* fix(engine,producer): media inputs refuse formats that open other files
Composition media read by ffprobe/ffmpeg now passes a format allowlist on every
input: the demuxers the installed ffmpeg reports, minus the ones that open other
files or URLs on their own (DASH, IMF and similar). hls and concat stay allowed,
since ffmpeg carries the list into every file they open. The HDR window,
alpha-plane, GIF, HEIC, audio-sidecar and snapshot reads that skipped it now use it.
The not-media check skips ID3v2 tags before it looks at the first byte, refuses a
playlist entry that is a document, and refuses JSON or non-SVG markup as an <img> src.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(engine,parsers,core,lint,cli): format allowlist reads old listings and covers every media probe
The demuxer listing is read from the binary that enforces the flag, accepts ffmpeg 4's
two-column output, and a failed listing is asked again next time. The lint HEVC probe,
the grade analyzer and the snapshot probes use it too. A nested playlist entry is refused
only for a real document signature, so encrypted segments and keys are kept; an image whose
prolog outruns the sniff window, or that is UTF-16, is no longer refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(parsers,producer): trim comments added with the media format list
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(engine): keep the format-limit probe's argv in the shape the ffprobe contract reads
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(engine): the image markup check reads a doctype without regex backtracking
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(parsers): raise the parsers size budget for the media format list
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(studio): a sealed preview frame forwards Studio's playback keys
* fix(core): a key held while Studio re-sends its watch list still forwards its keyup
* fix(core): the key watch adds and removes its listeners by name, as the runtime typecheck needs
* chore(producer): raise the unpacked size budget for the runtime's key forwarding
* fix(engine): check the resolved address of a restricted download under Bun too
Bun maps a bare undici import to a built-in whose fetch ignores the dispatcher,
so the download lookup that refuses private addresses never ran there.
Load the package by file path.
* refactor(engine): the undici package import says why in its name
* fix(parsers): the timeline inliner stops at a whole-script expansion budget
* fix(parsers): charge bound arguments to the inliner budget and size it from the corpus
A large argument used many times in a helper body is cloned once per use,
which the body count did not cover. Charge the bound node's size for each
reference, set the budget to more than 10x the largest measured real script
(20,760 nodes), and test the hostile shapes from SECOPS-1698.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(parsers): trim inliner budget comments to keep the comment share from rising
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(parsers): charge default expressions and empty expansions to the inliner budget
Default parameter expressions were cloned before the budget was charged,
and once more during the droppable-helper preflight. Empty bodies spent
nothing although each expansion allocates a block. Charge both, and let
the preflight check only for spread arguments instead of building bindings.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Miguel Ángel <miguel.sierra@heygen.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(core): a .hyperframes folder in a subfolder stays private too
The private-file rule treated .hyperframes as the app's records folder only at
the project root. A project copied into a subfolder keeps its own
.hyperframes/, and its chat history, hand-off and other records there were
listed in the file tree and served by Studio's file and preview routes and by
the other file servers that share the rule.
A .hyperframes segment at any depth is now private, in every spelling Windows
opens. Studio's own files and the prepared-assets and requests folders stay
servable only at the root, the one place Studio reads and writes them. The file
tree also skips .hyperframes/backup at any depth.
* fix(cli): the capture --local-file page server answers only its own host
The page server that capture --local-file starts on a loopback port answered a
request whatever its Host header said. A site whose name resolves to 127.0.0.1
could then read the files it serves from the browser.
The server now refuses, before any file lookup, a request whose Host is not its
own loopback name and port, with the same isOwnLoopbackHost rule the static
project server and the engine file server use. Every response carries
X-Content-Type-Options: nosniff. The capture browser loads the page by
127.0.0.1 and its port, and fetchLocalPageUrl reads from disk, so both work as
before.
* fix(skills): a plugin manifest in the project pins no CLI release
The skills launcher read a plugin manifest from the folder above skills/
wherever the skills were installed. For a project-level install that folder is
the project itself or its .agents or .claude folder, so a plugin.json shipped
with a project could pin an old hyperframes release and turn off update checks
and standalone skill refreshes.
