* perf(engine,cli): a machine's GPU encoder probe runs once, not on every render
* refactor(engine): one shared probe per process, no answer kept on disk
* fix(cli): every managed browser closes within a second even if Chrome never exits
Commands called browser.close() directly, which waits for the Chrome process to exit without a limit. A headless shell that acknowledged Browser.close and then stayed alive left the CLI hanging after its work was done. The managed browser's close() is now the bounded close that kills a browser that has not exited.
* test(cli): browser stand-ins carry the close() a real browser has
Audio effects launched Chrome with its own two flags and no retry, so inside
an agent's macOS seatbelt Chrome died at startup with "Target closed" while
the CLI's own browser launch recovered. The sandbox launch policy now lives
in the engine; the engine's Chrome launch retries once in one process over a
pipe when the sandbox refuses it, and audio effects use the shared Chrome
flags in software GPU mode. The CLI imports the same policy from the engine.
* fix(cli): a browser's orphaned helpers no longer keep a finished snapshot from exiting
The CDP pipe sockets stayed open after Chrome exited whenever a helper process outlived it, so the CLI never reached natural exit. Destroy them when the browser process exits.
* refactor(cli): name the pipe close so launch.ts carries no new comment
* fix(cli): a browser stand-in without stdio no longer throws from the exit listener
* fix(cli): close every pipe of an exited browser, not only the CDP pair
* fix(cli): closing the managed Chrome also ends the helper processes it leaves behind
The close waited a second for Chrome to exit and then killed only Chrome's own process.
Under load that second is often not enough, and the helpers (the GPU and renderer
processes, crashpad_handler) were reparented to init and kept running after a finished
run. The close now records the processes under Chrome before it starts, and ends any of
them still running (under the same command) once Chrome is gone.
* test(cli): the helper-reaping test polls for the helper's exit instead of waiting 200 ms
* fix(cli): the Chrome process listing is bounded and no longer runs at import
The ps call that records Chrome's helpers ran with no time limit before the
bounded browser close, so a stalled ps made the close unbounded. It now has a
2 s limit, and a timeout records no helpers.
execFile is resolved when ps runs, not when the module loads, so tests that
mock node:child_process without execFile import it again.
* fix(cli): a Chrome that fails to start says how it ended, and a slow start says so
A failed launch only reported "Protocol error (Target.setDiscoverTargets): Target closed". It now
runs the browser once more on its own with the same flags and reports its exit code (Windows
codes named: missing runtime, out of memory, security software), its last stderr lines, or that
it stayed up. The CLI and the engine's render launch both use it. A start slower than 5 s prints
"Chrome is still starting...", and a --timeout that looks like seconds says it is milliseconds.
* chore(cli): cut the Chrome launch diagnosis comments to the comment limit
* fix(cli): the Chrome startup check runs Chrome as puppeteer does and can be cancelled
The check started the browser with only the caller's flags, so a full Chrome ran headful and
was blamed for a missing X server. It now uses puppeteer's default arguments (headless
included). The check is stopped when the CLI is stopped, and the GPU probe, whose failure
falls back to software anyway, skips it.
* fix(cli): a Chrome launch error is never replaced by a failure to explain it
The startup diagnosis asks puppeteer for its default arguments before re-running
Chrome. If that call fails, its error used to replace the launch error the caller
needed. The launch error is now reported as it was. The core package size budget
rises to cover the new chromeStartup module (about 4.7 KB unpacked).
* refactor(engine): the Chrome startup diagnosis lives in the engine, not core
Both callers launch Chrome through the engine (the CLI already imports it), so the
engine owns the diagnosis. Core's package size budget is back to its old value.
* fix(cli): check and snapshot fail fast on a missing browser path and say while Chrome downloads
* fix(cli): a non-interactive Chrome download gives up after one stall, naming the env var and doctor
* fix(cli): the Windows Chrome check launches Chrome instead of trusting chrome.exe --version
* refactor(cli): shorten the Windows Chrome version comment to fit the comment ratchet
* test(cli): pin the platform in the Unix Chrome preflight tests so Windows runs them as written
* fix(cli): the Windows Chrome check reads the file version, not chrome.exe --version
* fix(cli): drop added comments, the names carry the meaning
* fix(cli): a check that is stopped by a signal prints the report it has so far
Killed by SIGTERM or SIGINT, check closed Chrome and exited 143 or 130 with
nothing on stdout, so a caller that stops a long check got no result at all.
The signal now ends the check the way an expired deadline does: findings
already made are kept, one finding says the check stopped before finishing,
and the report is printed before the process exits with the signal's code.
* fix(cli): reap only orphaned browsers HyperFrames launched
Orphan cleanup on preview and render killed every PPID-1 process of the
current user whose args named chrome-headless-shell or a Puppeteer temp
profile, which included detached browsers started by other tools.
Every HyperFrames Chrome launch now passes an inert
--hyperframes-owned-browser switch, and the reaper only kills PPID-1
processes that carry it as a whole argument.
Fixes#6203
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(core): serve the browser marker from its own subpath
The GCP image's BeginFrame contract step imports the engine's
browserManager straight from source before core's generated runtime
exists, so pulling the marker from the core root entry broke it.
A dependency-free @hyperframes/core/browser-marker subpath keeps that
import chain as light as it was.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test(cli): expect the browser marker in the capture fallback launches
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* docs(cli): say orphan cleanup runs before preview and render
The cleanup only stops browsers HyperFrames launched, so the opt-out is
no longer needed to protect other tools' browsers.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): a Chrome download that stops sending is abandoned, retried, and releases the lock
* fix(cli): a stalled Chrome download is aborted at the socket so the process can exit
* fix(cli): the Chrome download runs in its own process and is killed when it stalls
* test(cli): ignore the expected socket reset in the download stall test on Windows
* refactor(skills): fold six domain skills into hyperframes and media-use guides
hyperframes-core, -animation, -keyframes, -creative and -studio move to
skills/hyperframes/{core,animation,keyframes,creative,studio}/, and
hyperframes-audio moves to skills/media-use/mix/. Each moved SKILL.md becomes
a frontmatter-free GUIDE.md so installers that scan recursively for SKILL.md
no longer see them as skills; the published list goes from 21 to 15.
