Repository navigation
OpenCut v1.56.0
OpenCut v1.56.0 is mostly fixes for the Windows installer build. The packaged server now ships the generated manifest files it was missing, stops loading packages from another Python install on the PC, and won't start a second server on top of one that's already running. Issue reports, crash packets and log exports include the server log again, with passwords, keys and tokens taken out before anything leaves your PC. Pillow moves to 13 for a batch of image-parsing security fixes, exported captions stop dropping words, and EDL timecode comes out right at 23.976, 29.97 and 59.94.
Reports this release changes something for: #7 (RTX 50-series card reported as unavailable), #8 (missing route_manifest.json and "service unavailable" after install) and Discussion #10 (the health check failing on a broken torch install).
Changes
Security: Pillow 13 and newer floors for seven other packages
- OpenCut now needs Pillow 13. It fixes 43 advisories published after
12.3.0, five of them critical heap overflows in the JPEG, FlashPix and TIFF
code. Masks, mattes and thumbnails come out the same, since every place
OpenCut turns a NumPy array into an image already hands it 8-bit data. - The minimums went up for urllib3 (2.8.0), Werkzeug (3.1.9), certifi,
filelock, setuptools,mcpandrembg. certifi and filelock used to come
in only through huggingface-hub. They're declared directly now, so their
floors can't slip.docs/PYTHON_ADVISORIES.mdlists the advisory behind
each floor. - When background removal installs rembg for you, it now asks for 2.0.82 or
newer too, GPU build included. Before, that install took whatever version
the package index offered. - The release locks and
pylock.tomlpin the fixed versions. - Packaged builds use PyInstaller 6.22.3. Version 6.20.0 still looked for
SciPy's array helpers under their old pre-1.18 name, so in the packaged
server anything built onscipy.ndimagefailed, object removal included. - The release lock takes msgpack 1.2.4. Its C extension in 1.2.1 leaked
memory each time it rejected a map key (CVE-2026-114613). npm auditflagged 12 packages in the panel's build and test tools, 8 of
them high. Vite 8.3.5, Vitest 4.1.12 and pinned fixed versions of
minimatch, brace-expansion, js-yaml and source-map-js clear all of them,
and nothing is on the allow list.
Fixed: a broken optional package no longer takes down the server
- If PyTorch was installed but its DLLs wouldn't load, it raised
OSError
instead ofImportError. The server then wouldn't start at all, and the
capability check behind/healthfailed the same way. - Any optional package that fails to import now just shows that feature as
unavailable. OpenCut tries a broken one once per run, not on every health
poll (a secondimport torchafter a failed one makes things worse). GET /system/check-failureslists each one asimport:<module>with the
error it raised, so you can see why a feature is missing. A package that was
only missing a dependency gets another try after you install something.- The health checks for WhisperX, Kokoro, Edge TTS, Pedalboard, DeepFilterNet,
OpenCV, MediaPipe, InsightFace and DaVinci Resolve's scripting module now
get the same treatment. Before, a broken Kokoro also made Edge TTS show as
missing. - When several requests probe the same broken package at once, it's imported
one time instead of once per request. A check that lands while another
thread is still loading a package waits for it rather than getting a
half-loaded copy. Two packages that load each other from different
threads can't lock each other up, and a package that loads fine after you
install its missing dependency shows as available straight away. A
package that puts off loading until it's first used gets the same
treatment, instead of its error escaping the check or locking up every
check after it.