The launcher now reads a manifest only from a bundle of its own: a folder whose
name does not start with a dot and that is neither the folder the CLI works in
nor one that contains it. Plugin managers install bundles into their own
folders, so those keep their pin. A manifest anywhere else leaves the call the
same as with no manifest. A manifest that pins a release older than 0.9.34, the
launcher's floor, is refused. The agent-facing plugin rules and the skills
guide say the same.
* fix(skills): a plugin bundle inside the working folder pins only behind a dot folder
A package inside a monorepo could still pin the CLI: with the agent working at
the monorepo root, <mono>/packages/video/plugin.json sat beside a skills folder
the launcher took for a bundle.
The launcher now also refuses a bundle inside the working folder unless a
folder whose name starts with a dot lies between them, as .claude does in
~/.claude/plugins/cache when the agent works in the home folder. Bundles in
~/.claude/plugins, ~/.cursor/plugins and ~/.gemini/extensions keep their pin
from the home folder or a project. The plugin rules and the skills guide now
state the three conditions the launcher checks.
* docs(cli): a .hyperframes folder in a subfolder is left out and not served
The private-file rule now treats a .hyperframes/ folder at any depth as the
app's records. Files stored in one inside a subfolder, assets included, stay out
of uploads and are not served by Studio, preview or render. The cloud render
section now says so beside the rule it already describes.
* fix(cli): capture pages keep loading scripts with no or unlisted extensions
With nosniff on every response, a script the capture page server did not
know the type of (a .cjs file, or a file with no extension) was served as
application/octet-stream, and the browser refused to run it.
The server now names .cjs, .wasm, .xml, source maps, web manifests, CSV,
WebVTT, glTF and more audio, video and image types. A file whose type is still
unlisted is served without nosniff, so the browser sniffs it as it did before
and such a script runs. Every listed type and every empty error response keeps
nosniff.
* fix(share): sync and share leave out a .hyperframes folder below the root
The .hyperframes rules in the source path policy matched only at the project
root, so a copy's records in a subfolder, such as
sub/.hyperframes/agent-runs.jsonl, fell through to the default rule and
travelled with share and sync.
A new nested-hyperframes-state rule leaves out a .hyperframes folder at any
depth below the root. It sits before the first allow rule, as the contract's
rule order requires. The root folder's Studio files and other portable files
travel as before. Stored snapshot reads skip the new rule through
stored_manifest.legacy_rule_regexes, the shape the server and Desktop copies
already use, so snapshots and bindings captured earlier still read.
* test(studio-server): a sub-composition build past its deadline is refused off the request thread
* fix(studio-server): sub-composition previews build off the request thread within a deadline
* fix(studio-server): every parse of a sub-composition page runs on the build worker
* test(studio-server): a page built inside its deadline keeps the request thread free
* test(studio-server): a refused page leaves the next one built whole, hf-ids and all
* chore(studio-server): the build worker's comments fit its package's share
* fix(core): a graded video's opacity animation reaches its export
* fix(core): a graded clip keeps its own clip-path, so it never renders black
* test(producer): golden render of graded video dissolve and wipe
* fix(core): a graded clip stays clickable in the preview, hidden by a filter not a clip
* chore(studio): drop an import the reader no longer uses, and format the graded fixture
* fix(producer): the plan copy checks every project entry, whatever its name
A top-level name that starts with two dots, such as ..assets, is an ordinary
project entry. The plan copy treated it as outside the project and skipped the
link-resolving containment check, so a link with such a name had its target
copied into compiled/. Only the project folder itself now skips the check, and
the outside test compares whole path segments.
* fix(aws-lambda,gcp-cloud-run): untar leaves out symbolic links and special files
A project archive can come from another person. A symbolic link in it could
point anywhere on the render host, and the planner would copy its target. Untar
now unpacks only files, folders and hard links. node-tar already keeps a hard
link's target inside the destination folder. The CLI's site upload follows
symbolic links when it packs, so its archives hold none.
