The workspace cutter drop check waited a fixed 250ms before asserting
the populated cutter controls, but the drop triggers asynchronous IPC
and media probing, so loaded runners regularly lost that race and
failed with a misleading error. The check now polls the exact same UI
conditions with a bounded deadline before asserting. All window
screenshots in the cutter and workspace smokes carry an explicit
generous timeout because large two-times-DPR captures exceeded the
Playwright default under load. The focused Electron smoke and the
cutter media matrix additionally retry exactly once in both CI
workflows, mirroring the established packaging retries for transient
runner turbulence, and the CI contract now enforces those retries
instead of the former single-line steps.
Document the background tool provisioning after installation and the
dedicated tool preparation queue status in the changelog and set the
release version.
A fresh installation only provisioned Streamlink and FFmpeg when the
first download started, which delayed that download by the full tool
download and extraction time. Packaged builds now schedule a background
provisioning pass fifteen seconds after startup so the tools are
typically ready before the first download begins, reusing the existing
ensure functions with their in-flight deduplication so a download that
starts earlier simply joins the running installation. The timer is
tracked and cleared during shutdown, and development and test launches
keep the previous deferred behavior so isolated environments never
trigger network provisioning.
On a fresh installation the first download repairs the managed
Streamlink and FFmpeg copies, which downloads and extracts roughly two
hundred megabytes before the actual VOD transfer begins. The queue item
previously sat on a generic starting status for that entire window - the
old preparing hint only appeared when the streamlink command had not
been verified yet, and a prior System Check verifies the command without
making the managed repair any cheaper, so users saw no explanation for
delays of a minute or more. All three download paths - VOD, live
recording and merge groups - now report a dedicated localized
"Preparing download tools" status before the tool gates run, and the
misleading conditional hint was removed.
The installer smoke asserted a clean registry and shortcut surface
immediately after the silent uninstall returned and the installation
directory disappeared. NSIS uninstallers run detached through a
temporary copy, so shortcut and registry removal completes after the
invoked process exits, and on current runner images the immediate check
raced that cleanup and failed with leftover HKCU keys and shortcuts.
The smoke now polls the exact same surface assertions until a bounded
deadline, keeping a genuinely dirty uninstall failing, with contract
coverage for the tolerated asynchronous cleanup and the persistent
failure case.
The live managed-tool smoke and the real installer smoke gated on
RUNNER_OS being exactly Windows. The Gitea act runner reports the value
as lowercase windows, so both gates rejected the approved self-hosted
git.24-music.de runner and the corresponding verify steps could never
run there. The gates now compare RUNNER_OS case-insensitively while
keeping every other identity requirement unchanged, with contract
coverage for the lowercase Gitea identity.
The installer staged downloaded tool archives as archiveName.uuid, so
the temporary file no longer ended in .zip. The Expand-Archive cmdlet
shipped with stock Windows PowerShell (Microsoft.PowerShell.Archive
1.0.1.0) rejects any archive whose extension is not .zip, which made
every managed streamlink and ffmpeg installation fail with
extract-failed on systems without an updated Archive module - including
plain end-user machines and the self-hosted CI runner. The unique
download path now uses the UUID as a prefix and keeps the manifest
archive name with its .zip extension intact.
assertLockedTargetFailure compared the rename paths from the product
publish-lock diagnostic against the expected output file with
path.resolve only. On runners whose temporary directory surfaces as a
Windows 8.3 short path the smoke passes a short-path output file while
the product logs canonical long paths, so a correct atomic publish lock
diagnostic was rejected and the cutter media matrix failed. sameResolvedPath
now canonicalizes both sides through the existing resolveCanonicalPath
helper with a resolve fallback, covered by a ShortPath contract test.
recoverInterruptedMergeArtifacts resolved the artifact root through
fs.realpathSync.native but compared candidate artifact paths with
path.resolve only. On systems whose temporary or download directories
surface as Windows 8.3 short paths the queue-persisted artifact paths
never matched the canonical long root, so crash artifacts were treated
as outside the owned directory, nothing was removed and interrupted
merge items were incorrectly blocked. isInside now canonicalizes both
sides through the deepest existing path segment, keeping planned but
not yet created paths comparable, and a Windows ShortPath regression
test covers the recovery flow end to end.
assertPathInside compared managed-tool paths with path.resolve only, so
on runners whose temporary directory surfaces as a Windows 8.3 short
path (RUNNER~1) the canonical long executable paths appeared to live
outside the owned root and the cutter matrix provisioning contract
failed. Resolve both sides through fs.realpathSync.native, walking up
through not-yet-existing segments so planned directories keep working,
and cover the contract with a real ShortPath regression test plus
mixed-existence canonicalization cases.
