discardLatestPending() popped whichever batch happened to be at the
back of the queue. With a Stop → Record → Discard sequence, the
pending queue can have recording B's batch sitting in front of A's by
the time A's finalize callback resolves (because finalizeRecording
awaits fixWebmDuration), so the discard targets the wrong recording.
Tag each completed batch with the recording id supplied at
startSession() time and replace discardLatestPending() with
discardBatch(recordingId). takeNextBatch() now returns the full
{recordingId, samples} shape so prependBatch() can re-queue it on
write-failure without losing the id. The renderer already owns a
stable recordingId (Date.now() in useScreenRecorder) and the IPC
surface threads it through set-recording-state and
discard-cursor-telemetry.
Adds a regression test that mirrors FabLrc's scenario in PR #457:
two recordings finalize, A is discarded after B has already been
queued, and the buffer must drop A while keeping B intact.
- Replace anonymous Error in resolveImageWallpaperUrl with typed
UnsafeImagePrefixError, mirroring UnsafeAssetPathError so cause
chains stay discriminable.
- Replace `(err as BackgroundLoadError).cause` casts in wallpaper
tests with instanceof narrowing (no `as` per project rules).
- Remove unused `WALLPAPER_PATHS` re-export from projectPersistence;
consumers import directly from @/lib/wallpaper (SSOT).
Reviewer audit found two real risks in the prior amendment:
1. LEGACY_FILE_WALLPAPER_RE was too permissive. Any file:// URL
containing /wallpapers/wallpaperN.jpg would match — including a user's
own file at /home/me/wallpapers/wallpaper1.jpg that happened to share
the name pattern. Silent data-loss potential: user's photo replaced
with a bundled asset. In-app upload flow uses data: URIs today so it
can't actually produce such a value, but the regex should be tight
on intent. Now requires a known install-layout segment:
resources/[assets/]wallpapers/ (packaged) or public/wallpapers/ (dev).
2. No upper bound on \d+. A corrupted or future-schema project with
wallpaper99.jpg was silently rewritten to /wallpapers/wallpaper99.jpg
which 404s. Now validates against WALLPAPER_PATHS; out-of-set
bundled-looking values fall back to DEFAULT_WALLPAPER.
Also applied R2.2 defensive guard: resolveImageWallpaperUrl's catch
block now checks instanceof BackgroundLoadError and rethrows unchanged
instead of wrapping a second time. Current getAssetPath cannot throw
BackgroundLoadError so this is a future-proof against refactors.
Tests: 56 pass (up from 54). Added coverage for "user file outside
install dir stays untouched" and "bundled-looking but out-of-set falls
back to default".
R1 — Persisted wallpaper is now always the canonical /wallpapers/wallpaperN.jpg
form, never the resolved file:// URL. Swatch clicks pass WALLPAPER_PATHS[i]
(the relative path) to onWallpaperChange; the resolved URL stays in
wallpaperPreviewUrls for rendering only. This prevents machine-specific paths
from being written into project JSON and avoids break-on-upgrade /
break-on-share regressions. Legacy projects carrying resolved file:// URLs are
rewritten by a new normalizer in normalizeProjectEditor:
file://…(/assets)?/wallpapers/wallpaperN.jpg → /wallpapers/wallpaperN.jpg.
R2 — resolveImageWallpaperUrl now catches anything getAssetPath throws
(UnsafeAssetPathError, AssetBaseUnavailableError) and rewraps as
BackgroundLoadError with the original as cause. Callers (videoExporter retry
loop, gifExporter catch, VideoEditor toast) only need one instanceof check and
users always see the translated errors.exportBackgroundLoadFailed toast.
R3 — src/vite-env.d.ts no longer duplicates Window.electronAPI. The interface
had drifted — renderer declaration was missing readBinaryFile, getPlatform,
revealInFolder, getShortcuts, saveShortcuts, hudOverlay*, countdown overlay
methods that electron-env.d.ts already declares. Removed the duplicate and
kept the triple-slash reference so the authoritative declaration is the one
in electron/electron-env.d.ts.
