* fix(media-use): local models start outside the project, so a model id never names a project folder
MusicGen, parakeet-mlx and the LTX video runner take Hugging Face ids that their loaders read as
a folder when one exists in the working directory. Each now starts in the home folder; transcribe
still resolves a ./ or ../ --model path against the caller's folder.
* fix(media-use): a relative runner override still names the caller's folder
HYPERFRAMES_PARAKEET and HYPERFRAMES_PYTHON values with a folder in them are made absolute
against the caller's folder before the runner starts in the home folder. Bare names keep their
PATH lookup.
A preview made opaque by its server (Content-Security-Policy: sandbox) under a
frame with no sandbox attribute posts as null. frameOrigins now accepts null
for such a frame, only from that frame's own window and only while its window
refuses reads.
* fix(cli): history turn ids from the project folder are checked and sent encoded
* test(cli): history turn fixtures use the id shape the CLI mints
* fix(cli): a turn's label and agent name are kept as written; only ids are checked
* style(cli): format the turn reader
* refactor(cli): the turn reader checks a marker through small predicates
* fix(ci): publishing workflows run only for main
A workflow_dispatch on any branch ran the ClawHub sync with CLAWHUB_TOKEN and the
MCP registry publish with an OIDC token, publishing that branch's skills/ or
server.json under heygen-com. Each publishing job now skips unless github.ref is
refs/heads/main and checks out the triggering commit without stored credentials.
* test: split the main-only publish check into small checks
* test: close main-only publishing policy gaps
* refactor: separate publishing push filter classification
The static-dedup and at-risk predictors turn each GSAP tween into a frame
range and wrote every frame of every range into a Set. The page decides how
many tweens there are and how long they claim to run, so stacked tweens
multiplied the loop and the at-risk predictor never clamped a range's end at
all: one tween ending at 1e9 s walked billions of frames.
Ranges are now clamped to the frames the render captures, merged before any
frame is written, and the interval list is capped. Past the cap, static dedup
is skipped and every frame is treated as at risk (screenshot capture), so
output never changes for the worse.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(studio): the preview reads hf-preview messages only from its own frame
The preview message router acted on a message with no sender window, and on
any message that arrived before the preview frame existed. It now takes an
hf-preview message only when the preview frame exists, the message comes from
that frame's window, and its origin is one the frame can have: the page's own
origin, or "null" for a sealed frame. The element picker checks the same.
* fix(cli): present sandboxes deck pages another site loads and refuses framing by other sites
present now applies the same another-site decision as play to its
/composition/* route: a load by another site gets a sandbox CSP, its
validators are dropped, and the document is marked so its runtime offers
that site no editing port. Every present response also sends
frame-ancestors 'self' and X-Frame-Options SAMEORIGIN, so only the
presenter page itself can frame the deck. The route setup moves into
buildPresentApp so tests exercise the server as run() builds it.
Every skill now starts the CLI through hyperframes/scripts/plugin-cli.mjs, which starts npm in
the home folder. A standalone install, with no plugin manifest, runs the newest release that
moves into HYPERFRAMES_PLUGIN_CWD. The skills no longer prefer a project's own wrapper over the
CLI.
* ci(publish): publish prebuilt tarballs from a job that runs no release-commit code
Builds, lifecycle scripts and packing move to a pack job with a read-only
token and no secrets. The publish job, which holds the npm token and the
write token, checks each tarball's name and version against the release and
uploads it with lifecycle scripts off. A version already on npm with other
contents fails the first attempt instead of being skipped.
* ci(publish): prerelease tag pushes publish only after a maintainer approves
A prerelease tag can come from any prerelease branch, which no PR review
covers. Tag-push runs now use the npm-prerelease environment, which requires
a maintainer's approval. Merged release PRs keep the npm-publish environment.
* ci(publish): keep every publish in the approval-gated npm-publish environment
npm's trusted publishers are pinned to npm-publish, and that environment now
requires a maintainer's approval with self-review prevented. A second
environment would fail npm auth, so the workflow keeps npm-publish for every
trigger. The docs and test name it as the approval gate.
* fix(parsers): bound canvas scaffold dimension extraction
* chore(parsers): raise packed size budget for linear scaffold scanners
The bounded scanners replace three regexes with a small amount of
procedural code; @hyperframes/parsers packs to 210414 bytes (+1.2%).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): catch-up lists only history lines appended after the hand-off
The seen record now marks the history file (its identity and length) at
the hand-off, and catch-up and the end-of-command notice read only whole
lines appended past that mark. Lines a project already carried are never
shown, whatever dates they claim. After the app trims the history, the
place is found again by the digest of the last line shown; a history the
app set aside and restarted is read from its start; a history replaced or
cut short shows nothing until the next hand-off marks it again. Records
written before the mark keep the dated filter once, then switch to it.