Document the tool availability fallback, the mid-transfer download
failure handling, the managed tools status hint and the unified
auto-install error messages in the 1.0.18 changelog, move the release
date to 2026-08-14 and add the new src/tools.test.ts suite to the public
release file inventory.
ensureStreamlinkInstalled and ensureFfmpegInstalled forced a managed
reinstall whenever a bundled tool directory existed without passing
manifest verification. When that reinstall failed (offline, blocked
download, locked executable), both functions returned false without ever
probing the tool that was already present and runnable, so every
download path aborted with an auto-install error while the System Check
kept reporting the same tool as available. Both functions now fall back
to the same executability probe the System Check uses after a failed
managed repair, cache the verified command on success, and only report
failure when nothing runnable remains. The repair attempt itself is
preserved and retried on the next call.
The tool archive download also registered no error handler on the HTTP
response stream. A stream that failed mid-transfer settled neither the
writer finish nor the writer error path, leaving the install promise
pending forever and the download start hanging indefinitely. The
response stream now rejects the download on error and destroys the
writer.
Merge group downloads reported the bare "Streamlink is missing." text
while the VOD and live paths already explained the failed auto-install.
All three gates now share streamlinkAutoInstallFailed, ffmpeg gained the
matching ffmpegAutoInstallFailed message, and the now unused
streamlinkMissing and ffmpegMissing keys were removed.
The Managed tools panel showed "Missing / Unverified" with no hint that
downloads still work through a system-provided installation, which reads
like a hard failure on machines that never installed the managed copies.
getManagedToolStatuses now reports per tool whether a runnable
installation answers the version probe, and the panel appends a
localized "Downloads still available" note whenever an unverified state
is covered by a working tool.
Canonicalize shortcut targets across Windows short and long path forms while preserving exact target validation. Keep Gitea 1.22 push verification parser-compatible and harden CI contracts against skipped or weakened live gates.
Harden update, system-check, queue, cutter, streamer and shutdown state transitions.
Add multi-user installer recovery, secret-safe config migration, provider fallback handling, managed-tool validation and real media export coverage.
Refresh the English public documentation, release notes and 1.0.18 product screenshot.
Use clearly synthetic streamer fixtures in the workspace Electron smoke so the public source release contains no real account handles while preserving display-case and navigation-motion coverage.
Keep the loaded cutter layout stable across supported window sizes, preserve recovered encoder choices during capability discovery, and reject unsupported video selections without changing the active project.
Honor the Windows system theme, fully localize runtime metrics, invalidate imported System Check state safely, and bound child-process shutdown waits when close events never arrive.
Extend focused and Electron smoke coverage and prepare the v1.0.17 public release metadata.
Keep download progress inside the updater popover, slow changelog transitions, and improve dark-theme checkbox contrast. Expand diagnostics, prevent cleanup option clipping, improve navigation readability, and cover the affected states in Electron regression tests. Prepare version 1.0.15 release metadata.
Retry Electron binary provisioning before Windows smoke tests and retain bounded retries for package and installer builds. Bump release contracts to 1.0.14.
Retry the directory package once when Electron download transport is interrupted, keeping the Windows CI step bounded. Bump the release contracts to 1.0.12.
Mock Electron in the Node-only secure-storage unit test so its expected unavailable-path assertion never triggers Electron's binary bootstrap. Bump the patch release contracts to 1.0.11 after the 1.0.10 tag CI failure.
Ignore malformed update notifications that do not carry a version, preventing unusable release prompts in installed builds.
Keep the loaded cutter workspace free of the source selector and give cutter export selects an explicit non-repeating indicator after workspace styling.
Add regression coverage, bump the patch release to 1.0.10, and update public installation documentation.
Sucukdeluxe