N1 — GRADIENT_RE accepts optional "repeating-" prefix so
repeating-linear/radial/conic-gradient values classify as gradients instead
of falling through to color.
N2 — displayBasename returns "(unknown)" sentinel for URLs without a
meaningful basename (file:///, bare /) instead of leaking the original string.
N3 — electron-builder.json5 extraResources block gets an inline comment
pointing at preload.ts:assetBaseDir so the bidirectional coupling is
discoverable from either file.
Tests: 54 unit tests pass (up from 35). New coverage for repeating
gradients, displayBasename sentinels, BackgroundLoadError cause wrapping,
legacy file:// wallpaper normalization (5 cases).
Every consumer of /wallpapers/*.jpg — SettingsPanel, VideoPlayback,
frameRenderer — was doing async IPC round trips, useEffect dances, and
Promise.all for a value that is a build-time constant per process. Each
consumer showed briefly-empty or briefly-404ing state on first paint
until the handler's reply resolved.
The asset base URL depends only on process.defaultApp and
process.resourcesPath / __dirname — all available in preload at
context-bridge time. Compute once there, expose as a sync string.
- preload.ts resolves baseDir (process.resourcesPath packaged,
<appRoot>/public unpackaged) and emits assetBaseUrl synchronously.
- get-asset-base-path IPC handler + main-process branching deleted.
- getAssetPath() is now sync. Returns string, not Promise<string>.
Throws AssetBaseUnavailableError (new) when electronAPI.assetBaseUrl
is missing — catastrophic preload failure, not silent 404.
- resolveImageWallpaperUrl() sync; same sync throw semantics.
- SettingsPanel: Promise.all + useState + useEffect collapse to one
useMemo. First paint has real URLs, no 18× ERR_FILE_NOT_FOUND, no
flicker.
- VideoPlayback: wallpaper-resolve useEffect collapses to useMemo.
- frameRenderer.setupBackground: drops the await.
- electronAPI type decls updated in both .d.ts files.
- 35 unit tests updated to reflect sync signature + new
AssetBaseUnavailableError contract.
Silent-fallback behavior from getAssetPath (returning /relative when
electronAPI failed) is gone. Renderers now surface preload failures
instead of rendering 404s.
Adversarial review surfaced four defects and four drive-bys. All applied:
B1 (security, MEDIUM) — Path traversal via encodeRelativeAssetPath.
encodeURIComponent passed "." and ".." through unchanged; percent-encoded
"%2e%2e" got decoded by the URL constructor. Either form escaped the
asset root: new URL("../../etc/passwd", "file:///opt/Openscreen/resources/")
→ file:///opt/etc/passwd. Reject both at src/lib/assetPath.ts via a new
UnsafeAssetPathError thrown when a decoded segment equals "." or "..".
B2 (correctness) — classifyWallpaper returned { kind: "image" } for
conic-gradient(...), rgb(...), hsl(...), oklch(...), empty string,
and named colors like "red". Old frameRenderer's bare fillStyle = value
handled these; new code would throw BackgroundLoadError with misleading
message. Classification now anchors on regexes, accepts all CSS color
functions and all three gradient types, treats unknown strings as
fallthrough color (old behavior), and normalizes "" to "#000000".
B3 (SSOT) — DEFAULT_WALLPAPER, projectPersistence.WALLPAPER_PATHS, and
SettingsPanel.WALLPAPER_RELATIVE independently hardcoded the same
/wallpapers/wallpaperN.jpg pattern. Three drift sites collapse into one:
WALLPAPER_PATHS lives in src/lib/wallpaper.ts, DEFAULT_WALLPAPER derives
from WALLPAPER_PATHS[0], projectPersistence re-exports from the canonical
module, SettingsPanel imports it directly.