* fix(cli): catch-up notes the history's lines at hand-off instead of a file position
The hand-off now stores short digests of every whole line the history
holds, and catch-up and the notice show lines not among them nor shown
before, in file order. The person's turns keep showing after the app
trims the history or a sync rewrites it, and lines a project came with
never show. A hand-off over an older seen record always notes the
lines; such a record shows dated turns only when all are dated after it.
* fix(cli): catch-up shows nothing for another project now at the same path
The hand-off also notes the project folder's identity (device, inode and
birth time). When a different folder sits at the path, the seen record
counts as no hand-off: catch-up and the notice say nothing, and the next
hand-off notes that project's lines afresh.
Builds, lifecycle scripts and packing move to a pack job with a read-only
token and no secrets. The publish job, which holds the npm token and the
write token, checks each tarball's name and version against the release and
uploads it with lifecycle scripts off. A version already on npm with other
contents fails the first attempt instead of being skipped.
* fix(engine): a cached browser is used only when its files match the pinned build
* ci: install the pinned headless shell for the engine and producer test jobs
* ci: record the producer test job's new setup step for test reachability
* chore: release v0.9.55
* chore: release v0.9.55 notes include #7830
* chore: release v0.9.55 notes include #7859 and #7867
* chore: release v0.9.55 notes name the docs change and cloud render
* chore: release v0.9.55 notes include #7866 and #7801
* chore: release v0.9.55 notes match the style bar
* fix(cli): one trusted PATH rule for the engine and CLI; say why a project copy was skipped
The Windows PATH rules (quoted entries, the %SystemRoot% folders) now live in
the parsers' trustedPathDirs, so engine helper lookups and the CLI's PATH
rewrite apply the same rule.
When sam, aws, gcloud, docker, git or npx is not found but a copy sits in a
project folder the rule skipped, the message names that copy by its path
relative to the project and says why it was not used. cloudrun now reports a
missing gcloud or docker as not found instead of "exited with null".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(cli): join a simulated Windows PATH with the Windows separator
The shared trusted-folder rule now splits PATH on ';' whenever the platform
is win32, as Windows does. Two tests that pretend to be Windows on Linux
built PATH with ':', so the rule saw one folder instead of two.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): name skipped copies above the working folder; pin Windows PATH to the shared rule
The not-found hint now names any copy the trusted PATH rule skipped,
relative to the working folder, including ../node_modules/.bin when an
npm or pnpm script runs from a subfolder. A Windows test pins that quoted
PATH entries still go through the shared rule, so the project root's and
ancestors' node_modules/.bin stay skipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): trim projectCopyHint comments to the package's comment share
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
rootHeadContent and tagRootCompositionFile ran an opener, then [^>]* or a
lazy run, then a required literal, over the project's own HTML. When the
literal never came, each opener rescanned to the end of the page: about
10 s for 250 KB of "<head" or of "<a". Both are now plain scans that give
the same results: only the first <head can match, and each start tag is
checked up to its own > once.
* fix(cli): cloud render sends one zip when it fits and uploads only the largest media
* fix(cli): cloud render leaves unused media out only when the zip is over the limit
* test(cli): eight-at-once upload test waits for eight in flight, never a timer
* test(cli): eight-at-once upload test also fails when more than eight run
* fix(cli): upload archives never include credential or agent-state files
Publish, cloud render and feedback stage the project through one path.
It now always leaves out package-manager and Git credentials (.npmrc,
.yarnrc, .pypirc, pip.conf, .netrc, .git-credentials, .gitconfig), key
files (id_rsa*, id_ed25519*, *.pem, *.key, *.p12, *.pfx), .ssh, .aws,
.gnupg, .docker, .kube, .azure, .config/gcloud and .config/gh, agent and
editor folders (.claude, .codex, .cursor, .agents, .gemini, .vscode,
.idea) and .mcp.json, at any depth and in any letter case. This holds
whatever .hyperframesignore says and whatever the HTML or CSS references.
A reference can bring back a file the ignore rules skipped only when it
is an image, video, audio, font, CSS or JS file, or when it is not in a
hidden folder. Every file left out this way is reported, and a file the
folder walk already reported is not reported again for each reference.
