The crate's build.rs unconditionally adds the MSVC + Windows SDK
include dirs on Windows (for MSVC hosts); on the GNU toolchain that
breaks the bridge compile with MSVC-only headers. The runner's job
hook re-exports INCLUDE/LIB per step, so the in-step unset did not
help — patch the extracted build.rs instead (both the env-var failure
modes are now documented in the step comment).
- the warp runner's job hook re-exports MSVC INCLUDE/LIB per step, so
the GITHUB_ENV clear did not stick — unset in the Build/Test steps
themselves
- node_e2e_test segfaults only on the Linux runner; rerun the binary
under gdb on failure to capture the native stack
Each distro package builds inside that distro's container so declared
dependencies always resolve to native names: hand-rolled deb via
dpkg-shlibdeps + dpkg-deb, rpm via rpmbuild's auto-requires, Arch via
makepkg (non-root builder user). git/curl install before checkout
(container jobs). AppImage keeps cargo-packager on the Ubuntu runner.
The release gates on all four package jobs plus macOS/Windows.
- the Windows runner image exports MSVC's INCLUDE/LIB; cc-rs was
appending the MSVC SDK headers to MinGW compiles (vcruntime.h not
found)
- oak-worker handshake test helper advertised the input pool's total
byte size as per-slot data bytes (macOS tolerated the oversized
attach; Linux correctly rejects it)
- hw/sw decode comparison tolerance 0.05 -> 0.08 (VideoToolbox's
YUV->RGB legitimately differs by ~1 LSB of intermediate depth)
- the vendored OCIO source needs MSVC-only constructs (wide-path
ifstream); MSYS2's mingw build of the exact 2.5.2 the bridge targets
is the sane Windows path — DLLs get packaged next to the binaries
- oak-worker handshake test prints the error response on failure
(CI-only attach failure needs the message)
Each dimension of a parametric param renders as a CurveEditor in the
inspector (bezier handles map to the Hermite slopes of the host curve
model; edits serialize back through the JSON mirror — undoable and
project-persisted). The engine re-sync skips in-progress drags and
identical curves so the per-render sync neither steals gestures nor
loops. Also: physical-memory probe for the worker-count policy on
Windows (GlobalMemoryStatusEx).
- rust-toolchain sets CARGO_HOME to the Windows userprofile path while
the msys2 shell's HOME is elsewhere — the yaml-cpp patch targeted an
empty directory and the assertion ls failed
- examples/screenshot.rs uses the macOS-only VisualTestAppContext; its
items are now cfg-gated with a non-macOS stub main so workspace test
builds pass on Linux/Windows
- ValueType::Parametric; the node input carries the whole curve set as
NodeValue::Text(JSON) so undo and project serialization come for free
- translation pass builds the input with the default-curve JSON and the
dimension/range/ui-colour properties
- edits flow both ways: node input (UI) -> curves_from_json ->
set_ofx(Parametric) on the instance; plugin-side Set/Add/Delete ->
notify_instance_changed -> JSON written back to the input (undoable)
- screenshot example: gate the macOS-only offscreen capture items so
the workspace tests build on Linux/Windows
- oakplugin host: Win32 LoadLibraryExW/GetProcAddress/FreeLibrary
backend (LOAD_WITH_ALTERED_SEARCH_PATH so bundle-sibling DLLs
resolve), same dl_open/dl_sym/dlclose surface — the POSIX path is
untouched; OFX hosts now compile on Windows
- proxymanager: PATH split via std::env::split_paths (Windows ';'),
ffmpeg.exe name, Windows candidate locations; split logic unit
tested
SharedMemoryRegion gains a Win32 backend behind the unchanged public
API: Local\OakShm<key> names, OpenFileMapping for attach, VirtualQuery
for the size check, UnmapViewOfFile/CloseHandle for teardown. Semantic
differences from POSIX are documented: unlink_key is a no-op (the
kernel destroys the object with the last handle, so crashed owners
self-heal) and Create on a live name fails instead of replacing.
Windows CI builds the workspace again.
On a fresh runner registry/src has no hash subdir yet, so the unpack
glob never expanded and the step exited 2; derive it from the cache
dir and assert the patched file exists at the end.