Relative links, code paths and skill-name mentions are rewritten to the new
locations, the catalog surfaces, marketplace core-skills list, fallback core
set, module-copy targets and skills-manifest.json follow, and
`hyperframes skills update` now removes the retired skills (RETIRED_SKILLS)
when they are still installed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills): slim the core entries and give media-use a HeyGen and a free path
The four entries every run opens are rewritten to what an agent cannot infer:
hyperframes 21.8 KB to 7.7 KB, hyperframes-cli 24.8 KB to 6.9 KB,
hyperframes-registry 9.9 KB to 4.3 KB, media-use 9.7 KB to 4.9 KB. Detail moved
whole into references (routing close calls and creator edits, CLI pin upgrades,
catalog search tiers, CLI conventions and feedback reports).
- hyperframes: state table, routing, the render contract, a terminal work loop
(check, snapshot, preview before render), the desktop-app hand-off, and a
table of the folded guides.
- media-use: one table of HeyGen versus free local tools per need; HeyGen
sign-in, credits and the CLI's video-agent, avatar, clone, lipsync, translate
and clipping commands; every free path works with no account.
- scripts/check-core-skills.mjs holds entries and descriptions to budget and
checks every taught `npx hyperframes` command and path; runs in lint.
- Core entries are exempt from the formatter: its table padding cost ~1.5 KB
per entry on every run.
- Content tests follow the text to its new home.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): plugin runs skip skills update; media-use explains HeyGen sign-in and TTS setup
The smoke eval showed a presenter request never reaching media-use and the
free voiceover failing on a missing kokoro-onnx; plugin runs also ran the
skills update the plugin rules forbid.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore(skills): drop a committed Python bytecode file and ignore bytecode
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills): align the entry with the Studio skills and keep the workflows' measured facts
The entry drops routing and the intent interview and carries the contract, the
editable-project rules and the work loop. Studio's format, runtimes, motion,
look and captions references come over for terminal use, frame presets move to
assets/, media-use gains Studio's mixing and local-command references, and
film.md and overlays.md keep what the workflow skills measured.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills)!: retire the workflow skills and folded guides; publish seven knowledge skills
The skill pack now publishes exactly seven skills - hyperframes,
hyperframes-cli, hyperframes-registry, media-use, figma, slideshow and
remotion-to-hyperframes - carrying knowledge, not process. The eight
workflow skills, the folded domain guides and the routing/interview
references are deleted; their measured facts already live in the
hyperframes references (format, runtimes, motion, look, captions, film,
overlays) and media-use's mixing reference.
- bake-basemap.mjs and locate.mjs move to skills/hyperframes/scripts/
(bake-basemap gains --help); extract-audio-data.py moves to
media-use/audio/scripts/ for the caption motion reference.
- skills update: the eight workflows join RETIRED_SKILLS so existing
installs prune them (by lock attribution only, since their names are
generic); naming a retired skill reports where its knowledge went and
succeeds instead of failing as unknown.
- Capture prompts point at the hyperframes skill's references/film.md;
the live-map warning names the new bake-basemap path.
- Content tests that pinned knowledge are retargeted to the kept
references; tests of deleted workflows, scripts and process go.
- Catalog surfaces, templates and docs describe the seven skills.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): resolve the local.md captions link and allow capture's output files in citations
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): regenerate the skills manifest from tracked files only
The committed media-use hash included an ignored __pycache__ bytecode file present
in the local checkout, so CI's clean checkout hashed differently.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(docs): restore generated catalog pages and docs navigation to main
Catalog PRs commit item sources only; CI regenerates the pages from the edited
registry sources. docs/docs.json is catalog navigation, so its Workflows group
rename is left for a non-catalog PR.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* build(cli): raise the CLI size budget for the skill references it bundles
The CLI bundles skills/hyperframes, which now carries the folded domain guides and
frame presets (155 KB to 1.08 MB in git). Budgets are the CI-measured size plus 5%:
packed 7526634, unpacked 22593213.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test: keep repo-sweep guards meaningful after the workflow skills left
The last shell ffprobe caller and most root index.html examples lived in retired
skills. The shell detector is now pinned on a literal command line, and the
structure-rule sweep needs more than 5 roots (11 remain) rather than 30.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills): slim figma, slideshow and remotion to the Studio shape; restore the storyboard board
slideshow 33.5 KB to 4.9, figma 18.2 to 6.2, remotion-to-hyperframes 11.0 to
4.9, each keeping the facts only it holds (the figma motion helpers, verify gate,
storyboard decoding and usage beacons; present and hand-off; the Remotion lint,
SSIM check and API map). The board page format from the storyboard recipe comes
back as hyperframes/references/storyboard.md: the desktop app's Storyboard tab
reads boards/storyboard.html by its frame, title and time classes. The entry and
the init template now say to snapshot the moments an edit changed before
replying, which the terminal eval found skipped in 22 of 38 runs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): manifest without local bytecode; Studio's edit prompt names kept files
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): render when the person asks for the file or for no stops, not only after approval
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): keep the voiceover carve in media-use; presets stage their own fonts
Review follow-ups on #5651:
- carve.mjs and its tests move to skills/media-use/audio/scripts; the audio
docs, mixing.md and two Studio comments point there.
