* fix(cli): transcribe replaces a link at its output names instead of writing through it
whisper-cli now writes its JSON in a private temp folder, and the CLI puts
transcript.json in place by rename. The HeyGen engine, the SRT/VTT caption
file and the caption script patch also write by rename. A link planted at
transcript.json, transcript.srt or transcript.vtt is replaced by a regular
file, and the file it pointed to is left unchanged.
* test(cli): prove the link was replaced from file contents, not lstat
CodeQL reads an lstat followed by a read of the same path as a race. The
outside file staying unchanged while the output name holds the new
content already shows the link was replaced.
* test(cli): make a caption write fail the same way on every OS
A dangling link through a file was refused on POSIX but replaced on
Windows, where access() on it reports ENOENT and the rename takes its
place. Both are safe, so the failure-message tests now use a caption
name the disk accepts but whose temp sibling is too long.
* fix(cli): the credit limit counts no logged row below zero and resumes only keys this account sent
A job log row whose credits are negative or not finite now holds nothing,
so no row can lower the total spent against the limit. A logged job key
skips the limit check only when this account sent it before, as noted in
a per-account record in the config folder. A logged key it never sent is
priced as a new call, without the credits the log gives that key. Applied
to both copies of credit-limit.mjs.
* fix(cli): publish's prompt names the existing project it updates and says its visibility is kept
When the publish targets an existing project (an --update id, the folder's
.hyperframes/project.json, or this machine's earlier publish of the
folder), the confirmation now names that project and says it keeps its
current visibility, which may be public, instead of promising a private
upload. The id is shown as printable text only.
* docs(cli): credit-limit says a logged row never counts below zero
The earlier commit said no row can lower the total. That is not so: a
later row for the same key with credits 0 sets that key to zero, which is
how a refused call gives back its reservation and how a charge below the
estimate replaces it. What holds is that no row counts below zero.
* fix(cli): sent job keys are noted per account and project folder, privately and capped
A logged key now resumes unchecked only when this account sent it for the
same project folder, so a key sent for one project cannot skip the limit
in another. The record keeps its newest 10,000 keys, and the config
folder and key files are created readable by their owner only.
* fix(cli): publish's prompt names the existing project in every variant
Every prompt variant, --public included, now names the existing project a
publish would update. It says the update happens only when the person is
signed in and can access that project, and that otherwise a new project
is created, which is what a signed-out publish does.
* fix(cli): the credit limit reads this account's spending from its own ledger
Each account now keeps, in its config folder, a ledger of what its jobs in
each project folder hold: every charge, reservation and refund the CLI
logs is noted there too. The limit check counts this account's keys at
the ledger's value and ignores the project log's rows for them, so a row
a sync or a pull adds cannot lower them. Any other key in the project log
counts at the most it has been seen to hold, never less, and never lets a
call skip the check. This replaces the sent-key record; the ledger keeps
its newest 10,000 lines and is readable by its owner only.
* fix(cli): publish reads the sign-in before asking, so its prompt says what will happen
The credential is now read, and --update or --space without a login is
refused, before the confirmation prompt. The prompt names an existing
target project only for a login, which is the only case that updates it
in place; signed out or on an API key, the prompt describes the new
private upload that actually happens.
* fix(cli): creditsSpent skips a malformed or null job log line
One junk line in .hyperframes/heygen-jobs.jsonl no longer blocks every
priced call in the project; the reader skips it, as the account ledger
reader does. Applied to both copies of credit-limit.mjs.
* fix(cli): the credit limit refuses a paid call when it can't tell which account is signed in
Under a credit limit, a call whose signed-in account can't be named is
refused before anything is sent, since this account's ledger can't be
read or written then. With no limit set, nothing changes.
The ledger is trimmed through a temp file and a rename, and the project
being checked keeps its last line for every key it holds. A refusal that
counts credits from job log rows no longer in the project says so and
names the ledger file. The ledger scope notes that a moved or renamed
folder starts a new ledger. The private-mode check in the ledger test is
skipped on Windows, which has no POSIX modes.
* fix(cli): cache installs only use folders the current user owns
Code the CLI downloads or builds into its cache (Chrome, the TTS script,
audio helpers, whisper.cpp, OpenCV, the Lambda handler, optional packages)
now goes into folders created by the current user with owner-only write,
and each installed file is checked again before it is returned.
- A cache root that other accounts can add entries to, sticky or not, is
no longer used for code; the per-user cache is used instead.
- The ownership check reads links rather than following them, and a link
that leads out of the cache root is not trusted.
