The command palette's search field took the keyboard with `request_focus()`
on every frame it was drawn. egui's `Memory::request_focus` interrupts the
IME composition, and `egui-winit` acts on `should_interrupt_composition`
with `set_ime_allowed(false)` + `(true)` — on Windows that is
`ImmAssociateContextEx(hwnd, 0, IACE_CHILDREN/IACE_DEFAULT)`, which drops the
composition in progress. With Microsoft Pinyin the pre-edit text was
cancelled over and over, so the pinyin never accumulated and nothing could be
committed: Chinese (and Japanese/Korean) could not be typed into Quick Apply
at all, while Latin typing was unaffected.
Ask for the focus only when the field hasn't got it, the same guard the
inline editors use (`widgets::keep_focus`): the palette still takes the
keyboard when it opens, and takes it back when something else takes it away.
Tests: panels::dialogs::tests::the_palette_leaves_a_pinyin_composition_alone
opens the palette, feeds Preedit("n".."nihao") frame by frame and asserts that
the frame never asks the platform to cancel the composition, that the whole
pre-edit text is in the query, and that the field claims the keyboard back
when another widget takes it. Failed before: "typing `n` cancelled the
composition".
A single selected_layers list kept ids from every comp, so the active comp looked selected whenever another comp was. Opening a comp and running a script against a different comp now park and restore that selection.
sanitize_state dropped every selected layer not in the active comp at the
end of each script run, so a layer selected by script in a comp that was not
open was gone by the next script call. Every comp keeps its own selection as
in After Effects; only ids gone from every comp are dropped now. UI and
engine lookups resolve selections per comp and tolerate foreign ids.
The Japanese data point in #821 (zitann2020) offered to write its rows against whatever mechanism
this PR lands, so both languages share one catalog instead of inventing two. This makes adding a
language a data change and reserves `ja`:
- `Catalog` holds a language's effect rows and its category labels, and `catalog()` resolves a
language code to one; adding a language is two tables and one arm, with no change to the lookups
or to any panel.
- `JAPANESE_EFFECTS` and `JAPANESE_CATEGORIES` are declared, keyed by the same effect ids and by
the registry's English category names, and `ja` resolves to them. Both are empty, so a Japanese
interface draws exactly what it drew before — the slot changes no behaviour on its own.
- `a_pending_language_only_uses_known_ids` guards the pending tables (Japanese and Traditional
Chinese) as rows arrive: an id Simplified Chinese does not know, a different English name for one
of them, an empty translation or a repeated id fails. `languages_without_rows_draw_the_registry_name`
asserts each pending language has an arm in `catalog()` (a table with no arm would leave a
translator's rows dead) and is still row-less, so the day one gets rows the test says where its
coverage moved.
- `only_the_brand_names_stay_in_english` now documents that it pins *Simplified Chinese's* five
brand names and is deliberately not shared: AE's Japanese interface keeps the whole `CC`, `VR`
and `OCIO` sets in English while its Chinese one translates them, so repeating the English name
is a language's choice and not an unfinished row. The row check reflects that (a repeated English
name is accepted, an empty one is not).
docs/localization-parity.md records the two empty tables and that each language decides for itself
which names stay in English.
Refs #821
Verification against the installed After Effects 2026 (26.5.0.89) Simplified Chinese interface
(the checklist's `07-类别` sheet) finds two of the 21 category names differ from our first
suggestion:
- `Keying` is 抠像, not 键控 — the interface names the folder 抠像 while the effects inside it keep
键 (颜色差值键, 亮度键, 内部/外部键), so the two do not contradict each other;
- `Obsolete` is 过时, not 已废弃 — the effects in the folder say 旧版 for the legacy ones.
The other nineteen are confirmed as they were. `Color Correction` (颜色校正), `Matte` (遮罩) and
`Noise & Grain` (杂色和颗粒) were the three where the checklist's older machine-translated
suggestion differed; the interface's wording matches what the catalog already had, so those stay.
`CATEGORIES`' doc comment now records where the 21 names come from instead of calling them our
wording.
Refs #821
The interface below the menu bar is localized (618 fixed menu entries, ~490 panel rows) but an
effect's name still rendered in English everywhere it appeared: the generated Effect menu, Effect
Controls, the Timeline and Effects & Presets.
The registry's English name is the effect's identity: `effect.list`, `effect.apply {"effect": …}`,
scripts, expressions and the property tree's group name all match on it. Only the drawing changes.