- code-editorial's FRAME.md says to copy its bundled WOFF2s and add the
@font-face blocks, instead of promising the retired build-frame.mjs.
- media-use states the host rule before the HeyGen sign-in steps.
- storyboard-format marks transition_in as a note the agent builds itself.
- Two stale citations (creator-editing-recipes.md, sub-compositions.md).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore(skills): regenerate the manifest without local bytecode
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): the CLI render step follows step 4's rule, not approval alone
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(ci): run the core-skills check from lint again after the main merge
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills): films that tell a story start from three storylines; maps use real imagery
Opus 5.5 arm D read motion.md and applied it, but its launch film was a
feature list (pain, three feature cards, proof, logo): film.md left story
to the model. And the map case never opened film.md, so it drew Natural
Earth outlines instead of baking satellite tiles.
- references/story.md: the generic half of Studio's moment-film story and
craft (one thing, three storylines with world and device, the hook,
beats, worlds, copy, brand, tells of a generated film). The person picks
a storyline, or the agent takes its recommendation when asked for the
file or for no stops.
- The entry's work loop names the storyline step for films over ~15 s; the
references table routes stories to story.md and maps to film.md.
- film.md points at story.md and gives maps and real places a section.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): story.md follows the entry: no voiceover unless asked
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): bake-basemap starts from a globe and stops waiting out its idle timeout
The Opus 5.5 map run spent 13 of its 38 minutes on parallel bakes that
stalled and 15 more recording 300 frames, after rewriting the script because
it could only start on a flat map.
- PROJECTION=globe with a seeded starfield and atmosphere, and FROM/PAN to
turn the globe to CENTER before the dive.
- Each frame polls for loaded tiles instead of waiting for an "idle" event
that had often already fired: 30 frames went from 198-219 s to 10-13 s
locally with byte-identical frames.
- Progress prints seconds a frame and minutes left; film.md shows the globe
call and says to run one bake at a time.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): render in the foreground; wait on a detached render's log, never pgrep -f
Both Opus 5.5 explainer runs started a detached render, then waited with
`until ! pgrep -f "hyperframes render"`, which matches its own loop: each
render took 88-117 s and the wait sat for the tool's full 10 minutes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills): route each kind of film; storylines carry look and sound; add review and sound
The entry routes by request (launch, explainer, music, data, vertical, 3D,
character) to a self-contained file under references/types/, and a film
over ~15 s now gets music and sounds on its visible actions unless the
person asked for silence. Each storyline names its look and its sound, so
picking one picks both; STORYBOARD.md frames carry `reads` and `sound`.
review.md is the whole-film pass: the check warnings, the review sheets,
seven questions to look for, P0/P1/P2 notes and REVIEW.md. motion.md gains
eases by role, six title entrances, a reading-time rule and nine measured
transitions; look.md a design plan before the first scene, with presets
kept for decks and named picks. story.md adds the second default that
generated design falls into, with what to do instead.
The blank template and format.md no longer teach the fade-and-slide-up
entrance on a dark canvas.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(cli): check warns on unread lines, dead stretches, small type, empty frame 0 and silence
A coding agent cannot watch its film, so check now measures what a viewer
notices over time. A new `craft` section holds warnings only (exit 0
unless --strict), sampled on the dense 8 fps grid the overlap pass
already seeks, skipped for --range checks:
- line_unread: a reading-size line (>= 30 px tall at 1080) on screen
under 0.4 s for 1-2 words, else under 3/4 of 0.5 s + 0.3 s a word;
the film's last line needs 2 s. Short lines are re-measured between
samples to ~0.01 s before judging. Captions-track and
data-layout-ignore text are exempt.
- dead_stretch: more than 3 s where the keepsMoving liveness signature
does not change, outside the final 3 s and any data-hold element;
quiet when sweep_static fires or a whole-composition keepsMoving
assertion owns the limit.
- text_below_floor: text under 2.6% of the short side (28 px on 1080)
held 0.5 s, one finding per file listing every element.
- first_frame_empty: frame 0 paints only the background.
- film_silent: over 15 s with no unmuted audio or audible video.
Each message names the fix and the skill reference to read. keepsMoving
now also leaves the film's last 3 s (a still end card) out of its
static runs, so the two agree.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(cli): snapshot --review captures a beat sheet, seam strips and a phone sheet
snapshot --review picks the frames that show a film over time instead of
evenly spaced ones, in the same single browser session:
- review-beats.jpg: one settled frame per scene at its start + 0.3 s,
plus one per timed data-beat="name" element, labelled time and id.
Scenes are the composition's top-level timed clips and sub-composition
hosts, read like `timeline` reads rows (captions and audio excluded)
and timed by the runtime's start resolver over the live page.
- review-seam-NN.jpg: frames at -0.2, -0.1, 0, +0.1, +0.2 s around
every cut (each scene start, and any scene end no scene starts at).
- review-phone.jpg: the beat frames at 360 px wide.
