You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Bug]: Long recordings open stretched in the editor: screen dimension probe times out after 5 s, so asset.video is never set #1610
After a long recording, the editor stretches the screen video to fill a 16:9 frame instead of keeping its real aspect ratio. Changing the editor's aspect ratio only reshapes the outer frame; the screen footage inside stays stretched. Short recordings made on the same machine and window look correct.
The recorded MP4 itself is fine (2816×2064, undistorted frames). The difference is in the project file: the short recording's asset has
while the long recording's asset has no video entry at all.
Likely cause:probeVideoDimensions in src/lib/ai-edition/timeline/duration.ts loads the file into a hidden <video preload="metadata"> and gives up after DEFAULT_TIMEOUT_MS = 5000, resolving null. In useTimeline.ts (~line 297) a null result just leaves entry.video unset, and the asset ID is already in probedAssetIdsRef, so it isn't retried in that session. The Linux recorder writes fragmented MP4 (ftyp, moov, then thousands of moof/mdat pairs with no sidx/mfra). For an 11-minute, 2 GB file, Chromium doesn't reach loadedmetadata within 5 s, so the dimensions are never stored and the editor falls back to its default frame.
Adding the video entry to the project file by hand (with the editor closed) fixes the stretch for that project.
Possible fixes:
Get the dimensions from the recorder, which already knows them when it stores the capture, and write them into the asset at creation.
And/or raise the probe timeout or scale it with file size, and retry on later opens instead of marking the asset as probed when the probe failed.
And/or write a sidx (or a final mfra) when the recording is finalized, so players can read the metadata without scanning fragments.
Expected behavior
The screen recording keeps its real aspect ratio in the editor and export, whatever the recording length.
To Reproduce
On Linux (KDE Plasma Wayland), record a window for about 10+ minutes (here: 11 min 19 s, 2.0 GB, 2816×2064 H.264 via VA-API).
Open the recording in the editor.
The screen footage is stretched to 16:9; changing the aspect ratio doesn't fix it.
Look at the project's .openscreen file: assets[0] has no video field.
Make a 15-second recording of the same window: it opens correctly, and its asset has video.width/height.
Screenshots
Screenshot to follow in a comment.
Environment
OS: Linux, Fedora 44 (kernel 7.2.14-200.fc44.x86_64), KDE Plasma on Wayland
OpenScreen: 2.2.2 (RPM). Also seen on 2.3.0-rc.5, where even a 23-second recording opened stretched.
Displays: 5120×2160 at 1.6× scale (recorded monitor) and 3840×1600 at 1.3× scale
Capture log of note:window-crop-missing: a window was granted but no crop rectangle arrived in 508 ms; recording the full 2816x2061 stream instead (from the rc.5 session)
Thanks again for the detailed report. #1618 is now in main: the background dimension probe waits up to 60 s and retries a failed probe, so long takes should no longer open stretched. It ships in the next release.
One open question: your 23 s take reads in about 260 ms on our side, so its failure may have another cause. Once you are on a build with #1618, could you tell us whether that take still opens stretched? If it does, the editor log from opening it would help.
Closing this as fixed in main by #1618: long takes no longer lose their dimensions, and the fix ships in the next release. If a take, including the short 23 s one, still opens stretched on a build with #1618, please reopen with the editor log from opening it and we'll take it from there. Thanks again for the thorough report.
Activity
kaenamiller commentedon Nov 17, 2026
Screengrab from raw recording for comparison
kaenamiller commentedon Nov 17, 2026
Stretched recording in editor (persists into export)
fix(editor): count only failed dimension probes and reserve the batch
fix(editor): treat a path toFileUrl cannot encode as a failed probe
fix(editor): retry the dimension backfill and give it time on long takes
fix(editor): count only failed dimension probes and reserve the batch
fix(editor): treat a path toFileUrl cannot encode as a failed probe
fix(editor): serialize the dimension backfill's save so overlapping b…
fix(editor): keep the backfill save inside useTimeline for the write …
fix(editor): retry the dimension backfill after a failed save
fix(editor): bound the dimension backfill save so a hung save cannot …
fix(editor): retry the dimension backfill and give it time on long takes
fix(editor): count only failed dimension probes and reserve the batch
fix(editor): treat a path toFileUrl cannot encode as a failed probe
fix(editor): serialize the dimension backfill's save so overlapping b…
fix(editor): keep the backfill save inside useTimeline for the write …
fix(editor): retry the dimension backfill after a failed save
fix(editor): bound the dimension backfill save so a hung save cannot …
EtienneLescot commentedon Nov 19, 2026
Thanks again for the detailed report. #1618 is now in main: the background dimension probe waits up to 60 s and retries a failed probe, so long takes should no longer open stretched. It ships in the next release.
One open question: your 23 s take reads in about 260 ms on our side, so its failure may have another cause. Once you are on a build with #1618, could you tell us whether that take still opens stretched? If it does, the editor log from opening it would help.
EtienneLescot commentedon Nov 21, 2026
Closing this as fixed in main by #1618: long takes no longer lose their dimensions, and the fix ships in the next release. If a take, including the short 23 s one, still opens stretched on a build with #1618, please reopen with the editor log from opening it and we'll take it from there. Thanks again for the thorough report.