`crates/ui-egui/src/i18n/effects.rs` keys the names by effect **id** rather than by the English
string — a repeat is stored as `Gaussian Blur 2` (an English key would miss it) and upstream
renames a display name now and then, while the id is stable. It carries 306 rows for Simplified
Chinese, one per effect in the registry, taken from the installed After Effects 2026 (26.5.0.89)
Simplified Chinese interface, with the `CC` series following Adobe's Chinese documentation (which
names 68 of the 69 that its interface leaves in English). `Cryptomatte`, `EXtractoR`, `IDentifier`,
`Key Light` and `Unmult` stay in English as brand names; `Face Measurements` and `Face Track
Points` are EffectCraft's own effects. The 21 category folder names are in the same file, worded
after the effect names (channel → 通道, noise → 杂色, matte → 遮罩 …).
What draws the catalog now:
- **Effect menu:** the entries and the category submenus. The command and its `{"effect": <id>}`
params are untouched, so the menu still applies by id.
- **Effect Controls:** the header, its Rename field, the accessible name and the About popup.
- **Timeline:** effect group rows, and the search (a Chinese interface is searchable in Chinese).
- **Effects & Presets:** the rows, the drag ghost, the folder rows and the search.
Kept as they are: the effect id, the property tree's `PropGroup.name`, every automation id and
label, `effect.list`'s `name`/`category`, and the undo label (`Apply Gaussian Blur`). A name the
user typed in Effect Controls stays verbatim, and a repeated instance keeps the index the engine
appends (`Gaussian Blur 2` → `高斯模糊 2`). No other language has rows yet, Traditional Chinese
included, so those draw the registry's English name.
Tests: nine unit tests under `cargo test -p effectcraft-ui-egui --lib i18n` cover the data (every
effect has a row, the row still carries the registry's English name, ids are unique, only the five
brand names repeat their English name, every category has a row, a row names an effect the registry
knows) and the behaviour (the instance index, user renames, English and untranslated languages,
search) plus a scan pinning the drawing sites. `crates/ui-egui/tests/ui_effect_names.rs` checks the
Effect Controls header, the generated Effect menu and that the ids, automation labels and
`effect.list` stay English.
`docs/localization-parity.md` and `docs/gaps.md` no longer record effect names as deliberately
English; effect *parameter* names are still untranslated and stay the next piece.
Refs #821
FootageItem had no replaceWithSequence, setProxy, setProxyWithSequence or
setProxyToNone, useProxy and proxySource were fixed stubs, and mainSource
was a new object on every read with no conformFrameRate. They now run the
existing commands (file.replaceFootage, file.setProxy, file.useProxy,
file.setProxyNone, file.interpretFootage); the item info reports the proxy
and the conform/native frame rates, and mainSource is one live
FootageSource per FootageItem. replaceSource is not covered: it needs an
engine command that changes a layer's source.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keeping OK and Cancel in view took away the outer scroll area that also
scrolled the dialog sideways, and the controls only scrolled up and
down. In a narrow window (960x600 at 200 % UI Scale leaves the controls
about 190 points next to the 200-point list) they ran past the dialog's
right edge with no way to reach them.
The controls now scroll both ways in the room left under their heading,
with no minimum scrolled size, so a small budget can't hide the end of
them; the list scrolls in its height the same way. Each control also
registers dialog.layerStyle.value.<style>/<path> next to its label's id.
Tests:
- layer_style_controls_stay_reachable_in_a_small_window: in a 960x600
window at 200 %, and opened large then narrowed, Noise is scrolled to,
painted whole and dragged; fails before (it lies past the window).
- layer_style_buttons_stay_in_view now checks both buttons at 100 % and
200 %, each case in a fresh app with a real edit: OK keeps it, Cancel
puts the project and its undo history back as before the dialog.
After Effects scripts save a comp frame with saveFrameToPng(time, file);
EffectCraft only had the comp.saveFrameAs command, so the method was
undefined. It now runs comp.saveFrameAs at the given comp time, after the
same file-gate check as other script writes (Allow Scripts to Write Files).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
epaint's has_glyph compares face keys, so a glyph drawn by the replacement
face reads as missing. With no face in the stack covering egui's `◻`, Inter's
`?` becomes the replacement and Inter's `•` fails the check, which happens
once another test (the #1141 one) has loaded a system face without `◻`: the
fallback tries loaded faces first. Compare the drawn glyph with the
replacement glyph instead, as assert_drawn already does.
The Layer Style dialog's OK and Cancel came after the style list and
the style's controls in one scroll area, and at its usual size they
were laid out past the dialog's bottom edge: clipped, and a click on
them did nothing until the dialog was scrolled.
The dialog now uses above_footer: OK and Cancel sit at its bottom right
and the list and the controls scroll above them, also at 200 % UI
Scale.
Test: layer_style_buttons_stay_in_view (both buttons painted whole and
clicked, at 100 % and 200 %; fails before: Cancel isn't painted).
Settings keeps its buttons in view by laying them out first, at the
bottom of its body, and giving its pages the height left above them,
clipped to it. That layout is now dialogs::above_footer, so other
dialogs can keep their buttons in view the same way. Settings looks and
works as before.