--composition scopes the review to one sub-composition; with fewer than
two scenes the beats spread evenly and a note says so. snapshot also
gains --json: one envelope with every frame's path and time and, under
--review, the scenes, beats and seams behind the sheets; progress moves
to stderr.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): craft warnings stay out of --strict; a hold gets a real next action, not a caret
Craft findings (unread lines, dead stretches, small type, an empty first
frame, silence) are judgment calls a film may break on purpose, so
--strict keeps gating defects only. The check reference no longer
suggests a blinking caret to keep a still frame moving.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): review spreads beats when every scene starts together; map credits skip the type floor
A map film whose basemap and overlay both start at 0 s got one beat
frame; the even spread now keys on distinct scene starts. film.md tells
agents to mark the imagery credit data-layout-ignore and back it, since
the new type floor flagged it on the first real run.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): review frames each scene at its middle too, not only 0.3 s in
On the first real explainer, 0.3 s into a scene showed a bare colour
field before its title entered. Each scene longer than about a second
and a half now adds a frame at its middle, so the sheet shows what the
scene builds.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills): a film is built from a numeric SPEC.md, with defaults banned by name and a look test
Studied in full: flxrnc's Motion Lab (vague vs directed prompt), 0xMovez's
course, EianLu's riso skill, agentchengfeng's style samples and the two
Opus 5.5 guides. Their common point: adjectives are where the default look
lives, so the agent now writes SPEC.md before any scene (GOAL, OUTPUT,
TIME, LOOK, MOTION, SHOTS, BANNED, QA), with a number for every adjective.
spec.md bans the defaults by name, unless the brand or reference uses
them (a dark void with small things floating, glow and particles, the
cream-serif-mono-label look, fade-and-slide-up, corner brackets), and
grows that list after the first render. Each storyline gets a 3 s test of
its hero shot before the pick; a look is written as the rules of one
medium. Composition numbers: the hero fills 40% of the frame, one hero
move per shot, the shot's information in within 2.5 s then a held poster.
motion.md adds springs by damping ratio, shorter exits, overlap, stagger
and the beat grid; review.md reviews as a harsh director and grows BANNED;
vertical.md takes the riso skill's measured safe band; explainer.md the
shapes of information.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): story before look; every object reads as itself; posters only for info cards
The first spec-driven C02 swapped the dark default for a flat Swiss
poster, and the story went with it: two black bars nobody read as mirrors,
slogans instead of cause and effect, a storyline picked for its picture.
The spec now serves the story: a storyline is picked for the story a
stranger would follow, the look test asks first whether a stranger
understands what they see, and SPEC.md gains an OBJECTS line giving each
thing the material and details that make it read without a label. The
2.5 s held-poster rule is limited to information cards; in a story film
every shot causes the next. Explainers start from a person or situation
the viewer knows. Review names every object as a stranger would and
treats one that needs explaining as a P0.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): storylines are three films in words, each with a device and a form
The three storylines are pitched as text and the person picks from them;
nothing is built for a storyline before the pick, so the 3 s hero-shot
look test is gone. Ideas pair a device with a form, five or more first,
keeping the three that differ most, and the one stop states the rest in
an assumption line that names every paid item beside its free option.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* ci: list main's new embedded-captions hf-cli test among the retired workflow files
The merge with main removed it with the rest of the retired skill, and the
deletion guard rejects any removal it has not been told about.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(core): keep the audio preset comment within core's size budget
The comment's new citation to media-use's mixing reference made the published
package 14 bytes larger than its budget, which sits at main's exact size.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): craft keeps its own 8 fps grid; comments back within the checks
Review on #5651 found that the merge moved craft onto the overlap pass's
times, which main caps at 180, so a brief line in a long film fell between
samples and line_unread missed it. Craft now builds its own grid at 8 fps up to
600 samples, as before the merge, and a 60 s film with a line shown for 0.2 s
pins it.
The Comments check passes again: REVIEW.md and STORYLINES.md join the files a
skill writes into a project, the comment lines this branch added come back
under each file's share at the merge-base, and the audio preset comment drops
a word the check reads as history.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): a disk that fills mid-unzip fails the browser install at once instead of hanging
* test(cli): the unzip tests hold on Windows, and a cancel stops unzip
* fix(cli): a first render on a nearly full disk refuses at once instead of hanging
* fix(cli): name the low-disk guards instead of commenting them
* fix(cli): a killed install's leftovers never count against the free-space check
* test(cli): the render command's mocks carry the disk check it now runs first
* fix(engine): renders use the browser the CLI picked; Linux system Chrome is not beginframe
* fix(cli): state the browser env rule without its history
* fix(engine): renders from the CLI launch the browser it picked, not the engine config's
After SIGKILL the test slept a fixed 5 s, then required every Chrome process
gone. On a loaded Windows runner the watchdog took longer and two processes
were still exiting. It now waits up to 30 s for none to be alive, so a real
orphan still fails.
* fix(cli): snapshots, compare and check wait for late images the way render does
* fix(cli): check runs its browser pass inside a macOS sandbox that does not name itself
Every CLI Chrome launch has used a pipe since the warm-check change. A pipe
launch that a seatbelt kills fails with only "Protocol error
(Target.setDiscoverTargets): Target closed". The one-process fallback keyed
on Chrome's bootstrap_check_in stderr, which a pipe launch failure never
carries, so it never fired. On macOS a pipe launch that fails at the
protocol level now relaunches as one process over the pipe.
* fix(cli): retry only a pipe launch whose connection closed, and keep puppeteer lazy
* fix(cli): the browser the CLI launches exits with it, even on kill -9
* fix(build): the root build runs on Windows
* perf(cli): a repeat check runs in a process that already has the CLI and Chrome up
* fix(cli): a warm check delivers all its output before it exits
* fix(cli): the warm check only talks to a server that proves it is the user's
The warm socket now lives in a private, owner-verified directory (XDG_RUNTIME_DIR,
~/.cache or LOCALAPPDATA) next to a 0600 token. Client and server prove the token to
each other over fresh nonces before any request is sent, the request carries an
HMAC of the environment instead of the environment itself, and frames are capped.
* fix(cli): a warm check prints its stdout even after something went to stderr
The client stopped writing frames once any frame had been written, so a check that
printed a notice to stderr first lost its JSON report.
* test(cli): list the warm-check client fixture as a fallow entry instead of an inline ignore
* fix(cli): a check started while a warm one is busy runs cold at once on Windows
The warm server stopped listening when it accepted a client. On Windows the pipe name stays alive through the served connection, so a second client's connect saw a busy pipe and waited 30 s for a free instance before running cold. It now keeps listening while serving (accept already declines anyone else) and stops when the served check finishes.