Parametric (curve/LUT) parameters: ParamValue::Parametric holds one
ordered control-point curve per dimension (identity default over the
declared range), evaluated as piecewise cubic Hermite with auto
(centered-difference) slopes; the full suite — evaluate / count / get /
set / add / delete / delete-all — with the spec's error codes, descriptor
defaults copied to instances, and instanceChanged notifications on
edits. paramDefine accepts OfxParamTypeParametric; the dimension/range
and UI-colour properties round-trip. 148/148 real plugins discovered,
135 registered (one more than before: the parametric-suite consumer).
The first processor build has no prior state to be stale against; the
spurious bump made the next cpu_frame call drop the freshly cached
image (playback_display_tracks_the_playhead regression).
- cargo fetch does not extract sources; the yaml-cpp <cstdint> patch
now untars the .crate into the registry src dir first (the glob
found nothing and the step failed with exit 2)
- oakaudio: the watchdog-wrapped audio test called Self::... from a
free-function test module (compile error in lib test)
- Linux: libxkbcommon-x11-dev for the gpui X11 client link
- Windows: patch <cstdint> into the vendored yaml-cpp (a cached cmake
configure ignores CXXFLAGS; the patch is idempotent and runs after
cargo fetch)
- macOS: the hw-decode test skips its VideoToolbox engagement
assertions on hosts where VT cannot initialize (headless/virtualized
runners) instead of failing
- display color management: the display ICC (system or custom) is
applied to viewer frames at present time (F32 in place, or in place
on the BGRA staging copy with the R/B swizzle baked into the OCIO
chain); preferences get a Color section (mode + custom ICC file); on
macOS the Metal layer is tagged with the display colorspace when
self-managing so ColorSync passes pixels through (no double
correction); frame caches track the transform generation so a mode
or profile change drops stale pixels
dpkg-shlibdeps over the three shipped binaries resolves every NEEDED
library to exact build-distro package names (FFmpeg/OCIO are static so
only base-OS packages appear) and rewrites the deb's Depends. Distros
with divergent package names (openKylin) get their own build instead
of a wrong-name dependency list.
Static FFmpeg + static OCIO leave only base-OS libraries; the audit
step prints objdump NEEDED for each packaged binary so any accidental
dynamic dependency (and any distro-specific package-name surface) is
visible in the build log.
- OCIO_RS_LINK=static everywhere: the vendored OCIO is linked into the
binaries statically — the package carries no OCIO dependency
- Linux: libasound2-dev (alsa-sys), libpulse-dev, libsndfile1-dev —
the full audio dev set
- Windows: -include cstdint for the vendored yaml-cpp (pre-GCC-13
transitive includes)
- Windows: install MSYS2's own Rust (x86_64-pc-windows-gnu host); the
rustup MSVC toolchain is not on the msys2 shell's PATH and the MSVC
linker rejects the Unix-style link args the build scripts emit
- Linux: libpipewire-0.3-dev + libspa-0.2-dev for libspa-sys (gpui's
Linux screen-capture/audio stack)
The distro OCIO is too old for the bridge's API floor where it matters
(Ubuntu 24.04 ships 2.1; the bridge uses 2.4+ APIs), and version drift
across platforms is a support hazard — enable ocio-rs' bundled feature
and drop the OCIO_INSTALL_DIR/system-package wiring from CI and CD so
Linux, macOS and Windows all build the same vendored OCIO. cmake/make/
diffutils added where the runners lack them (Windows FFmpeg build needs
make + cmp).