Fixed: issue reports and crash packets carry the server log again
- The server writes
~/.opencut/server.log, but the issue report, the crash
packet and the job diagnostics all read~/.opencut/opencut.log, a file
nothing writes. Every bundle went out with no log at all, and the help text
sent people looking for that missing file too. - All of them now read
server.log. When the live file is short because it
just rotated, the older lines come fromserver.log.1and the files after it. - Settings > Export logs used to download only the crash log. It now saves one
text file with the recent server log and the crash log, with your home
folder shown as~. - Before any of those leave your machine, passwords, API keys, bearer tokens,
SRT passphrases and packager keys that ended up in the log are replaced with
[REDACTED]. The issue report already did this, but the crash packet, the
log export and job diagnostics only hid your home folder. - That redaction now reads each log line as the JSON it's stored as, so a
header in the middle of a message, a quoted password and a home folder
written with doubled backslashes are all caught. It also catches
Authorization: Token ..., URLs with a password in them (encoded or not),
--api-keyand-passphraseflags, names likellm_api_keyor
llmApiKey, header lists and request environments, and passphrases with
spaces in them when they're quoted or encoded. Frame patterns like
img_%04d.pngand URLs with a port come through untouched. - Job diagnostics keep a job's parameters as real data instead of a string,
with secret values swapped out. The live log view in Settings is redacted
the same way as the export, and its job filter only matches the redacted
text. - More forms are caught too:
token = '...',api_tokenandoauth_token
in a query,HTTP_AUTHORIZATION=from a request environment, a list of
passwords,-encryption_key, a URL that was encoded twice, a login URL
passed inside another link's query, a query password holding an@, and
a camera URL whose password starts with digits, holds an@, ends in/
or runs past 1,024 characters. So are Negotiate, NTLM and Digest
credentials, cookie lists (even aftercookie=on.), the byte-string
headers an ASGI server logs and atoken:right after a logger name or a
file path. A CORS header list keeps its other names,hasPassword=true
andhas_api_key: falsestay readable, and an IPv6 address is no longer
taken for a password. - Ordinary text stays as it was: a CDN link with a port and
logo@2x.png,
cookie=onorcookie=off: disabledand the rest of its line, an email
address in a link's query, tokenizer settings likebos_token
andnext_page_token, a parser'sUnexpected token: puncand a drawtext
Token: 5caption. A bare host with a port and an@in the next path
segment (nas:445/media@archive) still reads as a password, since it
looks just likeadmin:2024/pw@cam. - A long log line used to take up to 18 seconds to clean. It's now a small
fraction of a second, including lines with a repeatedeyJ-ortoken:[. - Job diagnostics still come out when a job parameter can't be printed, and
the live log view no longer splits a record at an unusual line separator.
Fixed: SRT and VTT captions no longer lose words
- A cue that fit the character budget on paper could still need a third line
once it wrapped at word boundaries, and the exporter cut that line off with
"...". On a 23-word sentence the SRT came out with 21 words. The shot-aware
subtitle cleanup dropped extra lines the same way, without the ellipsis. - Cues are now sized by how they actually wrap. A cue that needs more lines is
split into another cue at a word boundary, timed from the word timestamps
when there are any and by character count when there aren't. A word too long
for one cue is split across cues instead of being clipped. - Positioned WebVTT from the broadcast caption export splits long cues the
same way. It used to keep the first two lines and drop the rest, so one
23-word caption came out with 11 words. - When shot-aware cleanup splits a cue, a short leftover piece no longer
flashes by. Each piece keeps the profile's minimum duration (0.833 s for
Netflix) when the cue is long enough for that, and the pieces split the time
evenly when it isn't. - A cue time a fraction of a millisecond short of a whole minute could be
written as00:00:60.000, which players reject. Positioned WebVTT, TTML,
the SDH, multi-language and shot-aware SRT and VTT exports, project note
exports and the silence summary now round the whole time first, so it
reads00:01:00.000. A very long time keeps its exact milliseconds, and
half a millisecond rounds up.
Fixed: EDL and ADR timecode at 23.976, 29.97 and 59.94
- The EDL export, the transcript edit EDL and the ADR cue sheet counted
29.97 as 29 frames per timecode second (23.976 as 23, 59.94 as 59). An hour
of 29.97 footage ended at01:02:00:12instead of00:59:56:12, so late
events landed about two minutes off while the header said non-drop. - All three now share one converter that counts frames at the real rate and
labels them at 24, 30 or 60. EDL import reads them back to the same frame.
A rate that has no SMPTE timecode, like 12.5, is refused with an error
instead of being written wrong. - Auto rough cut, template assembly and the review bundle's marker EDL and
Premiere marker CSV had their own copies of the old math and now use the
same converter. At a rate with no timecode, or past 99 hours, rough cut and
template assembly still put the edit together and only leave out the EDL.