* test(cli): the orphan test reads the check's own Chrome and watchdog instead of racing a poll
On Windows each process listing is a PowerShell call of about a second, so a check that finishes quickly could start and close Chrome between two polls, and the test failed its own 'Chrome was seen' precondition. A fixture now records the Chrome tree and the watchdog from inside the check when its first page opens.
* fix(cli): check and snapshot use the pinned Chrome, not an older one another tool cached
* fix(cli): keep the browser-order comment within the line limit
* fix(engine): renders that resolve their own browser follow the same pinned-first rule
* fix(engine): keep the browser rule's reasoning in the PR, not new comments
* fix: a shared-cache shell is used only if it starts, after the pinned build
* fix(engine): import the shared-shell pick, and probe shells with a blank page
* fix(skills): caption scripts launch the browser the CLI chooses
* chore(skills): regenerate the skills manifest for the caption scripts
* perf(cli): check reads a solid text backdrop from styles instead of a screenshot
* perf(cli): check starts Chrome while lint and bundling run
* perf(cli): the CLI reuses compiled bytecode between runs
* test(cli): browser shutdown tests stop their Chrome when a case fails
The seatbelt launch runs Chrome as one process, where there is no GPU process
to composite. Capture's flags left GPU compositing on, so every screenshot
failed and heavier pages stalled until the protocol timeout. The one-process
launch now also disables GPU compositing, as software capture already does.
Inside Codex's seatbelt the GPU probe's Metal Chrome is always killed, and
when it dies after puppeteer's setAutoAttach is answered, puppeteer waits
forever for its initial targets. The probe now answers software in a named
seatbelt without starting Chrome, and every CLI browser launch is bounded
by its timeout, so a stuck launch rejects and a signal can end the command.
* chore(docker): pin FFmpeg 8.0.1 in the Docker images and host CI
Install one pinned static BtbN build (n8.0.1-66, month-end release) in
Dockerfile.test, the CLI render image, the Cloud Run image and the k8s
example, and bump the host CI install-ffmpeg-linux action from 7.1.3 to
the same build. Regression baselines and the shipped images now use the
same FFmpeg version instead of Debian's apt ffmpeg 5.1.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test(producer): regenerate six regression baselines on FFmpeg 8.0.1
The Docker test image now renders with FFmpeg 8.0.1 instead of Debian's
5.1. Six fixtures fail against the 5.1 baselines; everything else passes
unchanged, including vfr-screen-recording.
- style-2-prod, style-10-prod: FFmpeg 8 honors the MP3 gapless header,
so the probed SFX duration is 1.0s instead of 1.044898s. Regenerate
output.mp4 and compiled.html.
- vignelli-stacking: one checkpoint (8.06s) now shows source frame 201
of the 25fps a-roll, which is the right frame for that time; the old
baseline showed frame 202.
- missing-host-comp-id, variables-prod, speed-ramp: AAC encoder output
differs enough to drop the RMS-envelope correlation on flat signals
below 0.9. Sample correlation against the source stays at 0.9999.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): read the ffmpeg version from n-prefixed release builds
Release-branch ffmpeg builds, including the BtbN build now pinned in CI,
print "ffmpeg version n8.0.1-66-...". extractMajorVersion skipped the
"n8" token and returned 0 (and 1 for n7.1.3), so the CLI telemetry
smoke check failed. parseToolVersion fell back to the whole banner line.
Accept an optional "n" before the version number in both.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): keep the preflight comment share flat
Fold the n-prefix note into the existing doc line instead of adding new
comment lines, which the comment ratchet rejects.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(docker): pick the FFmpeg build per arch and fail on a bad binary
Dockerfile.test, the k8s example and the Cloud Run image always extracted the
amd64 archive on multiarch bases, and the piped version check passed even when
the binary could not run. All four images now pick linux64 or linuxarm64 by
TARGETARCH, fail on any other arch, and grep both ffmpeg and ffprobe for the
pinned version.
style-2-prod and style-10-prod compiled.html go back to the previous golden
with only the SFX data-duration/data-end changed, instead of a full
recompile that also picked up unrelated compiler output.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): bound checks of non-finishing compositions
* fix(cli): account for cold startup in deadline witness
* fix(cli): wait for profile owners before test cleanup
* test(cli): separate portable deadline and Linux process witnesses
* test(cli): preserve per-launch browser profile ownership
* fix(cli): start check deadline after browser readiness
* fix(cli): the browser download unpacks on machines without an unzip command
* fix(cli): a corrupt browser download on the yauzl path is retried and named
* revert(cli): drop the cause-chain reader; @puppeteer/browsers never passes the cause through
* fix(cli): snapshot, check, layout and thumbnails honour software WebGPU like render
* test(cli): stub the local WebGPU helper in the studio server tests
* fix(engine): webgpu compositions are recognised after the bundled runtime script
* fix(engine): a hung GPU probe falls back to software in 15 s instead of 120 s
* fix(engine): the GPU probe keeps the configured launch timeout and bounds only its page
* build(parsers): declare the html-attribute-spans subpath in its descriptor
* build(gcp-cloud-run): the BeginFrame contract stage copies parsers, which the engine now imports
* test(cli): the screenshot launch pin follows the shared local WebGPU helper
* fix(engine): the GPU probe's wait for Chrome's first page falls inside its 15 s limit
* fix(cli): name the unzip errors when every Windows unzip tool fails
* fix(cli): a cached empty or folder Chrome counts as missing, not found
* fix(engine): skip an empty cached Chrome the same way the CLI does
* fix(engine): keep scanning the Chrome cache past a build it cannot read
* test(engine): a folder at the cached Chrome path is skipped
* fix(cli): browser ensure fails when the unzip leaves no Chrome executable
* refactor(cli): keep install debug output and report only the last unzip attempt
* fix(cli): browser ensure exits after a failure or a --force download
* fix(cli): browser checks name why Chrome could not be installed
* fix(cli): render shows why its setup step failed instead of a crash dump
* refactor(cli): build the browser download label and staged test path once
* refactor(cli): one owner for the render setup result line
* fix(cli): render setup failures report through one error line
* refactor(cli): one reader and writer for both render setup lines
* fix(cli): close every browser on SIGTERM and SIGINT through one launch owner
* fix(cli): own SIGTERM before the GPU probe browser launches
* fix(cli): keep tests on the managed GPU mode and drop the engine signal flag
* test(cli): give the browser fallback fake a process handle
* test(cli): signal the GPU probe case only once the signal owner is installed
* test(cli): mark the launching case only after the signal owner is installed
* fix(cli): stop browser --force from deleting chrome under running renders
* style(cli): format browser manager test
* fix(cli): sweep leftovers of killed browser installs and on browser clear
* docs(cli): state what browser clear returns without history wording
* fix(cli): route FFmpeg exec-failure hint through getFFmpegInstallHint()
checkFFmpeg()'s "cannot start" branch (binary found but fails to run
-version) hardcoded a Windows/DLL-flavored install hint on every
platform instead of delegating to the already platform-aware
getFFmpegInstallHint() used by every other branch in this file.