* fix(cli): validate cache ancestry before filesystem side effects
* fix(core): check links before parent path segments
* fix(cli): validate cached libraries and clarify private parent requirements
* fix(cli): use the shared staging directory helper for Apple builds
The lambda commands read the bucket and state machine from
.hyperframes/lambda-stack-<name>.json, which arrives with any project
someone sends. A file naming another account's bucket made render,
render-batch and sites create upload the project there.
- renderToLambda pins ExpectedBucketOwner to the state machine ARN's
account on every S3 request; deploySite takes expectedBucketOwner,
which sites create and render-batch pass.
- The state file is read only as a plain file in a real .hyperframes
folder, and its bucket, ARN, region and other fields are checked.
* fix: hidden and credential files stay out of Studio, history and lint
- The private-file rule now refuses every hidden file or folder in a project,
except Studio's own files, the root .media folder, and pictures, video,
sound, fonts, stylesheets and scripts a page can load. It also refuses
Terraform state and variable files and more key-store types.
- Project history never tracks, lists or serves a private file, and the blob
route serves only bytes a listed file had, so copies kept by an older
version are no longer readable.
- Lint reads a linked stylesheet, script, import map or sub-composition only
when the private-file rule allows it and the file is the kind its link
names, so no finding quotes a private file.
- The rule moves to @hyperframes/parsers/private-path so lint can use it;
@hyperframes/core/safe-path re-exports it unchanged.
* fix(studio-server): a refused history blob reads as not a history blob
* fix(core): keep realpath in core, so its file-system callers and tests see one module
* fix(cli): text from projects, pages and registries prints control characters as visible escapes
catch-up, capture, figma component, timeline, keyframes, layout/inspect,
catalog and add printed strings that come from a project folder, a captured
web page or a registry manifest straight to the terminal. A crafted value
could move the cursor, rewrite earlier lines, set the window title or write
the clipboard. Those values now go through terminalText(), the same helper
lint, check and snapshot already use. The CLI's own colours stay, and --json
output is unchanged.
layout/inspect also keeps only well-formed issues with a known code from the
page, the way check already does, by reusing check's parseLayoutIssue. The
browser-script loader moves to its own module so layout and check can share
the parser without importing each other.
* test(cli): catch-up test names files with control characters only where the OS allows it
* fix(cli): import terminalText from core after main moved it
* fix(cli): escape layout banner and failure, top-level errors and installed paths; keep line breaks
* fix(core): a page marked for another site gives that site only playback
The runtime control bridge trusted window.parent alone. In a document the
preview server marked as framed by another site, that parent is the other
site, so it could turn on pick mode, flash elements, set runtime data and
colour grading, and it received the timeline, diagnostics and picked
elements, which describe the composition.
In a marked page the bridge now takes only playback controls (play, pause,
seek, volume, mute, rate) and posts only the ready, state and stage-size
messages. Pages served to their own site are unchanged.
* fix(core): keep the bridge's handler lookup in the shape code scanning accepts
* fix(studio-server): no write route touches files other tools run or trust
A composition previewed in Studio runs same-origin with the API, and the write rule
refused only private paths, node_modules and hidden names. Its script could create,
overwrite, rename or delete package.json, lock files, agent instruction files and
scripts, which a package manager, build tool or coding agent then runs with the
developer's rights.
Core now names these files in one predicate, toolTrustedFileKind, read in every
spelling a case-insensitive file system or Windows opens. Studio's write rule and the
registry installer's target rule both use it, so every write route (PUT, POST, PATCH,
DELETE, upload, duplicate-file, file-mutations and gsap-mutations) refuses such a
target. Deleting or moving a folder that holds one is refused too. Ordinary
composition, stylesheet, script-asset and media writes still succeed.
* fix(core): fold look-alike letters before matching files other tools trust
A case-insensitive macOS volume opens "AGENTſ.md" (long s) as AGENTS.md
and "Makefile" (fi ligature) as Makefile, but lower-casing alone does not
map either letter. Project path segments now fold names the way such a
volume does: NFKC, upper-then-lower case, and invisible characters
dropped. Every rule built on the segments gets the fold.