- Homebrew renamed libtheora->theora and libwebp->webp; the old names
no longer resolve, failing the macOS dependency step
- retry the MSYS2 pacman install (3 attempts, --needed resumes): CI
mirrors stall mid-download ("Operation too slow") often enough to
matter
- serializer resolves node types through the factory's dynamic
(runtime-registered OpenFX) entries, so a project carrying plugin
nodes loads again (was: "unknown node type"); covered by a new
CI-gated round-trip test driving the real fixture plugin
- built-in effect nodes expose their inputs as inspector parameters
like the C++ parameter editor: localized input names from the
behavior, combo option tables via the new
NodeBehavior::input_combo_strings (16 nodes, string-for-string from
the C++ set_combo_box_strings), connection/data inputs excluded
- effect library: live drag-and-drop — onto the inspector's effect
stack (lands at the indicator position) and onto the node editor
canvas (creates the node at the drop point); double-click still
appends to the selected clip
- inspector parameter controls are no longer recreated per render
(gpui stack view caches them per effect), so sliders drag and
checkboxes click; the view observes the engine and silently re-syncs
values (undo/redo land on the widgets)
- timeline: left-press selects clips (plain/keep-multi/Ctrl-Cmd
toggle); clip moves clamp the shared delta so no clip of a linked
group lands before frame 0 instead of failing with "invalid move
target"
- oakplugin: createInstance-rejected instances skip the destroyInstance
notification (the plugin never owned them); vendor-suite fetchSuite
misses moved behind OAK_OFX_TRACE; the worker logs the discovered/
registered plugin counts
- CI: the OFX probe step also runs the serialization round-trip test
- gpui submodule: params view caching, clip click-select, library
drag payload, graph_position_at
- new minimal C OFX plugin fixture (ci_test_plugin.c) compiled into a
real .ofx.bundle by build_fixture.sh; the CI step points
OFX_PLUGIN_PATH at it and asserts the scan_probe example discovers
AND registers it (Linux/macOS)
- host bundle binary search now also covers the OFX-standard Win64
platform directory
- cd.yml: restore the Windows NSIS packaging job (obsolete oakengine
cdylib prebuild dropped) and repair the job indentation that had
silently detached the macos/release jobs from the jobs: map;
releases now gate on all three platforms
- save() writes <config>/config.toml (atomic temp+rename), flat keys
at the top level and group/sub keys as [group] tables, values as
native TOML int/float/bool/string (non-finite doubles degrade to
strings and restore via the declared type)
- load() prefers config.toml; a legacy config.ini (C++ or pre-TOML
Rust builds) is read once and immediately re-persisted as TOML; the
INI file is left in place; a corrupt TOML is reported, never
silently discarded
- cd.yml: drop a stale oakengine comment (the crate is retired)
- NLE track growth is now a display concern: video/subtitle track
lists render reversed (a new track lands on top), audio lists render
in order (a new track lands at the bottom); the graph list always
appends. Track-add undo removes THIS track by id instead of blindly
removing the last one
- add_track returns the actual index of the new track (diffed against
the pre-command list) instead of assuming append-at-end
- status bar proxy segment reflects the real Use Proxy Media switch
instead of a static "Proxy: Off"
- proxy transcode PROGRESS events no longer invalidate the rendered
frame cache on every tick (only completion does) — progress updates
used to keep the playback cache permanently cold while generating
- effect library: live search box (name/type-id substring), Built-in
group header, and the addable-effects table is sorted alphabetically
(built-ins first, then OFX sub-category groups)
Real openfx-misc/CImg/Shadertoy bundles (148 plugins at
/Library/OFX/Plugins) all failed to load before; every failure was
silent. Root causes found one by one with a probe example + lldb:
- property suite rejected propSet on undefined properties and
propGetDimension on empty ones, and disallowed the index==size
append — OFX semantics are create-on-set and appendable dimensions
(this alone failed every plugin's describe)
- host property set missed the mandatory OfxPropType/OfxPropAPIVersion
and the capability props ofxs' fetchHostDescription reads with
throwOnFailure=true (IsBackground, TemporalClipAccess, MaxPages,
PageRowColumnCount, host SupportedContexts, ...) — one missing prop
aborted the read chain and left a half-initialised host description,
which made every temporal plugin refuse to load
- MultiThreadSuiteV1 lacked the five mutex functions (the plugin reads
past the short table — UB); implemented as a real counting-semaphore
registry
- the OfxHost struct was a stack local; ofxs keeps the POINTER past
setHost, so describe/render-time fetchSuite calls dereferenced a
dangling stack address (bus error once plugins actually loaded) —
the struct is now a leaked process global
- General is a standard OFX context and is no longer filtered out
(Roto/AppendClip/STMap declare only it)
- every scan/load/describe early-out now logs its reason; suite entry
points report non-OK statuses with caller location under
OAK_OFX_TRACE
- examples/scan_probe.rs: scans the real plugin dirs and prints
discovered/registered counts (also usable from CI)
Result: 148/148 plugins discovered, 134 registered as node types (the
remaining 14 need vendor suites — Vegas stereoscopic etc. — and are
logged, not silent)
- string tables live in assets/i18n/<lang>.yaml, loaded at runtime
(user pack dir ~/.oak/i18n, app bundle Resources/i18n, dev checkout)
with the compiled-in English/Chinese tables as fallback
- new status.proxy.on/off keys; inspector.params copy no longer says
"placeholder"; effect_library.group.builtin key
- bundle the packs as cargo-bundle resources; dev profile dep opt-level
dropped to 1 for faster iteration builds
- Thumbnails never appeared because the PNG was written to a .part
file with format inferred from the extension (always failing); the
writer now uses an explicit PNG encoder, and an e2e test proves the
pipeline yields real files.