* test(cli): pin the ffmpeg cannot-start hint through the async probe mock
* fix(cli): keep the runtime-DLL cannot-start hint on Windows
---------
Co-authored-by: Miguel Angel Simon Sierra <miguel.sierra@heygen.com>
* fix(cli): a Studio render refuses with the browser check's own message
* test(cli): the Studio browser refusal test no longer loads the real producer
* fix(capture): never silently render a webgpu composition blank
A composition declaring data-requires-webgpu only got the Chrome WebGPU
flag when hardware GPU auto-detection succeeded. On a software-GPU host
it launched without the flag and rendered blank with no error.
compositionRequiresWebGpu (static declaration check) and
assertWebGpuAdapterAvailable (live per-page adapter probe) now live once
in packages/engine's browserManager, the shared choke point every launch
site already goes through. All six launch sites (frame capture, the four
CLI diagnostic commands, the studio thumbnail server, and the render
pipeline's capture-options builder) route through those two functions;
the old CLI-only static-heuristic guard is deleted, not kept alongside
the new one. A software host that cannot obtain any WebGPU adapter now
fails loudly naming the requirement instead of producing a blank frame.
* fix(capture): thread requiresWebGpu through the distributed chunk-render path
The distributed renderChunk() path built its CaptureOptions without
requiresWebGpu, so it fell through to frameCapture's own fetch-based
detection, whose failure path silently treats "requires webgpu" as
"doesn't" if that fetch ever fails. Every other launch site (frame
capture, all four CLI diagnostic commands, the studio thumbnail server,
and the in-process render pipeline) already reads the value from HTML
already in memory or on disk instead of relying on that fallback.
Reads the compiled composition's index.html the same way probeStage
already does and derives requiresWebGpu from it before building
CaptureOptions. Adds a regression test that pins the exact path
renderChunk() reads at render time, since the composition doesn't
actually exercise WebGPU so an end-to-end render can't distinguish the
two cases on this host.
* test(producer): observe the real CaptureOptions renderChunk() builds
The previous regression test called plan() and compositionRequiresWebGpu()
directly, bypassing renderChunk() entirely, so it could not detect a
regression to the wiring it was meant to protect (confirmed: it stayed
green with the fix reverted).
Exports the existing createCaptureSession dependency seam so a test can
substitute a spy that delegates to the real implementation while
capturing the CaptureOptions object renderChunk() actually constructs.
Verified non-vacuous by reverting the fix and confirming the test fails.
* fix(gcp-cloud-run): resolve puppeteer-core from a real file in the webgpu probe
The image build's WebGPU adapter probe imported puppeteer-core from an
inline `bun -e` eval string. The eval's synthetic /app/[eval] location
can't resolve the workspace-hoisted dependency, so every build of this
stage failed closed with "Cannot find package 'puppeteer-core'" before
the probe itself ever ran.
Moves the probe into a real script file next to probe-beginframe.ts,
which already imports puppeteer-core dynamically the same way and
resolves fine from its real workspace location. Verified locally against
a cached chrome-headless-shell binary.
* fix(engine): bound the webgpu adapter probe so a stuck driver can't hang the caller
requestAdapter() has no native timeout. A broken software GPU driver
that never settles it would hang assertWebGpuAdapterAvailable
indefinitely, including inside the GCP image build's probe RUN step,
producing an opaque build timeout instead of a clean probe failure.
Races the in-page request against a bounded timeout so a stuck adapter
reads as unavailable and fails the same way a genuinely absent one does.
* test(cli): mock the engine's webgpu exports in the browser-fallback test
* test(producer): run the typegpu adapter fixture only on hosts with a hardware WebGPU adapter
The fixture declares data-requires-webgpu, so a software-only host now refuses it
instead of silently rendering the shader blank. The old golden had no shader in it.
Tag it webgpu, skip that tag in the CI image and the test scripts, and re-record the
golden from a Metal render with the shader visible.
* fix(engine): bound the composition fetch that decides the webgpu flag
createCaptureSession fetched the served page with no timeout before launching the
browser, so a hung server stalled a render forever and hung a fake-timer test. Give it a
five second limit, and let that test say it has no WebGPU requirement the way the
producer does.