B4 (privacy) — BackgroundLoadError.message and the translated toast
surfaced full file paths like file:///home/<user>/…/wallpaper.jpg —
leaks the user's home directory in copy-pasted bug reports. Added a
displayUrl getter that returns just the basename (or "data:…" for data
URIs), wired into the toast. Full URL remains in console.error and
error.url for debugging.
N1 — resolveImageWallpaperUrl now rejects image paths that don't live
under /wallpapers/ (throws BackgroundLoadError). Narrows the blast
radius of the returned <resourcesPath>/ base so the renderer can only
request files within the wallpapers directory, regardless of what the
project JSON claims.
N2 — videoExporter retry loop no longer calls cleanup() twice in the
BackgroundLoadError branch; the finally handles it.
N3 — Browser tests assert BackgroundLoadError.url contains the failing
path. Guards the {{url}} i18n interpolation contract.
N4 — VideoPlayback wallpaper resolve effect now catches resolver
throws (UnsafeAssetPathError, BackgroundLoadError from /wallpapers/
prefix enforcement). Prevents the new strict-rejection logic from
silently leaving the preview without a background.
Tests: 35 unit tests pass (up from 20); new coverage for all color
functions, all gradient types, empty string, named color fallback,
whitespace trimming, /wallpapers/ prefix enforcement, traversal
rejection, percent-encoded traversal rejection, displayUrl basename
and data-URI abbreviation.
Three independent defects plus one SSOT violation caused reported symptom
of image wallpapers rendering solid black in exported MP4/GIF while
appearing correctly in the editor preview.
Bug A — Dev-mode IPC handler returned <appPath>/public/assets/, but
wallpapers live at public/wallpapers/. No assets/ subdirectory exists in
source.
Bug B — FrameRenderer.setupBackground bypassed getAssetPath and did
window.location.origin + wallpaper, producing file:///wallpapers/*.jpg
404s in packaged Electron.
Bug C — setupBackground silently caught any background-load error and
filled black. Masked Bug B from the export pipeline; why the bug shipped.
Smell D — Asset layout asymmetric: public/wallpapers/ (dev) vs
resources/assets/wallpapers/ (packaged). assets/ subdirectory had no
other consumers.
Fixes:
- Unify asset layout. electron-builder extraResources now copies to
resources/wallpapers/ (no assets/). Main handler returns
<resourcesPath>/ packaged and <appPath>/public/ unpackaged. Same
convention in both modes: /wallpapers/x.jpg maps to <base>/wallpapers/x.jpg.
Nix package.nix mirror updated.
- New src/lib/wallpaper.ts module owns the wallpaper contract:
DEFAULT_WALLPAPER, classifyWallpaper (color/gradient/image), and
resolveImageWallpaperUrl (pure URL resolver, wraps getAssetPath).
BackgroundLoadError typed error for short-circuit detection.
- FrameRenderer.setupBackground uses the new helpers. Silent black
fallback removed; rethrows as BackgroundLoadError. Export pipeline
(VideoExporter + GifExporter) short-circuits encoder-retry loop on
BackgroundLoadError. VideoEditor catch site dispatches to translated
exportBackgroundLoadFailed toast.
- VideoPlayback editor preview consolidated onto the same helpers.
Three default-wallpaper path literals (useEditorHistory,
projectPersistence, VideoPlayback) collapsed onto DEFAULT_WALLPAPER.
- i18n: new errors.exportBackgroundLoadFailed key added to all seven
locales (en, zh-CN, zh-TW, es, fr, tr, ko-KR).
- Tests: 20 unit tests for wallpaper module (classifyWallpaper +
resolveImageWallpaperUrl branches + BackgroundLoadError).
videoExporter.browser.test.ts and gifExporter.browser.test.ts extended
with image-wallpaper happy path and BackgroundLoadError failure path.
Migration note: packaged users upgrading in place may retain an empty
resources/assets/ directory from the prior layout. Unreferenced at
runtime; cosmetic only. DMG/AppImage fresh installs get the new layout
directly.