Dashes reported its children as `dash`, `gap`, `offset`, so
`property("ADBE Vector Stroke Dash 1")` was null and addProperty threw.
They now report ADBE Vector Stroke Dash 1-3 / Gap 1-3 / Offset.
addProperty with one of those names returns the existing property
(EffectCraft's Dashes always holds Dash 1, Gap 1 and Offset) or adds the
Dash/Gap 2 or 3 pair through shape.dashes.add. Lookups by the old ids
still work.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Stretched layers copied the nearest source sample for each output sample, which adds error: at
48 kHz the worst error on a 440 Hz tone was 0.025, and 0.15 at 8 kHz. They now blend the two
source samples either side. Layers at 100% and 50% speed keep their exact samples.
The render tests cover a plain precomp, which mixes the same audio as its layers, and the stretch
error against the exact tone.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Adding a marker replaced the comp, so every cached frame stopped matching and the preview had to
re-render before its audio could carry on. Plain markers change no pixels; protected time markers
with a duration still do (Responsive Design). Markers count in full again when a project uses
expressions, which can read them.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A script asks a light's options for "ADBE Casts Shadows", which returned null. The light's option
is named "Casts Shadows" here, so the lookup never matched. The name is now an alias for the
castsShadows property, and a regression test checks the lookup returns the property.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Shift+Spacebar now plays every frame by default, so it is no longer a gap in docs/ui-parity.md.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Inside a function wrapper such as (function () { ... })(), boa places a ReferenceError at the start of
the function, so a script's error was reported at the wrapper's first line rather than the line that
failed. The first use of the undefined name in the code is now the position, and boa's frames remain
the fallback. Thrown errors and object-model errors already report their own lines. A regression
test covers a thrown error, an undefined name, and an object-model error inside a wrapper.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Pro Levels2 keeps its Alpha controls one group down, under Alpha, and a script's
effect.property("Alpha Input White") only searched the effect's direct children, so it returned null.
The lookup now searches the groups nested below the direct children, up to three levels, after the
direct match fails. Direct children keep their priority. A regression test checks the name on the
effect and the same control through the Alpha group.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The Gradient Ramp wrote alpha 1.0 to every pixel, so a shape layer's transparent area filled up with
the ramp and the layer's bounds showed instead of the shape. The ramp now uses the pixel's own alpha
on the CPU and in the GPU kernel, as the #939 fix does for the noise effects. A regression test
checks that transparent pixels stay transparent at two Blend With Original values.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Every key call (removeKey, keyTime, keyValue, setInterpolationTypeAtKey…)
looked its key up through `__keys()`, which serialised the whole key list,
eases and tangents included, on each call. Removing keys one by one was
therefore quadratic: about 100 ms per removeKey on a 2,000-key property.
The `keys` query now takes an optional `index` and returns just that key
with the count; `__key` uses it. removeKey on 2,000 keys drops from
103.7 ms to 0.4 ms (release, Windows 11 aarch64).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
uniform_parent_pose took the Y angle from asin, which loses precision within about 1e-6 degrees of
+-90. The Y angle now comes from atan2 on cos(Y), as the X and Z angles already do. The gimbal
branch and the principal-range output are unchanged.
A new test parents a static Orientation with Y 1e-6 degrees off gimbal lock under a uniform
parent. Its largest corner displacement was 3.3e-8 px before this change and 1.5e-14 px after.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The layer bar Quick Offset was followed and ended in its row's code, and
rows out of view aren't drawn. Released with its row scrolled away, it
stayed active: the next drag of that bar, even a plain one, ran as a
Quick Offset from the old origin, into the old undo step.
Like the key drag, it now follows the pointer below the rows and ends
when the button is released, wherever the bar's row is.
Test: a_quick_offset_released_off_screen_ends (release with the row
scrolled away; a plain drag after it moves both layers together as its
own undo step; fails before).
keys.quickOffset and keys.move stored the drag's new total before the
edit ran. When the edit refused the step (keys.quickOffset with a key
landing past the end of time), the project stayed as it was but the
total kept the refused delta, and with incremental steps (merge without
fromStart) the next step moved by it too.
The drag state is now stored only after the edit applies.
Test: a_refused_step_does_not_count_towards_the_drag (+1 s, a refused
step, -1 s: the keys end where they started; fails before: the last key
ends near the end of time).
keys.move and the nudges snapped each key's destination to the nearest
frame, so moving keys a Quick Offset had staggered onto subframes put
them back on frames: four keys at 0, 1/3, 2/3 and 1 frame, moved one
frame, came out at 1, 1, 2 and 2. Layer bars kept theirs, because
layer.timing snaps the move, not the times.
The move is now rounded to whole frames and every key translated by it.
layer_shift turns a comp-time move into a layer-time move through the
layer's start and stretch (reversed too) without a round trip through
comp time; keys.quickOffset uses it as well.