A frame rate of 0, below 0 or not a number is refused before any media is
read. Review markers still get a label at 12.5 fps, and at 100 fps and up
or past 99 hours their frames are counted at the real rate. At 24, 25,
30 and other whole-number rates their frames now round to the nearest
frame instead of always rounding down, so some events move by one frame. - The multicam timecode sync EDL uses the same converter, so it refuses
12.5 fps and counts past 24 hours too. Rates of 100 fps and up are
refused, since their frame numbers don't fit an EDL. A time that isn't a
number, or that runs past 99 hours, is an error instead of00:00:00:00. - EDL import refuses a timecode with a three-digit frame field instead of
reading00:00:01:119as frame 11. A 120 fps EDL with two-digit frames
still imports. A minute, second or frame that's out of range is refused,
and the error names the event it's in. - A transcript edit on footage whose rate has no timecode (12.5 fps, or an
odd phone rate) renders again. It used to fail at the very end over the
EDL. Now it skips just that file and still writes the OTIO sidecar. The
same goes for an EDL that can't be written, and an EDL left over from an
earlier render of the same name is removed. - Timecode past 24 hours keeps counting (
25:03:04:05) instead of wrapping
back to01:03:04:05, so a long timeline reads back to the right frame. - The other writers that still did their own 30 fps math count at the real
rate now. That covers the AI rough cut EDL (at the rendered file's rate),
the transcript editor's EDL export, the paper edit EDL and text list (at
the clip's rate when the request names one), the narrative assembly's
edit list, and SCC and MCC captions, which always run at 29.97. With a
rate that has no timecode, the rough cut skips its EDL, the EDL exports
refuse before creating a file, and the text lists fall back to clock time.
The same happens past 99 hours. A transcript edit builds its EDL before it
renders anything, so a cut or pause that would land past 99:59:59 is
refused first. - SRT times from the broadcast caption tools round to the millisecond. 2.3
seconds used to come out as00:00:02.299.
Fixed: the packaged server shipped without its generated manifests
- Every release built with PyInstaller was missing all sixteen files under
opencut/_generated, becausecollect_data_filesgathers one subpackage at
a time and the spec named onlyopencut.data. Route validation, feature
readiness, extended MCP tools and the CLI route allowlist all fell back
to defaults, and the only visible sign was one warning about
route_manifest.json. Reported in issue #8. - The spec now derives the data-bearing subpackages from the source tree, so a
new one cannot be forgotten, and the server reports every missing manifest at
startup instead of degrading in silence. A test compares the built artifact
against the source tree.
Fixed: saved editing defaults are applied again
- A survey keyed on the stored file rather than the accessor found seven
settings under~/.opencutthat reached no consumer. Changing the chapter
count, the minimum multicam cut length, the footage-index transcription model
or the loudness target in Settings saved the value, served it back, and then
the job used a literal instead. - Those four now reach the work. Two more, colour-match profiles and auto-zoom
presets, have no reader anywhere on either side and are recorded for removal
with a migration that cleans them off disk. - A new registry states what every persisted setting is for, and a test fails
when a stored setting has no consumer or a route stops reading one.
Fixed: the live-updates bridge reported success before it was listening
- Starting the bridge answered "running on port 5680" the moment the thread was
spawned. If the port was taken, the bind failed a moment later, the error was
logged and dropped, and the panel sat waiting on a socket that never existed.
/ws/statusreported a plain "stopped", so a port collision looked the same
as a bridge nobody had started. - Starting now waits for the bind and reports what happened, including the port
actually bound rather than the one requested. A failure comes back with its
reason and stays readable afterwards. The CEP panel connects to the reported
port instead of assuming 5680, which mattered because the backend searches
5680-5689 and skips the HTTP port.
Fixed: the release provenance document described a policy the code had abandoned
- It told readers a tagged FFmpeg release
>= 8.1.1was acceptable on an open
release lane. The code had closed that lane outright, raised the floor to
8.1.3, moved the snapshot floor to 2026-07-06, and the installers had moved to
a 2026-08-03 git snapshot. Four statements, all wrong, and the test guarding
the document asserted the stale strings literally so it kept passing. - The document now states the real policy, and
python -m opencut.tools.check_provenance_docs --checkcompares it against
the code on every release. Changing a floor, a lane, or a bundled pin without
updating the document fails the gate and names the field. The same check
covers advisory waivers and the recorded dependency floor raises.