* fix(capture): refuse a software fallback WebGPU adapter instead of rendering on it
swiftshader reports a WebGPU adapter but cannot render canvas WebGPU, so a
data-requires-webgpu composition on a software-only host crashed on its first frame.
Refuse fallback adapters with the same clear WebGPU-unavailable error. The GCP image
probe now serves its page over http (file:// hides navigator.gpu) and reports a
software-only host instead of failing the image build.
* fix(engine): parse the composition root tag without a backtracking regex
* fix(engine): keep quoted attribute values whole when reading the composition root tag
The test polled a pid file written by `echo $$ > path`, whose shell
redirection creates (truncates) the file before the echo writes to
it. A read landing in that window parsed an empty string to 0, which
satisfied "not undefined" and exited the wait loop early, failing
the next assertion with "expected 0 to be greater than 0" — the
exact shape of the intermittent CI failure. The loop now only
accepts a parsed value that's a positive integer, reproduced with a
busy-wait harness that hit the empty-file race in most iterations
and confirmed clean after the fix.
* feat(telemetry): add output-shape request facts to render events
output_resolution_preset, output_format, hdr_mode, video_frame_format,
and gif_fps_capped join render_complete/render_error, all threaded
from already-resolved RenderOptions/RenderPlan fields rather than
recomputed: createRenderPlan already computes gifFpsCapped (clamping a
requested --fps above 30 to 30 for --format gif) but only used it for
a console warning; this threads that same decision into RenderOptions
and telemetry. All five are known before the pipeline starts, so both
events carry them even when a render fails before a perfSummary
exists.
Also wires output_format/output_resolution_preset into Studio's
separate emitStudioRenderComplete/emitStudioRenderError path, since
Studio's request genuinely varies those two per-request. hdr_mode,
video_frame_format, and gif_fps_capped stay CLI-only: Studio exposes
no HDR-mode, video-frame-format, or GIF output control to users at
all, so there is no per-request value to report for those three.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* feat(telemetry): add adapters_used to render events
adapters_used lists the runtime deterministic adapters
(packages/core/src/runtime/adapters/) a composition exercises.
Investigated the actual codebase rather than the ticket's guessed
7-adapter list and found 12 real registered adapters (gsap, css,
waapi, animejs, lottie, three, typegpu, d3, leaflet, mapbox, maplibre,
google-maps) — the map/data-source adapters share the exact same
window.__hf<Name> registration mechanism and are equally real
"adapters used" facts, so all 12 are included.
Detection unions a live page.evaluate probe (piggybacked on the same
probeSession resolveCompositionElementCount already uses) with a
static regex fallback over the compiled HTML, since this is purely
observational with no gating role (unlike compositionElementCount) —
union strictly increases recall rather than needing one source of
truth. Threaded through the same plumbing chain as the already-merged
composition_element_tags/aroll_video_count/heygen_video_count.
Studio's emitStudioRenderComplete/emitStudioRenderError get this field
automatically, with no Studio-side code change, via the same shared
capture-observability spread heygen_video_count already uses.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* feat(telemetry): add audio/image/sub-composition/color-grading counts
audio_count, image_count, sub_composition_count, audio_group_count,
color_grading_count, and has_lut join composition_element_tags in the
same single scanElementTags pass — audio/image/audio-group counts read
from the scan's uncapped tag map so a 50+-distinct-tag composition
can't push a named count into the "other" bucket. sub_composition_count
counts data-composition-src mounts directly rather than reusing
collectSubCompositionSrcs, which dedupes by src and drops
placeholder/remote mounts (a different, smaller fact than the mount
count this field reports).
has_lut decodes the HTML-entity-escaped form of data-color-grading
before parsing, since linkedom's serializer re-emits it
"-escaped on every compile-pipeline round-trip — without the
decode this silently reported false for every LUT-graded composition.
The LUT-presence check now mirrors normalizeLut's own definition of
"has a LUT" exactly, closing a gap where a blank or empty lut value
would have reported true.
Threaded through the same plumbing chain as the already-merged
per-tag/provenance/adapter fields; Studio's emitter picks these up
automatically via the shared capture-observability spread.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* feat(telemetry): add browser_version_major and ffmpeg_version_major
Both are already-run subprocess calls whose output was discarded:
readToolVersion (checkFFmpeg's preflight) already runs `ffmpeg
-version`, and chromeLaunchOutcome already runs `--version` as a
Chrome launch sanity probe on every render. A shared
extractMajorVersion helper pulls the first X.Y-shaped number from
either banner, threaded onto EnvironmentCheckOutcome/
EnvironmentCheckResult, then RenderOptions right after
runEnvironmentChecks resolves in the local render path only.
Docker renders never populate these two RenderOptions fields (the
containerized producer runs its own preflight the host CLI never
sees), so both properties come back absent there rather than
reporting the host's own toolchain versions for a render that didn't
use them.
Studio does not call runEnvironmentChecks or this ffmpeg/Chrome
preflight at all before creating a render job, so its render_complete/
render_error events will not carry these two properties unless
Studio's own startup path is separately instrumented — a genuine
parity gap, not an oversight in this change.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* feat(telemetry): add authoring_skill_source and authoring_skill_invalid
Both derive from resolution logic createRenderPlan already runs:
authoring_skill_source names whichever step actually won attribution
(an explicit --skill flag, or the skill persisted in hyperframes.json)
using the same flagSkill/projectConfigSkill values already computed
for authoringSkill itself, not a second lookup. authoring_skill_invalid
surfaces the raw --skill value when it fails slug-shape validation
(normalizeSkillSlug rejects malformed input, not unregistered skill
names) — this value already exists as RenderPlan.invalidAuthoringSkill
and already drives a CLI warning in present.ts; this change threads
the same field into telemetry rather than adding a parallel one.