* fix(cli): anchor both alternatives of the key-file pattern
* fix(cli): a page's own scripts in a hidden folder still upload
* fix(cli): the static project server answers only loopback and marks pages another site loads
check, snapshot, validate, layout, compare, grade-compare and motion-shot serve the
project from a loopback server. It now refuses any Host other than loopback or its
in-process name before reading a file. A page or file that another site loads comes
back sandboxed, and its documents carry the another-site mark, so the runtime offers
that site no editing port. The capture browser's own loads are unchanged.
* fix(producer): the render file server answers only loopback and offers no editing port
The render file server now refuses any Host other than loopback before it opens a
file; the capture browser asks for localhost. Every HTML page it serves carries the
another-site mark, so its runtime answers no editing connect. A render never sends
one, so capture is unchanged.
* fix(engine): the exported file server answers only its own host and offers no editing port
createFileServer from @hyperframes/engine now refuses any Host other than loopback or
the hostname it was bound to (on a wildcard bind, this machine's own addresses too)
before it reads a file. Every HTML page it serves carries the another-site mark, so
its runtime answers no editing connect.
* fix(cli): never open another computer's share while staging an upload
publish and cloud render now leave out a reference to another computer's
network share or a Windows device path before any file-system call on it,
and report it as left out with the network_share reason. A page that came
in with an imported project has its outside references left out before the
files they name are checked.
* test(cli): the Windows staging test's lstat mock honours throwIfNoEntry
The received-files lookup above a rendered folder now treats a record path
that exists but is not a readable file (a folder, a dangling link, any
other kind of file) as an unreadable record, the same answer the folder's
own record gives. Only a missing record means nothing was received.
A listed file now counts as inside the rendered subfolder when one of its
parent folders is the same folder on disk, so naming the subfolder in
another letter case on a case-insensitive disk keeps its received files.
* fix(parsers): helper lookups skip the project above the working folder, not only the folder
* fix(parsers): compare a missing PATH folder by its nearest real parent, so Windows short names match
* fix(engine): extraction-time video downloads get the not-media check
A video src that is still remote at frame extraction is downloaded again and
then probed. That copy now gets the checks other media downloads get before
anything reads it as media:
- the not-media sniff, so an HTML, XML or JSON body (including a DASH
manifest) is refused with the existing NotMediaPayloadError, classified as
invalid_media;
- on a render service (PRODUCER_RESTRICT_PAGE_NETWORK), the same
-format_whitelist an https input gets, so a playlist, manifest or concat
list fails to open whatever its bytes say, and no file it names is read.
Real mp4, mov and webm downloads extract the same frames as before.
* fix(engine): the download format check stops when the render is cancelled
* fix(engine): the not-media sniff keeps mp4 files whose box size reads as text
An ISO BMFF file opens with a 4-byte box size and a box type. A size of 60
is the byte `<`, so after the leading NULs were skipped a valid mp4 with a
60-byte ftyp was refused as a text document. The sniff now recognises an
mp4/mov box header (size 0, 1 or at least 8, then ftyp, styp, moov, moof,
mdat, free, skip, wide, sidx or pdin) before it skips NULs. HTML, XML,
JSON and playlist refusals are unchanged.
* fix(engine): the mp4 exception in the not-media sniff needs a box that fits the file
The sniff only skips its document check for an mp4/mov opening when the
first box is ftyp or styp, its declared size (including a 64-bit size) fits
the file length, and the next box header, when the sniff window holds it, is
a known media box that also fits. A document that only spells a box header
falls through to the document and playlist checks as before.
* fix(engine): the not-media sniff keeps mp4 files with leading padding or extension boxes
An mp4 or QuickTime file may open with free, skip, wide, uuid, pdin, moov,
mdat or moof as well as ftyp or styp, and may carry any box type after it.
The sniff now accepts any of those first boxes, then walks every box header
the sniff window holds: each needs a printable four-character type and a
size that fits the file. A document that only spells a box header still
falls through to the document and playlist checks.
The outside-project media check skipped any src starting with data:, blob:,
about: or http(s): in any letter case, while the readers resolve everything
but exact http:// and https:// URLs as a local path. A src such as
data:x/../../outside/clip.mp4 passed the check and the readers opened the
file it named.
The check and the media type preflight now use the readers' remote test,
isHttpUrl, and check every other src with a shared containment helper,
isInsideAssetRoots. Frame extraction, the audio mix, and the HDR and video
sound probes also refuse a local source outside the project, compiled and
download folders on their own when outside files are not allowed.