* fix(aws-lambda,gcp-cloud-run): untar turns hard links into separate plain files
The site upload packs two hard-linked project files as one file and a hard
link. Untar still unpacks both, then gives each hard-linked file its own copy,
so the extracted project holds no links of any kind and loses no file.
* fix(aws-lambda,gcp-cloud-run): copy hard-linked files without mkdtemp; untar test in its own file
* fix(cloud): keep the hard-link copy name short so long file names still unpack
* fix(producer): render compile never inlines private project files named by @font-face
* fix(producer): a font link leading outside still warns before the private-file check
* fix(studio): the preview picture takes a new box size in the same frame
The preview stage measured its box in a ResizeObserver and stored the size
in React state, which rendered a frame later. A box that grows at once (the
window going full screen) painted one frame with the old, small picture, and
macOS animated that frame for its whole full-screen transition.
* fix(studio): the preview picture takes a new box size in the same frame
* fix(studio): the preview picture is sized from its box's layout, not its on-screen rect
* style(studio): format the preview stage size test
* fix(engine,producer): render media readers treat private project files as missing
A video, audio, GIF or HEIC src that names a file the file servers refuse
(git metadata, env, credentials, tool state, app records), as written or
through a link, is now treated as missing by every render-time reader:
the compile-time duration probe and keyframe advisory, GIF and HEIC prep,
the media drop step, the HDR probe, frame extraction and the audio mixer.
The engine's isPrivateUnderAssetRoots judges a path from every asset root
that holds it, so a render's own work folder inside the project's
.hyperframes/ keeps its downloaded and prepared files.
* fix(engine): probe refuses a text file that ffmpeg reads as text-art video
ffmpeg's tty and bintext demuxers turn a text file (such as .asc or .nfo)
into a video stream that draws the text into frames. Every probe now
refuses an input whose container is tty or whose video codec is ansi,
bintext, xbin or idf, as not media. Probing, frame extraction and audio
reads all go through the same probe, on every render. No demuxer for real
video or audio produces these, so local renders keep every other format.
* fix(engine,producer,cli): playlist entries and check probes skip private files
- A media playlist (HLS segment, key or map URI, ffconcat file) that names a
private file inside its own folder is refused as not media, like one that
names a file outside it.
- `hyperframes check` no longer probes a media source outside the project or
private in it: mediaPath returns null for such a file and the audio,
heard-clip and HDR checks skip it.
- The compiler's duration probe and keyframe check judge privacy from every
asset root, as the render-time readers do.
* docs(skills): style-sampling pulls the template every time, so a received copy is checked first
* chore(skills): regenerate the skills manifest for the style-sampling change
* test(skills): style-sampling must pull before using a template, resuming included
* fix(studio): a host that shows a still gets a preview reload with one page at a time
A full reload kept the live page beside the reload's own until it was promoted, so a heavy project held two
whole pages at once. A host that can show the live frame as a still (holdPreviewStill) now has the live page
dropped behind it while the reload's page loads alone; the still goes once the new page shows the live time,
and a reload that gives up brings the live page back at its time. With no still, nothing changes.
* fix(studio): keep the reload hook's comment share at the base
* fix(studio): pause before the still so a held reload resumes from the frame it shows
* fix(studio): drop a give-up's still on the live page's own load, not a polled load event
* fix(studio): a host that throws while taking the still falls back to the two-page reload
* fix: warnings that name project files print control characters as visible escapes
The CLI already escaped control characters in the text its own commands
print, but warnings from core, producer and engine went to the terminal
raw. A composition whose sub-composition path held ESC or BEL still
drove the terminal through the bundler's "Skipping sub-composition"
warning in `info --media` and `snapshot`.
terminalText moves to core so there is one sanitiser, the CLI imports it
from there, and the library warnings that carry project text (the
bundler, the producer logger, the file server, render path skips, asset
load failures, FrameCapture, the import map, fonts, figma and atomic
file warnings) pass that text through it.
