry
* fix(studio-server): serve the preview mark as its own meta, apart from the GSAP fallback script
* fix(studio-server): capture thumbnails and frames from the preview without lazy images or the mark
* fix(core): let the full readiness scan wait on every image except lazy ones off the first frame
* fix(studio-server): keep an authored loading attribute in any letter case on preview images
* fix(core): leave template content out of findStartTags so preview images match the DOM
* fix(studio): send Ask agent the live markup without the preview loading marks
* chore(core): drop the template note from findStartTags, its test carries it
* perf(studio-server): splice lazy attributes into the preview in one pass
* fix(studio): keep authored image loading on copy and Ask agent by marking preview-made lazy images
* test(studio-server): follow the capture preview URL back to its sub-composition for every project id
* refactor: name the upcoming-clip and GSAP fallback marks once
* fix(studio-server): keep capture builds out of the on-disk preview store
* feat(studio): let a host app mount the timeline toolbar, undo buttons, meters and clipboard
* test(studio): lock in history button rules and trim duplicate export tests
* fix(producer): renders use the font named in a var() fallback
When a font-family starts with var(--x, A, B) and --x is not defined, the
render now reads A, B the way the browser does. A system font at the head of
that fallback gets Inter placed in front of it, the same rule a direct
system-font stack already follows. The rewrite stays inside the var()
fallback, so a defined --x still wins. A defined property whose value is
itself a var() is followed too. The distributed plan check reads the same
answer.
Fonts named only in such a fallback are best-effort: they are embedded when
they resolve, and one that does not resolve is logged instead of failing a
distributed render. Fonts named directly, or through a defined property,
still fail closed as before.
CSS-wide keywords such as inherit are no longer treated as font names to
fetch.
Compiled the way a distributed render compiles them, 28 registry items now
embed a font they name this way, and none newly fails.
* refactor(producer): fold var() fallback resolution into one branch
* fix(producer): fonts behind a var() chain or empty fallback add no new render failures
A family reached through a chain of defined var()s is optional like a
fallback-only one: fetched fail-closed, skipped only when no source serves
it. An empty var() fallback makes the declaration invalid, so the browser
inherits and the plan check no longer rejects the families after it.
* test(producer): a family named after a leading var() stays required
* fix(producer): a transient error on an optional font warns and renders the fallback
* chore(catalog): record the font requests var() fallbacks now make
* fix(producer): required fonts fetch before optional ones can spend the time budget
* fix(shader-transitions): page-side shader renders keep scene layout and fonts and leave no ghost
* refactor(shader-transitions): one loop stages both scene copies
* fix(shader-transitions): stage scene copies inside copies of their ancestors
* refactor(shader-transitions): pin the staged root in one assignment
* fix(shader-transitions): keep ancestor copies out of scene layout in renders
* fix(shader-transitions): scene copies keep wrapper effects and live animation times
* fix(shader-transitions): hold copied animations with one lookup per scene and drop the no-op shift
* test(shader-transitions): pin the element and pseudo-element parts of the animation match
* docs(shader-transitions): keep the staging comments to two lines
* fix(studio): read and write a clip's in-point the way playback does
* fix(studio): a group head trim keeps the minimum clip length off the grid
* fix(studio): a group head trim stops where a single-clip trim does and never goes below zero
* fix(studio): group members on the grid move together when a trim stops at its limit
* fix(lint): a missing image, audio or mask file names the composition that references it
* test(cli): a missing file is reported once per composition that references it
* 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
* fix(layout): flag text a clipping card cuts off
A short card with overflow hidden cut the words off, and the check
dropped text_box_overflow because that parent clips.
The text element's own box does not clip, so clipped_text stayed quiet.
A deliberate mask still clears with data-layout-allow-overflow.
* fix(layout): ignore a text line parked fully outside its box
A reel row that never enters the window was failing with the cut-off card.
Only a line that still meets the box and sticks out of it is an overflow.
* fix(layout): report visible text outside non-clipping cards
* fix(layout): distinguish line boxes from parked font extents
* test(layout): type browser audit test interfaces
* fix(studio-server): background removal writes only where it checked, even through planted links
* test(cli): read the cutout through a no-follow handle instead of checking then reading
* fix(producer): embed the Google font the page linked
The weight-only download drops axes the page's link asked for.
That file is written after the link, so the film draws a wider cut than the preview.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(producer): keep every weight the page's font link asked for
Google answers a 400..900 link as font-weight: 400 900, and only 400 was kept.