- Mock engine: the CardSelected -> SelectionChanged echo no longer
re-expands a card the same click just collapsed.
- Dragging a clip moves its linked audio/video partners by the same
frame offset in a single undo entry; the dragged clip may change
tracks while partners keep theirs.
- Root view paints the near-black base so dock gaps match.
- Waveform trace follows the deep-green design color.
- actions.rs: reset the shortcut overrides in
save_writes_only_entries_that_differ_from_default (pre-existing
flake: a leaked override from a previous test poisoned the shared
lock under parallel test order).
- With a clip selected, the node editor shows that clip's context
chain (footage -> effects -> clip) instead of the global graph; the
clip's node is highlighted. No selection keeps the full graph.
- Node clicks in the graph select the node and expand/highlight the
matching effect card in the inspector; clicking an inspector card
highlights the node in the graph (single source of truth: the
engine's selected_graph_node).
- The Window menu lists every panel, checks the open ones, and toggles
visibility on click (closed panels reopen at their last dock target,
falling back to the default group) — a panel closed by accident
(e.g. the inspector) is one menu click away again. The menu refreshes
on dock structure events.
- OFX plugin scanning now covers the full standard location set:
per-user (~/.OFX/Plugins, ~/.local/share, ~/Library/OFX/Plugins on
macOS), system-level (/Library/OFX/Plugins, /usr/OFX/Plugins,
/usr/local, %ProgramFiles%\Common Files\OFX\Plugins), app-relative,
and the OFX_PLUGIN_PATH environment variable.
TrackPlaceBlockCommand::redo now homes the block's in point to the
placement target (capturing the original for undo): the Rust block
stores its position on the block, so a fresh clip that never had its
in point set always rendered at the timeline zero — the 'drops always
land at zero' bug. The original in point is captured on the first redo
and restored on undo, keeping the sync re-place round-trip exact. The
A/V drop test now asserts the clip lands at the drop frame.
- Footage drops no longer clamp to the sequence length — that clamp
squashed every drop on an empty or short timeline to frame zero.
Dropping past the end now extends the timeline, so the clip lands
where it is released.
- While dragging, a translucent ghost (35% opacity, accent border)
previews the resolved track, start frame and footage length at the
cursor; a new AppEngine::footage_length_frames (probed duration x
frame rate) feeds its extent, with a mock implementation for demo
mode.
- Main-process interact instances for the selected OFX effect card
(create on selection change, describe, destroy on deselect/close),
coexisting with the render-worker plugin instances per the OFX
multi-instance model.
- Program viewer composites the interact's overlay: draw into a GL
FBO via gl_bridge, read back, straight-alpha 'over' composite onto
the displayed frame; cached and only re-rendered on frame/time/
viewport/instance change or plugin redraw requests.
- Event forwarding: picture-area pointer maps through the contain-fit
letterbox inverse to OFX pen coordinates (pen_motion/down/up);
Keystroke to OFX key symbols (ASCII, navigation, F1-F35) for
key_down/up; a 50ms idle pump; global shortcut consumption keeps
precedence.
- e2e with the real test plugin: lifecycle marker assertions, pen/key
event records, and macOS GL overlay compositing verified (265 tests
green incl. gpui_widgets viewer suite).
- gl_bridge: macOS CGL offscreen context (process-wide singleton,
serialized GlGuard), real GL output textures/FBOs, glReadPixels
readback with vertical flip and format conversion; use_opengl now
really engages for OpenGLRenderSupported plugins (verified with real
GL rendering: C smoke 11/11, unit tests, GL e2e).
- OfxColor: color params get a swatch button plus a real picker popup
(RGBA sliders, live preview, hex input, undoable commit) replacing
the four spinboxes.