Test: moving_staggered_keys_keeps_their_subframe_spacing (Quick Offset,
then keys.move and keys.nudgeForward; fails before: 1/3 frame lands on
a frame).
wgpu validates indirect draw and dispatch arguments with a compute shader it
compiles when the device is created. Metal on an Intel Iris Pro (MacBookPro11,3,
macOS 12.7.6) fails to compile it ("Compiler encountered an internal error" in
MTLCompilerService), the device is lost, and the window never opens.
EffectCraft draws and dispatches nothing indirectly, so leave the check out of the
instance flags. WGPU_VALIDATION_INDIRECT_CALL=1 turns it back on. VectorCraft
does the same (storytold/vectorcraft#651).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
prop.reset cleared every keyframe before setting the default value. As in
After Effects (and as effect.reset already does for effect parameters), an
animated property now keeps its keyframes and gets the default as a key at
the current time; a static property just takes the default. The expression
is still removed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In After Effects a marquee selection staggers in the comp's layer order,
not in the direction the box was drawn. A Shift-click range of layer
bars and a box of keys already come out top layer first; this test makes
both with the pointer (the range from the bottom layer up, the box from
the bottom right to the top left) and checks the Quick Offset that
follows: the top layer stays and the bottom one moves the whole delta.
Validate saved stroke frames against the current layer limits before clamping a span edit. Keep project and undo state intact when trimming has left strokes outside the layer.
Co-Authored-By: OpenAI Codex <noreply@openai.com>
The Chinese UI fallback was chosen by probing 文, which Japanese faces
cover too. Fallback takes a face already loaded first, so after the
English interface had loaded a Japanese face for kana, switching to
Simplified Chinese reused it, and the simplified-only hanzi (选, 删,
层, 设, 页 ...) drew as boxes. A Simplified Chinese interface now probes
页, as wordcraft does for its interface font (storytold/wordcraft#248).
Ctrl/Cmd+Alt dragging a selected layer bar or a selected key now runs a
Quick Offset (layer.quickOffset / keys.quickOffset with fromStart, one
undo step). The mode is chosen when the drag starts and needs at least
two unlocked layers; the last selected layer moves by whole frames with
the pointer. A readout by the pointer shows the Total Offset and the
offset Per Layer, translated.
A plain drag still moves the selection together, Alt alone still scales
a key group in time (it now leaves Ctrl/Cmd+Alt to Quick Offset), and
Ctrl/Cmd+Alt+click still toggles Hold. The key drag's readout is drawn
by a shared drag_tip. Keys also register timeline.keyAt.<uid>.<ticks>,
exact where keys on subframes can share a frame.
Tests (ui_quick_offset): staggering layer bars in selection order and
keys per layer, one undo each; plain drags and Alt scaling unchanged; the
readout while dragging.
After Effects 25.4's Quick Offset staggers selected layers, or each
layer's selected keyframes, in the order they were selected: the first
stays, the last moves by the drag and the ones between by equal steps,
landing on subframes.
layer.quickOffset {delta, layers?} moves each layer's start, in and out
point by total * i / (n - 1), worked out once in ticks with integer
arithmetic, so 1 frame over 4 layers gives 0, 1/3, 2/3 and 1 frame
exactly. keys.quickOffset {delta} does the same per layer for the
selected keys, the layers in the order their first key was selected;
every selected key of a layer, across its properties, moves by that
layer's comp-time offset through its start and stretch (reversed too),
keeping values, interpolation, easing and labels. Locked layers are
skipped and at least two layers are needed. With merge and fromStart,
delta is the whole drag and each step starts again from before it, so
the drag is one undo step and a key passed over survives. A delta or a
destination out of the range of times is an error, not a clamped move.
Docs: agents.md; architecture.md notes Quick Offset as the exception to
times snapping to frames.
keys.move (and nudging) rebuilt the key selection in layer and property
order, so the order the keys were selected in was lost after a move. A
Quick Offset that follows staggers layers in that order (#1012).
The move now goes through move_key_refs, which maps each selected key to
its new time and returns the selection in its original order, with the
same collisions (a key landing on another replaces it), roving retiming
and locked-layer handling as before. key_gesture is the drag-step logic
(start again from the keys and project as they were before the drag)
split out of shift_keys so other key drags can share it.
Test: moving_keys_keeps_the_selection_order (fails before: the keys come
back sorted by layer).
The 2^28 pixel limit applied per image; one composite or layer decode kept
the ZIP output, a Vec<f32> per channel and the result alive together, and the
bake size came from the file's quad. A decode now charges every buffer
against a byte budget (4 GiB, 1 GiB on 32-bit), checks its whole expected peak
before allocating, writes channels straight into the result and returns an
error when over budget. Bakes are at most 4x the document across and 2^24
pixels.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
WaWu