Fixed: the console no longer warns about a production deployment
- A loopback server printed "WARNING: This is a development server. Do not use
it in a production deployment." on every start, plus a second copy of the
address OpenCut had already printed. One user pasted that into a bug report
about something else entirely. - Loopback traffic still goes through the same threaded server, now built
without the banner./system/statusnow reports which server is running, so
a bug report can say. Remote binds still use Waitress.
Added: the Adobe version snapshot can go stale loudly
adobe_premierepro_versions.jsonrecords when it was captured and nothing
ever read it, so the file sat untouched for months while the UXP migration
was planned against it. The release gate now fails when it is more than 90
days old, and distinguishes that from a snapshot whose upstream fetch failed.- The snapshot is current again (Premiere Pro 27.0.0, beta 27.1.0), and the
UXP panel is now checked against the 27.0.0 type definitions and Adobe's
27.0.0 lint rules instead of 26.3.0. - Those rules caught silence removal in the UXP panel holding on to a clip
selection after Premiere'screateEmptySelectioncallback returned, which
the 27.0 typings say isn't valid. The selection is now filled and used
inside the callback, under the same project lock, and a callback that
arrives late changes nothing.
Added: the UXP panel now installs
- Every installer shipped only the CEP panel while the README advertised a UXP
one for Premiere 25.6+. Adobe ends ExtendScript support in Premiere Pro in
September 2026 and the CEP panel routes every host mutation through it, so the
only installable panel was the one being retired. install.py,Install.ps1, the Inno Setup script and the Windows installer
all place the UXP panel now, and all of them remove it again on uninstall.
python -m opencut.tools.build_uxp_ccxbuilds a distributable package.- Premiere ignores a plugin installed this way until you enable Settings >
Plugins > Developer Mode, which is a preference no installer can set. The
README says so. Signed.ccxdistribution through Creative Cloud needs an
Adobe identity this project does not have yet.
Fixed: the plugin registry pointed at a namespace the project does not own
- The marketplace fetched its index from
raw.githubusercontent.com/opencut/....
That GitHub organisation exists and belongs to someone else, and the
repository did not exist at all, so whoever created it would have become the
plugin index for every OpenCut install. - The default now sits in a namespace the maintainer controls, and the index
itself has to carry a signature from a key shipped with the release before a
single entry is read. An index that fails to verify is refused and does not
replace the last one that passed. Since each entry supplies its own publisher
key, and publisher trust is granted on first use, nothing installs from an
index that was not authenticated.
Fixed: a packaged server no longer borrows another Python's packages
- The frozen build used to run the first
pythonit found on PATH and import
whatever site-packages that interpreter reported. When the versions did not
match, a native module compiled for the wrong Python aborted the process with
no traceback, which is how the server "just stopped" in issue #8. It also
meant any writable folder on PATH containing a python.exe got executed at
startup. - Optional dependencies now come from the bundled runtime and
~/.opencut/packages. Borrowing an external interpreter is opt-in through
OPENCUT_EXTERNAL_SITE_PACKAGES, and even then its Python version has to
match the build's.
Fixed: add-on installs in the packaged app landed where the server couldn't load them
- The packaged server imports add-ons only from
~/.opencut/packages, using
its own bundled Python 3.13. Installing one, Whisper for example, ran the
firstpythonon PATH. A 3.12 interpreter filled that folder with 3.12
wheels the server then crashed importing, and the plain and--user
attempts put packages where the server never looks. With no Python found at
all, it ran the server executable itself in place of pip. Also reported in
issue #8. - Installs now look for a Python with the same minor version, the
py
launcher first, and install only into~/.opencut/packages. When there isn't
one, the error says which version to install. Reinstalling Whisper no longer
uninstalls anything from your own Python. - The faster-whisper fallback that upgrades pip first handed pip, setuptools
and wheel to pip as one name, which pip rejects, so that step never ran. It
names them separately now, and the packaged app skips it rather than putting
a copy of pip in~/.opencut/packages.