Two follow-up fixes for CodeRabbit feedback on the docs commit:
- CursorTelemetryPoint JSDoc previously described cx/cy as 'device-pixel
positions'. The producer sampleCursorPoint() in electron/ipc/handlers.ts
clamps them to the [0, 1] range after dividing by the source display's
width/height, so they are normalised ratios, not pixel values. Correct
the doc comment accordingly.
- createCursorTelemetryBuffer now sanitizes maxActiveSamples and
maxPendingBatches: non-finite, zero, or negative values fall back to
safe positive-integer defaults. Without this, a caller passing Infinity
or NaN would hang the trim loops.
New test covers the sanitisation path for both options.
Add JSDoc to every public export in cursorTelemetryBuffer so the module
meets the 80% docstring-coverage threshold, and make two silent-drop
paths observable:
- endSession() now returns the number of pending batches evicted by the
maxPendingBatches cap and emits console.warn when any are dropped.
- prependBatch() defensively trims and warns if an unusual retry pattern
would push the queue past the cap (normal retry after takeNextBatch()
stays a no-op).
Tests cover both drop paths.
Address review feedback on #471 from @coderabbitai. The BMP-only
codepoint ranges missed two classes of characters:
- Non-BMP Han extensions (CJK Unified Ideographs Extension B, C, D, E, F)
such as 𠀀. A long string of Extension-B characters would still be
tokenized as a single unbreakable unit and overflow the box.
- Halfwidth Katakana (U+FF65-U+FF9F) such as ｶ. Same failure mode.
Switch to Unicode script property escapes (\\p{Script=Han},
\\p{Script=Hiragana}, \\p{Script=Katakana}, \\p{Script=Hangul}) which
cover these cases without enumerating ranges. tsconfig target is ES2020;
property escapes require ES2018+ so this is safe.
Verified coverage: 漢 あ ア 가 𠀀 ｶ all match; A and digits do not.
renderText split each line on whitespace, which works for Latin text
but leaves CJK strings as a single unbreakable token because CJK
scripts have no word-separating whitespace. Result: CJK annotation
text overflows the clipped annotation box even though the editor's
HTML preview wraps it correctly via CSS word-break: break-word.
Replace the ad-hoc whitespace split with a tokenizeForWrap helper
that emits each CJK character (Hiragana, Katakana, Hangul Syllables,
CJK Unified Ideographs + Extension A, and CJK Compatibility
Ideographs) as its own token, while keeping Latin words + whitespace
intact. The existing width-measurement wrap loop then handles CJK
per-character, matching the editor's behavior.
Closes#449
AudioProcessor.process and renderPitchPreservedTimelineAudio accepted
validatedDurationSec as optional, so the speed-aware path fell back to
media.duration when it was absent. HTMLMediaElement.duration can be
Infinity for the same MediaRecorder/Chromium Linux containers this PR
targets, which would make effectiveEnd and the playback stop checks
unreliable.
The only caller (VideoExporter.process) already threads
streamingDecoder's validatedDuration through, so make the parameter
required. Drop the media.duration fallback, the Number.isFinite guard
on readEndSec, and the two `!== undefined` checks in the tick loop.
While here:
- Document that +0.5 on readEndSec mirrors streamingDecoder.decodeAll's
read window so trim-only and speed-aware paths stay in sync.
- Replace the unreachable silent-blob fallback at the end of
renderPitchPreservedTimelineAudio with a loud invariant throw, so a
broken recorder contract surfaces instead of yielding empty audio.
The earlier NaN/Infinity guard collapsed both duration hints to 0 when
the container reported invalid values, which turned scanEndSec into
0.5s. The packet scan then read only the first half-second, scannedDuration
capped there, and validateDuration fell back to that wrong value for the
entire export — exactly the Chromium Linux case this PR is meant to fix.
Use a 24h sentinel as the read endpoint when no hint is usable. An
explicit end is still required (some containers are truncated without
one, per prior comment), but the sentinel is large enough to exceed any
realistic recording so the scan reaches real EOF.