* fix: every library warning that carries project text escapes control characters
* fix(studio): recovery drafts are kept by each project's own server, never the browser
* fix(studio): a restored draft stays offered when Studio settles which composition is open
* fix(studio): a draft read never waits on a pipe or loads an oversized record
* test(studio): the draft read's child fixture is an entry point, and its note sits in the test
* chore(studio-server): size budget covers the recovery-draft route
* chore(cli): unpacked size budget covers the recovery-draft route
* fix(media-use): audio engine reads and writes only inside the project folder
A voice line id must be a plain file name and a recording path must stay
inside the project. The engine never writes through a link or into a folder
that leads outside the project. wait-bgm reads only the renderer's own log in
assets/bgm, and a carried-over BGM job keeps its pid and log only when that
log is the engine's own.
* fix(media-use): engine outputs and recordings stay inside the project
A recording that is a playlist or list naming other files is refused,
and ffmpeg is pinned to the probed format with file-only input. The
media manifest, its index, bgm_status.json and an in-project
audio_meta.json are never written through a link, and a .media folder
that leads outside the project is refused. Index cells keep each record
on one line. wait-bgm ignores a track path outside the project and
drops a pid whose log is not the engine's own.
* fix(cli): the published skills leave their test files out of the package
The pack job no longer downloads Lambda assets already attached to the
release; it always pins the handler zip and video image this run built.
The publish job decides what the release keeps. Once
@hyperframes/aws-lambda is on npm for the version, the release must hold
exactly the assets that package pins. Before then, an asset this run did
not produce fails a first attempt, and a re-run of the same workflow run
replaces it. An asset identical to the bundle's is left alone, so a
re-run of only the publish job also works.
* fix(cli): batch outputs from a template that starts with a placeholder stay in the working folder
* fix(cli): the bare {path} template is named, not commented
* fix(hdr): bound HDR image decode and resample sizes
The native HDR image path sized its host buffers from the PNG header and
the element's page-measured layout box, with no ceiling.
- PNG decode checks IHDR dimensions and bit depth before allocating,
inflates with maxOutputLength = height * (stride + 1), and rejects image
data whose inflated length does not match the header.
- HDR images are limited to 8192x8192 pixels; resampleRgb48leObjectFit
refuses targets above that.
- A layout box larger than the output frame is shrunk uniformly to fit it.
- Decoded and resampled HDR image buffers count against the HDR extraction
budget (HDR_EXTRACTION_MAX_BYTES / cgroup limit) for the render's lifetime.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* build(engine): raise the packed size budget to 510000 bytes
The PNG header checks and bounded inflate add about 500 bytes to the
packed engine, which was already within a few hundred bytes of its
1% tolerance over the 500000-byte budget.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): the CLI trusts a preview only once it proves this user started it
* test(cli): a listener that cannot prove it is the user's preview is never trusted
* test(cli): stand-in previews answer the challenge as the user's own do
* style(cli): format the preview proof changes
* fix(cli): a preview's proof names its own endpoint and config, and the token file must be private
* style(cli): format the proof changes
* test(studio): the fake request in the host check has a socket, as every real one does
* fix(studio): a Studio started by hand proves it is the user's, so the CLI still finds it
* fix(cli): the challenge travels in a header, so Studio's routes keep their exact paths
* style(studio): keep the config producer compact
* fix(cli): fixed-name outputs and history turn records replace a link, never write through it
transcript.json, hyperframes.json patches, the grade-compare sheet, the batch manifest, diff and
measure outputs, lambda stack state and the history turn records are written through a temp file
renamed into place. A linked hyperframes.json is left alone, and the default .hyperframes/<tool>
output folder is used only when every folder on the way is real.
* fix(cli): measure and diff pictures and sbs.mp4 replace a link, never write through it
writeImage and the diff side-by-side now have ffmpeg write into a new private folder beside the
output, then rename the file over its name. measure frames already staged and renamed its PNGs.