The CLI static project server and the producer and engine file servers bind
127.0.0.1, but answered any Host header. They now refuse a request whose Host
is not 127.0.0.1, localhost or [::1] at their own port (or the in-process
origin the CLI hands out when listen is refused), and their responses carry
frame-ancestors 'self' and X-Frame-Options: SAMEORIGIN so another site cannot
frame the served page.
The notice a render prints for a received project, and the left-out warnings and
outside-files list that publish, feedback and cloud render print, now show control
characters from references, file names and paths as visible escapes instead of
sending them to the terminal.
While it names the references a received project makes outside itself, the render
check no longer looks at a path on another computer's share or a Windows device
path. It names such a reference in the notice without reading it. A Windows device
prefix such as \\?\ is no longer cut short as if its ? started a query.
A HEYGEN_API_KEY or HYPERFRAMES_API_KEY that a project .env sets now comes
after every credential the person saved or exported. The CLI resolver, the
media environment and the media-use skill use it only when there is no saved
login, no saved API key and no key in the shell. When one is used, a single
stderr line names the .env file.
An expired, foreign-client, unreadable or refused login now raises the login
error instead of switching to a project .env key, whether or not the folder
has a received-files record. A key exported in the shell, or set by the app
that launched the CLI, keeps its current place, and the host gateway path is
unchanged.
* fix(parsers): helper programs never come from the project a subfolder sits in
Run from a subfolder, the trusted PATH check skipped only folders inside the
working directory, so the project's own node_modules/.bin, which npm run puts
on PATH from every ancestor, still supplied helpers. The check now skips every
folder inside the outermost ancestor that marks a project (.git, package.json,
node_modules, hyperframes.json), below the home folder.
A CLI test also runs a real bare-name spawn of an executable planted in the
project folder: it runs under the raw PATH and is not found after the PATH trim.
* fix(parsers): keep the trusted lookup's comment share at main's level
* fix(cli): the audio caches never write or read through links in .hyperframes
The read-audio cache, its stems, mix and descriptions, and the audio-kind cache
now use only real folders under the project's .hyperframes. A link or a file in
place of a cache folder stops the write with a clear message. Every cache file is
written to an exclusive temp file with a random name and renamed into place, so a
link at a temp or final name is replaced, never written through. ffmpeg writes the
voice-free mix into a private staging folder. Stem downloads use the new
downloadToFile replaceLink option. Cached files are read only when they are plain
files, and a stem list may only name files in its own folder.
* fix(cli): the kind cache reads no entry through a linked cache folder
readKindCache now checks that .hyperframes and audio-kind are real folders before
it reads an entry, as writeKindCache already does before it writes one.
* test(cli): read the replaced download before checking its type and mode
The replaceLink test now reads the file's bytes and lstat result once each, so
no read follows a check on the same path.
A download whose response declares a Content-Length re-arms its attempt
timeout on every chunk, so a server that trickles just above the throughput
floor was never cut off by time. Parallel downloads also had no shared limit
on the disk they could fill.
- Each download attempt now has a deadline that incoming data never moves:
30 minutes by default, HYPERFRAMES_MAX_DOWNLOAD_MS to change it. Passing it
fails the download without a local retry.
- The default per-file cap drops from 20 GiB to 8 GiB
(HYPERFRAMES_MAX_DOWNLOAD_BYTES still overrides it).
- Downloads keep 512 MiB free on their volume
(HYPERFRAMES_DOWNLOAD_MIN_FREE_BYTES, 0 turns it off). A declared size is
claimed before the body is read and counted against every parallel
download on the same device, and free space is measured again every 4 MiB
as bytes arrive. Free space that cannot be measured does not block.
- docs/packages/engine.mdx lists the download limits and their variables.
* feat(studio): hosts can seal the preview frames NLEPreview draws
* refactor(studio): Player builds its frame src in a helper, keeping its complexity at main's
* fix(cli): snapshot shows control characters in composition text as escapes
* fix(studio-server): making or serving a still sweeps the transcode cache
Render, capture and the HTML bundler used a narrow private-file rule that
only refused Git data and env files. A composition's script could read the
app's own records under .hyperframes/ (chat history, hand-off, comments,
job log) from the file server, or a data-composition-src could inline one.
Every file server a composition reaches now uses the rule Studio, play and
present already use: Git, env, and .hyperframes/ apart from Studio's own
files and the prepared-assets and requests folders. The rule now matches
Windows spellings too: any case, trailing dots or spaces, backslashes and
stream suffixes.
* feat(studio-server): an opaque preview loads its images, media and fetched files
* refactor(studio-server): the fetch grant's reach is named, not commented
* test(studio-server): the main preview loads under the frame-token path