Also:
- name JSON, YAML and TOML tool config, compose files, turbo.json,
nx.json, pyproject.toml and gulp or grunt files
- treat a ~N name as a short name only when it has the 8.3 shape and an
extension a trusted file's could shorten to, so take~2.mp4 stays
writable
- refuse to delete or move a folder holding node_modules or .git, walking
it without the listing's ignore rules
- report an upload refused by the write rule in `skipped`
* feat(cli): diff --hide scores a project's layer by what it adds
A project folder is captured a second time with the layer hidden (a clone, so its
fonts and sub-compositions stay). Where the captures differ is the layer's footprint;
over it, benefit = 1 - MAE(original, with) / MAE(original, without), and the
original's copy of the layer is fitted by position (+-40 px) and size (80-120%)
into an instruction such as "move down 12 px". Moments where leaving the layer out
already matches the original are not scored. Reported in the result's `layer`;
`pass` is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): diff --hide scores a layer over most of the frame without running out of stack
The footprint's box came from Math.min(...pixels), which throws RangeError past about
150k pixels (a 500x300 card at 1080p), aborting the whole diff. The box is now one loop,
and the position/size fit reads an evenly strided sample of at most 20k footprint
pixels; benefit still reads all of them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(lint): run Studio and producer lint off the host's own thread
A composition that sends the linter's parser into a runaway froze or
crashed the long-lived hosts that lint it in their own process.
CLI Studio's single-page lint now runs in a child process with a heap
cap and a 15 s limit. Producer /lint (runHyperframeLint) runs on a
worker thread with the same limits. A page that times out, runs out of
memory, or whose child or thread ends without an answer gets one
lint_aborted error finding and is never linted in the host process.
Studio's project lint keeps its timeout and gains a heap cap.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(producer): cap concurrent /lint workers at two
Each lint worker can hold up to 512 MB, so parallel /lint requests
multiplied that without bound. At most two run at once; the rest wait
in order, and a request still waiting when its own 15 s runs out gets
the lint_aborted result.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(producer): probe sub-composition media defaults
Sub-composition defaults declared on <html> disappear from the DOM
during inlining. The browser-probe gate then misses data-var-src
bindings, causing renders to use the authored placeholder source.
Pass the inliner's variablesByComp table through CompiledComposition
and consult the nearest composition scope when checking media bindings.
Give anonymous mounts distinct runtime scopes and reserve IDs as
nested mounts are processed, preserving each instance's variable values.
Add regression coverage for declared defaults, repeated mounts,
nested scopes, and named/anonymous mount collisions.
Fixes#7505
* fix(core): keep anonymous hosts unnamed, key their variables by the authored id
The previous commit made an anonymous mount claim its sub-composition's id on
the host element, which changed the DOM contract the runtime mount path and
the `missing-host-comp-id` regression golden pin: the compile paths put the
host's start/duration on the composition while the mount path kept them on an
unnamed timed element, so the producer parity test and the regression shard
failed on the first CI run.
Restore the contract: an anonymous host carries no composition id and its
flattened wrapper keeps the authored one. For #7505 that wrapper is already
the scope the probe gate (and `variableScope.ts`) resolves through, so the
inliner now keys an anonymous mount's merged defaults by the authored id
instead of skipping them. A named mount of the same file shares that scope;
its explicit instance values win regardless of inlining order, which the
replacement unit test pins in both orders. The id-reservation machinery for
claimed ids goes with the claim.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* fix(cli): validate and keyframes print project text with control characters as visible escapes
* test(cli): the CSS case prints a payload the declaration split keeps whole
* fix(cli): escape project-supplied object keys too, so keyframes prints tween property names safely
When an <audio> clip's source failed to load (a 404, a format Chrome
can't play), the slot-fit audit put it with the clips whose metadata
was still loading and told the person to re-run with a longer --timeout,
next to the http_error for the same file. The audit's probe already
stops waiting on the error, so the timeout was never the cause. Record
whether the element or its probe errored, and report those clips as
failed to load, without the timeout advice.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
citty's help appends "(Default: <value>)" to every arg whose `default` is
set, and 22 descriptions also spelled the default out, so `--help` read
"(default: 5) (Default: 5)". Drop the hand-written copies, reword the
lint-verbose and `measure extract --audio` flags whose "(default: a
one-line summary)" sat next to "(Default: false)", and add a test that
fails when a description with "(default:" has a `default` too.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): render --json gives each missing tool, a held slot and an old Node its own code
Every failed `render --json` answered with `error.code: "runtime_error"`, so a
script could only tell a missing FFmpeg from a busy render slot by matching the
human message. A CliRuntimeError now carries a code, as CliUsageError does, and
the envelope reports it: chrome_not_found, ffmpeg_not_found, ffprobe_not_found,
render_locked. A Node older than 22 prints the envelope too
(node_version_unsupported) when the run asked for --json. Other failures keep
runtime_error.