- Push buttons route kOfxActionInstanceChanged (UserEdited) per the
OFX contract; test plugin asserts the callback.
- Worker-side plugin progress flows to the main-process progress
dialog over the NDJSON control channel, with cancel propagation.
- OFX Interact host: NewInteract/Describe lifecycle, Draw/Pen/Key/Idle
action surface with proper in-args, DrawSuite v1 host implementation
sharing the gl_bridge context; interact test plugin verifies the
event stream and real GL drawing.
- Right-clicking an unselected clip selects it first (C++ parity) —
this is what made Cut/Delete appear to do nothing.
- Cut/Copy/Paste clipboard: clipboard_copy/cut/paste on the engine,
clipboard clips keep footage/range/speed/track kind and stay linked
in the pasted group; paste lands at the playhead as one undo entry.
- Dropping a video-with-audio footage places the video clip plus a
linked audio clip at the same range in ONE 'Add Clip' undo entry.
- Add Video/Audio Track buttons in the timeline toolbar and the track
header context menu; new sequences start with 2 video + 2 audio
tracks (not an undoable edit).
- oaknode Graph::add_entry now reclaims the slot from the free list —
before, a detached-then-reattached node left its slot in the free
list, so node_count undercounted and the next add_node silently
clobbered the restored node. This was the user's 'undo, redo, undo,
redo and the result changed' bug; regression covered by cycle tests
(move/trim/delete/split/add-track/linked-placement all converge).
- Dropping a video-with-audio footage now places a video clip AND a
linked audio clip at the same range in ONE undoable 'Add Clip' entry
(the links live on NodeCore.links, the canonical links_of storage;
auto-creates the missing track kind).
- The timeline toolbar gains 'Add Video/Audio Track' buttons and the
track-header context menu offers the same two entries above
Delete/Delete All Empty.
- create_sequence now starts every new sequence with the default
2 video + 2 audio track layout (driven directly through the add-track
commands, not through the undo stack). Tests updated for the new
default track counts.
FFmpeg 8 removed the standalone hardware decoders (h264_videotoolbox/
vaapi/nvdec/d3d11va no longer exist in its configure) — hardware decode
now only exists as a hwaccel attached to the software decoder. The new
oakcodec::hwdecode module therefore opens the regular decoder with the
platform's hardware device context attached (VideoToolbox on macOS,
VA-API then NVDEC on Linux, D3D11VA then NVDEC on Windows): FFmpeg
engages the matching hwaccel, decodes into hardware surfaces, and we
transfer them to system memory (NV12/P010) ahead of swscale.
- HardwareDecoding config switch, default ON by mandate; a checkbox in
Preferences > Rendering (EN/ZH); device creation failure skips to the
next candidate and finally to software; a decode-time failure on a
hardware session reopens it as software and retries once.
- hw_decoder_name() observability hook plus a HW_TRANSFERS counter so
tests can prove the hwaccel really engaged (not silently software).
- Verification: demo.mp4 H.264 decodes through VideoToolbox with a
transferred hardware surface, and the pixels match the software
decode within 0.05; switch off forces software.
- build-ffmpeg.sh also enables nvdec when ffnvcodec headers exist.
The C++ viewer Playback Resolution menu, wired end to end: the radio in
the viewer context menu reflects and sets the PlaybackDivider config,
proxy_render_size renders the preview at 480/divider long edge, and
changing the divider invalidates the cached and in-flight preview
frames. This is the escape hatch for machines that cannot keep up with
playback (measured: a debug build of the worker pool reaches only 11
fps vs 152 fps in release on 1080p H.264, which no amount of
scheduling can make realtime).
Two compounding causes behind 'playhead advances but the picture stays
frozen' and 'pause freezes the app':
- The wall-anchored clock teleported the playhead past the pre-render
window during any long stall (the first render after pressing play
costs seconds while the worker pool spins up: measured +104 frames in
one 4.1s block). The window then started behind and, with stale
in-flight frames occupying the workers, never converged.
RealClock::tick now clamps the advance to 2 frames/tick and
re-anchors the dropped time (NLE drop-frames semantics).
- Window frames the playhead had already passed stayed pending/in
flight, burning worker time on frames that could never be displayed.
update_preview_window now cancels them per tick via the new
JobDispatch::cancel_preview_frame, keeping the workers on frames
around the playhead.