* wip: paused before review
* test: cover credential files on the preview routes and the deploy notice
* refactor: name the inspector's file lookup for what it serves
* fix(cli): a link to a private file outside the project keeps its own left-out reason
* fix: escape the upload notice and skip private files from old manifests
* fix(inspector): the inspection page reaches only its own server and public hosts
* fix(inspector): release the page server when the proxy fails to start
* fix(inspector): open each source session in its own browser context
Image prep (HEIF and GIF, in renders and the Studio preview) now applies the
other-network-root test to every candidate path, whatever its leading
character, and decides whether a candidate is allowed before checking that it
exists. Outside-project asset collection leaves out paths on another network
root with the existing network_share reason before any file lookup.
Tests run both under Windows path rules and record every path the disk is
asked about.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
On Windows even a stat or realpath of a \\host\share path connects to that
host and signs in as the current user. Three readers looked such paths up
before refusing them:
- `hyperframes compositions` resolved a data-composition-src with
path.resolve, so an absolute or share src reached readProjectFile. It now
uses resolveCompositionSrc, the resolver preview and render use.
- The timeline reader (`timeline`, `audio`, measure) realpath'd a
sub-composition before its containment check. It now refuses a path on
another network root first.
- Artifact bookkeeping for render, check and snapshot looked up the paths
recorded in .hyperframes/video-state.json. A recorded path must now be
spelled inside a .hyperframes/artifacts/<kind> folder on the project's own
root before cleanup looks at it, and a record naming another share counts
as gone without a lookup.
* wip: partial security fix, paused before tests finished
* fix(cli): templates pull checks every file of a kit and moves a changed one aside
A template folder counts as pulled only when every path under it matches the
file list a verified download unpacked to: the same paths, sizes and SHA-256,
no extra files, links or hard-linked files. That list lives in the user's
cache, never in the project. A folder that differs is moved to a dated
templates/<slug>.changed-<time> sibling and a verified copy takes its place.
* test(cli): drop a test comment the helper's name already says
* wip: partial security fix, paused before tests finished
* fix(cli): a received-files record in a folder under the project counts for a render
The received-files lookup now also walks the project's folders for records,
skipping tool and output folders and never following links. A nested
record's entries are received files named from the project folder, so a
received scene copied in with its own record makes the render leave out
outside files, as a record at or above the project does. A walk that cannot
list every folder fails closed.
* fix(engine): give each superset extraction a folder of its own
A superset group extracted into a folder named from the group index and the
process id, and deleted and recreated it by that name in the cache root that
every render shares. Two renders in one process could share the folder and
remove each other's frames.
Partial folders are now created with mkdtemp, so each extraction writes only
into a folder it made. A cache entry is published only when its frame count
is the one the extraction wrote.
* fix(engine): make the partial folders with createStagingDir; raise the engine size budget
* fix(producer): compile and render-telemetry tag scans run in linear time
The compiler's gsap script rewrite, crossorigin and preload strips, remote
media and image src scans and inline-script scan, and the render telemetry's
tag counts, now use forward scans that return exactly what the regexes they
replace returned. Those regexes rescanned to the end of the page from every
tag opener that never closed, or from every character of a long whitespace
run in one tag. The telemetry's script and style strip stops after 8 passes.
* test(producer): give the scan fuzz tests room on CI and raise the producer size budget
* fix(studio-server): history restores stay inside the project and use only intact copies
* fix(studio-server): a history path's trailing dots and spaces are trimmed by a scan, not a regex
* test(lint): density warning ignores lanes
* fix(lint): count dense timelines per file so agents keep scenes on one lane
* chore(skills): make room for the lane pointer in the entry skill
* fix(lint): plain comment on the per-file density count
* chore(skills): say why the entry skill budget grew
* test(skills): pin the one-lane wording agents follow
* style(skills): format the lane wording test
* chore(skills): regenerate skills manifest
* docs(skills): overlapping sound effects take a second lane
* wip: paused before review
* test: cover credential files on the preview routes and the deploy notice
* refactor: name the inspector's file lookup for what it serves
* fix(cli): a link to a private file outside the project keeps its own left-out reason
* fix: escape the upload notice and skip private files from old manifests
* wip: paused before review
* fix(core): keep the flash fade, with a checked duration set per element
* fix(core): find stylable flash targets with the realm-safe predicate
* 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