Startup trim-skip only consulted the first active region at t=0, so
back-to-back or overlapping trims starting at zero (e.g. [0,500ms]
followed by [500ms,1000ms]) left the second region un-skipped. The
in-flight tick loop would catch it, but MediaRecorder was already
running by then, capturing up to one rAF frame of trimmed audio into
the blob and shifting the downstream timeline.
Loop findActiveTrimRegion from the advancing startPosition until no
region matches or startPosition >= effectiveEnd, bounded by
trimRegions.length for safety. Recompute initialSpeedRegion from the
final startPosition so playbackRate reflects the true start point.
mediaInfo.duration from web-demuxer can be NaN or Infinity on Chromium
Linux (same MediaRecorder bug this PR otherwise addresses). That value
flowed straight into Math.max + demuxer.read() as scanEndSec, producing
an invalid range argument and breaking the ground-truth packet scan.
Guard both mediaInfo.duration and videoStream.duration with
Number.isFinite before Math.max; validateDuration() already handled the
downstream use.
Drop redundant WebDemuxer.read() / getDecoderConfig() type casts while
here — the generics infer the chunk/config type from the media string
literal, so the `as ReadableStream<EncodedVideoChunk>` and
`as AudioDecoderConfig` are no-ops.
- validateDuration returns 0 instead of NaN when both container is
NaN and scanned is zero
- Use Math.abs for divergence check so container under-reporting is
also corrected (not just over-reporting)
Some containers are truncated when read() has no end bound.
Use container/stream duration + buffer as scan range, matching
the same pattern used in decodeAll().
Two bugs in the export pipeline:
1. Container duration from WebM metadata can be unreliable (Chromium bug
on Linux — reports Infinity, 0, or inflated values). The pipeline
trusted this value, causing inflated exports, frozen video, and
"decode ended early" errors.
Fix: scan actual packet timestamps in loadMetadata() and compare
against container duration. Use packet-based ground truth when they
diverge.
2. The speed-aware audio path (renderPitchPreservedTimelineAudio)
recorded in real-time via MediaRecorder but never paused recording
during trim-region seeks. Seek dead time was captured as audio,
inflating the audio track beyond the video duration.
Fix: pause MediaRecorder during trim seeks, skip past initial trim
before recording starts, wait for seek completion before resuming.
Fixes#276, #433. Partially addresses #428.
Address two issues raised during review:
P1 – When a recording is cancelled or restarted, setRecordingState(false)
enqueues its cursor batch but store-recorded-session is never called,
leaving a stale batch that contaminates the next recording's telemetry.
Add discardLatestPending() to the buffer and a discard-cursor-telemetry
IPC handler; the renderer now calls it on the discard path.
P2 – takeNextBatch() dequeued the batch before fs.writeFile, so a write
failure would permanently lose the telemetry. Wrap the write in
try/catch and re-insert the batch via prependBatch() on failure.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously, the main process kept two module-scope arrays —
activeCursorSamples and pendingCursorSamples — and set-recording-state
on a new recording wiped BOTH. When a user stopped recording and
immediately started a new one before store-recorded-session fired,
the previous recording's pending samples were discarded or later
overwritten with the new session's data, producing empty or mismatched
.cursor.json files.
Replace the two arrays with a small FIFO buffer
(createCursorTelemetryBuffer) that:
- Keeps pending batches per completed recording, never wiping them on
a new session start.
- Yields batches in arrival order to storeRecordedSessionFiles.
- Caps pending batches (default 8) so a never-stored sequence cannot
leak unbounded memory.
Unit-tested directly in src/lib/cursorTelemetryBuffer.test.ts, including
the rapid-restart race that motivated the change.
Add requiredEndSec calculation to properly handle early decode termination by using the last segment's end time. This addresses issues with export processing on Windows platforms.
Add requiredEndSec calculation to properly handle early decode termination by using the last segment's end time. This addresses issues with export processing on Windows platforms.
Sid