Fixed: OPENCUT_FFMPEG_DIR pointing at the FFmpeg folder was ignored
- The Flatpak and AppImage launchers and the Linux build script set
OPENCUT_FFMPEG_DIRto the folder that holdsffmpegandffprobe, but the
server only looked for anffmpegfolder inside it. Those packages still
found the right binaries because the launchers also put that folder on PATH.
Setting it by hand did nothing. - It now takes either the folder with the binaries or a folder that has an
ffmpegfolder inside.
Added: crash records
~/.opencut/crash.logwas read by the crash packet, the issue reporter and
the crash-log download, and written by nothing. An unhandled exception in any
thread now leaves a record there, and a native crash leaves a faulthandler
traceback, both with the loaded paths that would explain a bad import.
Fixed: GPU install instructions named a CUDA build too old for current cards
- The README, requirements.txt, the source installer and the FlashVSR hint each
pointed at the CUDA 12.1 wheel index, which has no kernels for RTX 50-series
cards, so following the documentation produced the exact failure in issue #7.
All four now come from one place and name a build that covers current
hardware, with a test that fails if they drift apart again.
Fixed: two servers could bind the same port on Windows
- The port check set
SO_REUSEADDRbefore binding, which on Windows also
allows binding over a socket another process is actively using. A busy port
read as free, a second server started on it, and the second overwrote the PID
file that pointed at the first. The panel saw the result as intermittent
"service unavailable". Also reported in issue #8. - A live listener is now detected directly, TIME_WAIT sockets left by a killed
server are still reusable, and OpenCut refuses to start a second server
against a running one rather than moving to another port and sharing its
databases.
Fixed: a supported GPU reported as unavailable
- Selecting an RTX 50-series card failed every job with "GPU index 0 is not
available", while listing index 0 as available in the same sentence. The
adapter was fine. The installed PyTorch build had no kernels for it.
Reported in issue #7. - Adapters now carry their compute capability on both detection paths, and
OpenCut checks it against the architectures the installed build was compiled
for. An adapter the build cannot run is reported as such in Settings before a
job starts, and the message names the wheel that would work instead of
suggesting a different device. ONNX Runtime is no longer asked for a CUDA
provider that a CPU-only install does not have. - That check by itself still failed every job on such a card, FFmpeg-only work
and faster-whisper captions included, though neither one touches PyTorch.
Only PyTorch work is affected now. With the GPU left on automatic it runs on
the CPU and logs why once, faster-whisper and ONNX Runtime keep the card, and
an explicit selection gets an error with the install command in it. In the
packaged app that command installs into~/.opencut/packageswith the
matching Python, since a plainpip install torchnever reaches it.
Fixed: feature checks crashed on a refused FFmpeg
- When FFmpeg doesn't clear the security floor, looking up its path raises.
Several availability checks (loudness validation, deflicker, quality metrics
and VMAF, APV, VVC and SRT output) let that error escape instead of reporting
the feature as unavailable. They report it unavailable now. The ones that
cache their answer skip caching this one, so a replaced FFmpeg is picked up
without a restart.
Verification
The release gate ran 45 checks on the tagged commit and none failed. That covers 1,226 Python tests, 48 media conformance checks, 276 panel unit tests and 72 rendered panel tests. The standards check gave one warning, because the IMF and IMSC validators aren't installed on the build PC, so it skipped 9 cases. pip-audit and npm audit found no open advisories, and msgpack is now 1.2.4 for CVE-2026-114613. The 423 tests that cover redaction in logs and crash reports all pass.
The server was packaged from that same commit in a clean environment installed from the hashed release lock. All 16 generated manifests are in the bundle, and the packaged server started on its own and answered its health check, the startup failure check and the API spec.
Install
Download OpenCut-Setup-1.56.0.exe and run it. The installer isn't code-signed yet, so Windows SmartScreen will warn the first time. Click More info, then Run anyway. release-digests.json lists the SHA-256 of every file here if you'd like to check your download first.
SHA-256
| File | SHA-256 |
|---|---|
OpenCut-Setup-1.56.0.exe |
078ba78b5f6f0b4d7ad6c0d68b4279f41bec5cb740bb5f2c56df0d73f219f7e3 |
payload.zip |
4f598d32a4cd660b56a37d49b633924b88402926bfbc36000717688e2f1d47f5 |