Refs #7408
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
* test(cli): give the bin entry test's runtimeVersion stub the new JSON export
bin/hyperframes.mjs now imports runtimeVersionJson, so the stub dist file
binEntry.test.ts writes must export it too, or the bin fails to load.
---------
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
vite.config.ts imported createFsAdapter from @hyperframes/sdk/adapters/fs.
Vite leaves a bare import in a config file to Node, and that export points at
src/adapters/fs.ts, so Node had to load a .ts file itself. Node only does
that without a flag from 22.18, and on 22.0 to 22.17 build, dev and test all
stopped with ERR_UNKNOWN_FILE_EXTENSION, which also failed the last step of
the root build.
The config now imports the adapter by its relative source path, which Vite
bundles with the config on every Node version.
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
The Contrast section printed every finding with the error mark, so a
failure seen at a single sample (a warning) read as "✗ ... 0 error(s),
1 warning(s)". It now uses the same severity icon as the other sections.
A finding on an element without an id is named by an nth-of-type path,
which does not say which text failed. The line now quotes the text the
audit judged, as `validate` already does.
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
* fix(core): a clip the preview lets go of stays unloaded in Firefox
Releasing a clip dropped its src, called load() and set the src again.
Firefox treats preload="none" as "metadata" while an explicit load() is
pending, so every release fetched the clip's metadata and first frame. The
runtime then saw a far clip loading and released it again, in a loop: about
20,700 media requests over one 2:16 film, with 2-4 s stalls at cuts.
Setting src already runs the media load algorithm, which ends the fetch, so
the release now sets it again without load(). Chrome and Safari behave as
before: Blink clears its ignore-preload-none flag when src is set, and
WebKit's src set cancels the prepareToPlay that load() queued.
To reproduce: in Firefox, give a <video> with a src preload="none", then call
removeAttribute("src"), load() and setAttribute("src", src) in one task. It
loads to readyState 4; with the src set in a later task it stays at 0.
* fix(player): a film cut from one long recording plays through in Chrome and Firefox
A film whose clips are cut from one or two long recordings froze a few
seconds in, in Chrome and Firefox. Both keep a paused clip's request open
with its unread bytes, and every request to one host shares one HTTP/2 or
HTTP/3 connection window. With every clip of the file loading at parse, those
idle streams fill the window, and the request for the clip coming up is never
answered.
The player's frame now holds clips back on every engine, as it already did on
WebKit: each file keeps the clip on screen and the next one loading, and
frees the rest. The player also names its frame, because a composition in an
opaque sandbox cannot read frameElement to find the player's mark.
To reproduce: play a film of 37 clips cut from a 363 MB and a 180 MB .mov
served over HTTP/3 in <hyperframes-player>. On main it stops at 0:17 in
Chrome and at 0:09 in Firefox and never recovers.
* chore(player): raise the packed size budget for the frame name
Naming the player's frame adds 44 bytes to the packed player (158,706 to
158,750 bytes with pnpm 10.17.1). main was already 10 bytes under the 1%
tolerance of its 157,145-byte budget, so any growth fails the check. The
budget moves to 5% above the new measured size, as the check suggests.
* fix(core): seeking in the player stays as it was in Chrome and Firefox
Holding clips back in every player frame also moved Chrome and Firefox
player frames onto Studio's paused-seek path, which keeps the old picture
until every clip on screen has its new frame. In the player's scrub
benchmark the inline p95 rose from about 1.0 s on main to 1.9 s in CI,
and a seek could miss the benchmark's 1 s limit. In Firefox a paused seek
to a clip the player had let go showed no frame for about 5 s, because the
clip was released again each time it loaded.
Chrome and Firefox player frames now hold clips back and seek as they did
on main. Studio and Safari player frames keep the path they already had.
The button puts the whole player, controls included, in fullscreen and follows every
fullscreen change. Set disable-fullscreen to hide it.
It raises the player's package size budget. With pnpm 10.17.1, main packs to 158,705 bytes
(544,891 unpacked) and this branch to 160,137 (549,402).
* fix(engine): video warnings show control characters in the src as visible escapes
The missing-source and opaque-alpha warnings pass the authored src, and the path
built from it, through terminalText before writing them to stderr.
* fix(cli): snapshot never probes or decodes a private project file for a page video
projectLocalFile applies the same private-file rule as the render readers, judged as written
and with links resolved, so held-video probes and frame injection skip those files. Studio's
own prepared-assets and requests folders still resolve.