Heavier text then waited on the network. Large families are limited to the
characters on the page, @import is followed, and a failed link is not replaced
by a font found on this machine.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(producer): preserve italic supplements for bundled fonts
* fix(producer): preserve active authored font resources
* style(producer): condense font fetch comments
* fix(producer): preserve authored oblique font faces
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(engine): renders show the video frame the preview shows below the source frame rate
* refactor(engine): one condition owns CFR sampling at the output rate
* fix(engine): webm and mkv clips keep every frame when rendered at their own rate
* fix(engine): only fine timebases get the half-tick sampling shift, so coarse ones stay exact
* docs(engine): the sampling comment states the half-tick bound it keeps
* fix(engine): move the extraction cache to v6 so main v5 entries are not reused
* docs(engine): keep extraction comments within their share
* test(producer): regenerate four 25-to-30 fps baselines for the exact frame pick
* test(engine): name the v6 schema prefix in its test
* perf(studio): playback no longer relayouts the timecode and playhead every frame
The timecode shows MM:SS but was rewritten on every playback frame, and the
timeline playhead moved by style.left. Both forced a layout and a repaint per
frame. The timecode is now written only when its text changes, and the
playhead moves by a composited transform.
* fix(studio): the moving playhead stays on whole device pixels
As a composited translate the 1px playhead line is anti-aliased, so at a
half-pixel offset it drew as two half-bright columns, where the old left
position snapped it to one. The transform now rounds to device pixels. The
helper moves to its own module as a named function with full parameter
names, which also keeps timelineLayout under the file-size limit.
* refactor(studio): drop the timecode write comment
The comment ratchet does not let a file's comment share rise; the reason
lives in the PR body.
* fix(studio): the playhead moves smoothly while playing and rests on its tick
Rounding the playhead to device pixels on every update made slow playback
stair-step: at fit zoom on a two-minute film, 120 frames of playback gave
22 distinct positions where main gives 120. While playing the transform now
keeps fractional pixels (120 of 120). At rest (paused, or a seek while
paused) it rounds to device pixels the same way layout rounds the ruler
ticks, so the resting line is one full column on its tick at 1x and 2x.
Pausing snaps the line even when the paused time equals the store's.
* 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(studio): quick timeline saves no longer refuse each other as outside changes
* fix(studio): canvas edits and block writes wait their turn in the file queue
* fix(studio): animated clip moves and the last direct writes wait their turn in the file queue
* feat(plugins): ship portable agent bundles and installation guides
* fix(plugins): align new manifests with current release
* test(plugins): split caption fixture setup and cleanup
* test(plugins): pass render fixture path as an argument
* feat(studio-server): a host can make Cmd+Z undo the newest change whoever made it
* fix(studio-server): Cmd+Z for everyone leaves an agent turn that is still open alone
* style(studio-server): format the undo scope check
* refactor(studio-server): the undo scope option names itself
* fix(studio-server): an open turn's parts are known by id, and a new change ends redo for everyone
* style(studio-server): format the open turn check
* fix(studio-server): an open turn's part is known before its entry is announced
* fix(studio-server): history saved by 0.8.78 opens with every entry, and an unreadable id is never overwritten
* test(studio-server): a malformed history id is refused and kept, not replaced
* refactor(studio-server): name the 0.8.78 record check instead of commenting it
* docs(studio-server): hosts keep project histories in historyCache, which reopens a replaced folder
* fix(studio-server): an unreadable 0.8.78 record keeps its history, and another path to the folder is not a copy
* style(studio-server): format the 0.8.78 history tests
* fix(studio-server): name the history id error, and a recorded path that cannot be read is not a copy
* style(studio-server): format the 0.8.78 history fixture
* perf(studio): the lint error badge stops pulsing after three beats
The badge pulsed forever, so a Studio with a lint error painted a frame on
every display refresh, idle or paused. It now pulses three times and holds.
* fix(studio): the lint badge counts findings, never files
The badge showed the manual result's finding count while its dialog was open
and the number of files with findings once it closed, while its screen-reader
label always said findings. The hook now owns one count, the findings in the
list the error pulse also reads, and the badge shows it in both states.
* fix(studio): the lint badge pulses again when its finding count changes
A CSS animation does not restart on a text change, so after its three
pulses a background lint that found more problems left the badge still.
The badge is keyed on its count, so a new count mounts a new element and
pulses three more times.
* refactor(studio): the lint badge names its pulse count, and App reads the hook as one object
The pulse count moves from a comment into a constant name, so the file's
comment share does not rise. App takes useLintModal as one object instead
of a nine-line destructure, which keeps App.tsx under 600 lines once
merged with main.
* fix(cli): a failed command sends one cli_error, naming its real cause
* chore(cli): drop a comment the parameter name already says
* fix(cli): a failure caught by its own command is still reported, and only once
* test(cli): normalize-audio test passes the flags it requires
* style(cli): format the failure tracking test
* fix(cli): one function unwraps and reports a command failure, once
* docs(cli): shorten the command failure reporter's comment
Miguel Ángel