Threaded through both RenderOptions builders in execute.ts (single and
batch render) and all three telemetry call sites in render.ts,
including the Docker path — skill attribution is a request-level fact
independent of which backend executes the render.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* feat(telemetry): add root_body_mismatch and root_body_delta_px_bucket
Measures the composition root's authored data-width/data-height
against the scaffold's own html/body CSS size, straight from the
compiled document string scanElementTags already scans — the
source-level fact, not runtime DOM state. init.ts's applyCompositionSizing
(the #4001 fix) only patches the live DOM after the browser loads this
same HTML, so this measurement stays meaningful whether or not that fix
ran for a given render.
Reuses @hyperframes/parsers' canvasScaffoldPatterns regexes (the same
ones packages/lint's root_dimensions_mismatch rule reads) for the CSS
side. Root-tag detection here is a lighter first-data-composition-id-match
heuristic, not lint's more thorough findRootTag (which also handles a
leading decorative <svg> defs block) — that helper isn't part of
@hyperframes/lint's public API, and expanding it is out of scope here.
The delta is bucketed (0 / 1-10 / 11-50 / 51+ px) rather than reported
raw, and both fields come back absent — not a false/"0" default — when
the root or the CSS block can't be read at all.
Threaded through the same plumbing as every other static-scan field
this week, including the live-capture-observability fallback used on a
crash that never reaches perfSummary.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* feat(telemetry): add hf_env_overrides (names only, capped at 20)
Reports the sorted names of HF_*/HYPERFRAMES_* env vars present when
createRenderPlan resolves — never their values, since several of these
carry filesystem paths. Snapshotting at plan-resolution time (the very
first line of the render command, before any pipeline code runs)
matters: this render's own preflight later injects
HYPERFRAMES_FFMPEG_PATH/HYPERFRAMES_FFPROBE_PATH into its own
process.env, which would otherwise make those two always read back as
"operator overrides" whether or not the operator ever set them.
Reports an empty array, never an absent key, when nothing is set,
matching every other capped/bucketed field added this week. Capped at
20 names, documented alongside the cap constant.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* fix(cli): stop clobbering studio render_complete fields, fix an env leak
trackRenderComplete's explicit audio_count/image_count/root_body_mismatch/
adapters_used/etc. keys, sourced from perfSummary.drawElement and only ever
populated on the CLI's own render.ts path, always won the object spread over
the observability-capture fallback those same keys share with trackRenderError,
even when the direct value was undefined. studioRenderTelemetry.ts's
emitStudioRenderComplete never populates drawElement, only the capture
fallback, so 24 fields silently came back missing from every studio-triggered
render_complete event despite Studio having computed and sent the capture
value. Each field now falls back to its captureXxx counterpart via a new
directOrCapture() helper when the direct field is absent, with regression
tests proving both the fallback and that the direct value still wins when
both are present.
Also excludes HF_SHADER_WORKER_ENTRY from hf_env_overrides: cli.ts sets this
on every invocation to point the bundled shader worker pool at its own file,
sharing the HF_ prefix by coincidence rather than because an operator set it.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* style: remove em and en dashes from added code comments
No behavior change. Rewrites 22 comment lines this PR's own diff introduced,
using a comma or a period where the dash separated two clauses.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* test(producer): add a temporary diagnostic for a CI-only rootBodyMismatch failure
Not a real fix. The rootBodyMismatch test group fails deterministically in
CI (confirmed on two runs of the same head) but will not reproduce locally
by any method tried, including an isolated file run, the exact combined
producer:test:unit command, a worker-count-constrained run, and a fresh
clone with a clean install and build matching CI's own steps.
This dumps the intermediate regex-match state for the simplest failing
fixture to stderr so the next CI run's log says what actually differs.
To be removed once that answer lands.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* test(producer): widen the rootBodyMismatch diagnostic to compare import paths
Still not a real fix. The previous diagnostic (ca4e73983) proved the
imported HTML_BODY_CSS_WIDTH_FIRST_RE has 4 capture groups in the failing
CI run instead of the 2 the source and production code expect, but it used
a dynamic import(), which can resolve through a different codepath than
renderOrchestrator.ts's static import. This adds a static import of the
same binding, a require.resolve() of the resolved file path, and a direct
read of that file's own regex line, so the next run can rule in or out a
static/dynamic resolution split versus a genuinely different file on disk.
Locally both imports agree (2 groups, same object, correct file). To be
removed once the CI answer lands.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* fix(producer): read the merged 4-group scaffold-size regex correctly
main merged #4012 while this branch was in flight, consolidating
HTML_BODY_CSS_WIDTH_FIRST_RE/HEIGHT_FIRST_RE from 2 capture groups (digits
only) to 4 (prefix text, digits, suffix text, digits) so cli's
applyResolutionPreset could reuse the same regex for its in-place replace.
detectRootBodySizeMismatch still read groups 1/2 as the digits, now the
prefix/suffix text, so Number() on them was NaN and every render's
rootBodyMismatch/rootBodyDeltaPxBucket silently came back undefined. Fixed
to read groups 2/4, matching lint's core.ts readHtmlBodyCssSize exactly.
Removes the temporary diagnostic test (ca4e73983, e2e384c8e) now that its
CI log gave the real answer: both a static and dynamic import of the
regex resolved to the same file, and that file's own content had 4 groups
because this branch's merge-base predates #4012, not because of any
resolution or caching difference. A fresh CI run confirmed no CI-only
failure remains after merging main and fixing the indices.
Co-Authored-By: Miguel Ángel <miguel.sierra@heygen.com>
* chore: nudge CI (no workflow run triggered on previous push)
* chore: nudge CI again now that #4017 is out of draft
---------
Co-authored-by: Miguel Ángel <miguel.sierra@heygen.com>