* fix(parsers,engine): media inputs never select a text-art demuxer
The media format allowlist leaves out tty, bin, xbin, idf and adf, so ffmpeg and ffprobe
cannot draw a text file as a video frame, whether or not ffprobe is installed. ffprobe's
refusal of one of these formats still reports the source as a text file.
A link inside the project whose target is outside it is still read by snapshot: links a
person puts in their own project are accepted, imported projects cannot carry links, and
render reads them the same way.
* fix(cli): registry warnings show control characters as visible escapes
A project's hyperframes.json can name the registry, so item names and deprecation text
from it pass through terminalText before the studio install, remote template and
item-loading warnings write them to stderr.
* fix(cli): failed render and frame read messages escape control characters
The render failure box escapes each line of the message, keeping its line breaks, and
snapshot's frame read errors escape the path and the ffmpeg or ffprobe text they quote.
* test(cli): probe a missing path with control characters instead of creating it
Windows cannot create a file with those characters in its name; the failed probe's
message still carries the path.
The bypass entry for a restricted browser's page server now always names its
port, so a default port no longer admits every port on the host. Studio
thumbnails take the server's own origin, not a request's Host header. A stop
signal that lands while the temp profile is being made removes it and launches
no Chrome. The wiring test follows renamed and namespaced calls, requires the
callback to pass on the network's args and profile, and refuses an unset
pageServerUrl.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(core): describe replies carry the picker's element label, and studio exports standInHandle
* fix(core): label an element by its rendered text, and read a describe reply without a label
* style(core): format the rendered-text walk
* fix(studio): beat analysis never picks a recording longer than a song
* fix(studio): a long music track keeps its saved beats; the cap judges the file alone
* fix(studio): a music track's beat edits survive its file length arriving
* fix(cli): own render-slot lock dir; refuse cache files other accounts can write
Sharing one render slot across accounts no longer means sharing HYPERFRAMES_CACHE_DIR,
which holds the browser, packages and helpers each account runs.
HYPERFRAMES_RENDER_LOCK_DIR names a folder of lock records only, created sticky and
writable by all (1777); the per-user lock folder is created 0700. Help text, errors and
docs no longer suggest sharing the cache or chmod 777.
foreignWritablePath (core/cache-dir) refuses a cached file whose path up to the cache
root is owned by another account or is group/world-writable. Chrome (CLI, engine,
inspector), optional packages (incl. the Sherpa runtime), the whisper.cpp build, the TTS
script, the Apple audio helpers, opencv.js and the Lambda handler reinstall or rebuild
instead of running such a file, and trustedCacheRoot moves their installs to
~/.cache/hyperframes when another account can write HYPERFRAMES_CACHE_DIR.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): seal own cache installs against umask 002 and repair old lock folder modes
sealCacheInstall (core/cache-dir) takes group and world write off what this account just
installed and the folders above it up to the cache root, so an install made under umask
002 passes foreignWritablePath on the next run instead of being refused and fetched again.
Chrome, optional packages, the whisper.cpp build, opencv.js, the TTS script, the Apple
audio helpers and the Lambda handler call it after installing.
ensureLockDir makes an existing per-user lock folder 0700 again when an older release left
it open, and adds the sticky bit to an existing HYPERFRAMES_RENDER_LOCK_DIR folder, warning
once with the folder's name when it cannot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(cli): drop added comments to pass the comment ratchet
The explanations move into names: the mode bits in core/cache-dir are GROUP_OR_WORLD_WRITE
and WITHOUT_GROUP_OR_WORLD_WRITE, and the owner check is anotherAccountCanWrite. No change
in behaviour.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): verify the lock folder through a no-follow handle and stage installs owner-only
The render lock folder is opened with O_NOFOLLOW | O_DIRECTORY and checked and chmod-ed
through that descriptor: a link, or a folder another account owns (root may own a shared
one), is refused with RenderLockFolderRefused instead of being made world-writable or
used as is.