Includes a production-shaped regression test (real 1080p media on the
timeline, actual cpu_frame display path) that failed with the exact
production signature (playhead 240 / displayed 0 / 36 stale slots)
before the fix and passes after.
The main-process sample showed the UI thread spending 100% of its time
in TicketArena::wait from the painted frame's synchronous render: every
cache-missed playhead frame sync-rendered inline, and the seek-priority
ticket then stole worker capacity from the pre-render window while the
blocked tick loop could not feed it — a self-reinforcing loop that made
playback unusably choppy.
On a playback miss the viewer now shows the last displayed frame while
the pre-render window warms up/catches up (paused monitors and the very
first frame keep the synchronous path). Adds a gpui test driving real
playback that requires the window to supply playhead frames, and the
real-footage bench_playback example used for the measurements
(152 fps aggregate on 1080p H.264 at 480p preview, decode-bound).
A claim mixing audio and video tickets is delivered as the video
message first and the audio message second, and the worker pops one
free-ring slot per ticket in that message order, checking each pop
against the assignment. The dispatcher however assigned slots in the
scheduler's interleaved frame order, so every audio ticket inside a
mixed batch mismatched, and each mismatch consumed a worker slot
without recycling it — cascading into the 'slot assignment mismatch'
flood and failed frames during playback.
Slot assignment now partitions the claim: video tickets first, then
audio. The mixed_audio_video integration test forces mixed claims
(queue depth > slot count with immediate releases) and fails with the
exact production signature when the fix is reverted.
Pressing play froze the app: the 120-frame pre-render window could
hold every shm slot in the pool (e.g. 8 workers x 3 F32 slots = 24 <
120). Once the wall-clock playhead outran the renders, the UI's
synchronous frame wait had no credit to dispatch, and the
slot-releasing cleanup runs on that same blocked UI thread — a hard
deadlock.
The window is now capped to (workers x slots - workers), reserving one
slot per worker so interactive (seek/sync display) and audio tickets
always dispatch. preview_window_capacity is exposed through
JobDispatch; a unit test pins the reserve math.
- preview_footage_media decodes the original from the footage's actual
first stream of the kind instead of hardcoded 0/1, fixing audio-first
and other atypical stream layouts (with a unit test).
- spawn_modal now probes for a nested window update and defers the
build instead of silently dropping the dialog — the phase-7
Preferences/Action Search fix applied centrally to every modal
(export, proxy settings, project manager, progress dialogs).
- The gpui_wgpu atlas no longer double-copies identity-format uploads,
leaving a single CPU staging copy (gpui RenderImage ownership) plus
the GPU upload on the onscreen path; the residual copy and the
IOSurface route to true zero-copy are documented in the M15 design.
- Audio tickets join the process backend: render_audio_batch wire
message, workers mix straight into shm slots (SLOT_FORMAT_AUDIO_F32),
ShmAudio payload with release semantics, crash isolation covers audio
renders; playback audio uses an async 4-chunk prefetch drained on the
UI tick (also fixes the sub-60fps chunk truncation bug); oversized
ranges and dispatcher outages fall back to in-process inline.
- Per-ticket slot formats: force_format is honored (exports request
F32 slots, dropping the BGRA8 round-trip and its 8-bit quantization);
segments grow on demand via worker-idle rebuild with generation
handoff; the scheduler filters over-capacity tickets.
- Adaptive defaults: 128-256MB/worker segment budgets drive slots per
worker, batch size follows workers/slots; bench_process example
measures throughput and adjacent-frame completion deltas
(e.g. 4 workers: 841 fps, 4.6ms mean delta).
With the C ABI facade (oakengine) retired, the frozen-ABI rationale is
gone. UndoCommand now boxes a Send Command trait (new/from_closures/
multi), dropping OakUndoCommandVtable, the userdata trampolines, the
refcount shell, the handle module, and all undostack_* handle exports.
The global facade loses its raw-pointer out-params (can_undo/can_redo
return bool, command_name returns String). oaktimeline/oaknode/
oakplugin/oaktask construct commands directly via UndoCommand::new.
oakundo src is now free of unsafe; behavior (ordering, idempotence,
done flags, groups, observers, 200-row cap) is unchanged and pinned by
the rewritten tests.
Mike-Solar