Chrome staging, the optional-package staging folder, the whisper.cpp build folder and the
opencv, TTS, Apple helper and Lambda handler folders are created 0700, and Chrome's staged
tree is sealed before it leaves staging, so nothing is group-writable while it installs.
foreignWritablePath accepts a sticky folder owned by this account or root, as /tmp is, and
tolerates a statSync without options; tests whose cache sat directly in /tmp, and tests
that mock statSync, failed on both. Docs note that a record naming a live process of
another account holds the shared slot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(cli): skip the uid-based Chrome cache refusal cases where there is no uid
Windows has no process.getuid, so the ownership check never runs there and the two cases
that expect a refusal cannot hold.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(core): raise the unpacked size budget for the cache ownership checks
core grew to 4418013 bytes unpacked with foreignWritablePath, sealCacheInstall
and trustedCacheRoot; the budget follows the usual 5% headroom over that size.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): never write project records through symlinks under .hyperframes
The committed team descriptor (.hyperframes/project.json) and the HeyGen job
log (.hyperframes/heygen-jobs.jsonl) were written with path-based mkdir/write/
append calls that follow links, so a cloned project could point either file,
or .hyperframes itself, at a file outside the project and have publish or a
paid HeyGen job overwrite or append to it.
- projectRecords: add appendRecord (O_NOFOLLOW open, same plain file checked
by fstat/lstat, real folders from .hyperframes) and recordKind.
- writeTeamProject goes through writeRecord (exclusive temp + rename), so a
link at project.json is replaced, never written through, and a linked
.hyperframes is refused.
- readTeamProject refuses a linked or non-plain descriptor (or a linked
.hyperframes) instead of reading it as absent, which stops publish before it
would write one.
- The job log and project voice picks go through appendRecord and fail with
a Fix when refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): route every remaining .hyperframes write through link-refusing helpers
Writers that still made or wrote into a project's .hyperframes with
path-based mkdir/write calls now refuse a linked or non-folder .hyperframes
and never write through a link:
- project lock: folder via recordFolder (the exclusive lock create stays)
- lambda stack state: writeRecord
- received-files record: recordFolder + writeRecord, so a link to another
project file is replaced instead of overwritten
- check reports: folder via recordFolder (the wx create stays)
- check and snapshot frame folders: new projectFolder, which holds a folder
inside .hyperframes to recordFolder and makes one elsewhere as asked
- template pull staging: recordFolder
- share render record: writeRecord
- studio server state folder: recordFolder
- share storage: the atomic JSON temp file is opened with wx
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): drop doc comments the comment ratchet counts
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): refuse to append to a project record with more than one hard link
A hard link at .hyperframes/heygen-jobs.jsonl names an outside file as a
plain file, so the O_NOFOLLOW open and the same-file check let the job log
append to it. The append now refuses a file whose link count is above one.
Whole-file record writes go through a temp file renamed into place, which
replaces the name rather than writing to a linked inode.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): check and info --media leave remote media sources unprobed
mediaPath handed an http(s) src to ffprobe, so check's HDR and sound
probes and info --media's clip lengths opened composition URLs from the
CLI process. Remote sources are now left unprobed, as timeline and
check's clip audit already do. check's report lists the remote video
and image sources HDR inspection skipped under hdr.remoteNotInspected,
and a remote video that may carry sound keeps the film from being
reported as silent.
* fix(cli): drop credentials from remoteNotInspected and document unprobed remote media
* fix(cli): open starts the desktop app in the home folder, not the folder it runs in
* style(cli): keep launchApp's note on one comment line
* fix(cli): open starts the desktop app without the project's .env values
* fix(cli): studio thumbnails and Ctrl-C cleanup under a restricted page network
Studio's thumbnail browser now gets the preview's server as its page server, so
with PRODUCER_RESTRICT_PAGE_NETWORK on the proxy no longer refuses the Studio's
own preview. A restricted CLI browser's proxy and temp profile are released by
the CLI's stop-signal handler, which waits for them before exiting.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(cli): every composition Chrome launch goes through the page-network fence
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): keep the follow-up inside 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>
* fix(studio-server): a host that built a page on the build worker can exit
* test(studio-server): a dead build worker's exit code is left to the platform
* refactor(studio-server): every build-worker queue change sets its ref in one place
* fix(cli): commands that open a composition honor PRODUCER_RESTRICT_PAGE_NETWORK
check, snapshot, compare, grade-compare, layout, validate and the motion shot
launched Chrome with no page-network restriction, so a composition page
could reach loopback, the LAN and link-local addresses even when the caller
set PRODUCER_RESTRICT_PAGE_NETWORK=true, as a render honors it.
Every composition browser now launches through launchWithPageNetwork. When
the flag is on, it starts the engine's public-only capture proxy and profile,
the same ones a render uses. The page server is the only direct host, and the
proxy and profile live as long as the browser. The prelaunch path, which
starts Chrome before the page server exists, is skipped when the network is
restricted. A restricted launch without a page URL throws instead of running
unrestricted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): keep the page-network change inside 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>
* 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