Consume the pending changesets into a minor release. Users get the
terminal renderer, a headless Dialog with an Escape layer stack, default
Tab focus navigation, focus-visible styles and a headless Button,
createPortal, onLayout and onScrollChange, native scrollbars, new CSS
styles, registerFont, BGRA image uploads, getElementsAt, idle GC,
display-link frame pacing on macOS and Windows ARM64 builds.
Bump the native Cargo crate, napi loader version checks, native, React
and Solid packages to 0.11.0, @gpuix/cli to 0.1.2 and disktree to 0.2.1.
Fixes#73Fixes#74Fixes#75Fixes#76Fixes#77Fixes#78Fixes#84Fixes#91Fixes#93Fixes#95Fixes#96Fixes#97Fixes#98Fixes#99
Discord: https://discord.com/channels/1422625037164351591/1562091690376962111
Session: ses_ea68cd52affex3PxfjqF4PhYyY
Track visited element IDs during each locator collection and skip subtrees
already traversed. Preserve depth-first tree order and strict descendant
matching while preventing duplicate counts and false ambiguity errors.
Add a real native-tree regression covering count, all, element, waitFor,
multiple nested selector levels, sibling ordering, and genuine ambiguity.
Document the locator result contract and add the native patch changeset.
Validation: all seven automation tests pass against the release addon.
Fixes#99
Session: ses_ea76fa70affeF580gmCkDnrEtT
Discord: https://discord.com/channels/1422625037164351591/1562029343889428490
Lowercase the complete text and query consistently so Greek final sigma
matches exact uppercase words. Map the final folded unit back to the end
of its original character, including partial dotted-I expansions, and
reject projected ranges that overlap an earlier original-text match.
Treat every Unicode mark category as part of a word in both JS and Rust.
Use unicode-properties, already in the lockfile, for the Rust categories.
Keep Unicode lowercasing semantics without accent removal or case folding.
Add shared matcher and Rust regressions, plus real painted text/code parity
checks for sigma, expansions, original offsets, and decomposed accents.
Update the README contract and the pending native search changeset.
Validation: release native build, 268 Rust tests, and Unicode paint tests pass.
The full React suite has one separate display-link wake timeout.
Fixes#95Fixes#96
Session: ses_ea76fa70affeF580gmCkDnrEtT
Discord: https://discord.com/channels/1422625037164351591/1562029343889428490
`render(<App />, { terminal: true })` or `GPUIX_TERMINAL=1` runs the same
GpuixView on gpui_terminal, a GPUI platform in the Zed fork that paints with
the kitty graphics protocol. No native window opens.
```
macOS Linux
renderer.init ─► init_macos renderer.init ─► init_threaded
EmbeddedPlatform::Terminal gpuix-ui thread
TerminalPlatform::new_embedded() gpui_terminal::application().run()
tick() ─► pump_events()
```
- renderer.rs: `terminal_requested()` reads the option and the env var;
macOS swaps MacPlatform for TerminalPlatform inside the embedded path, so
tick() and the frame wake hook are unchanged. No per-method branches
- input.rs: `init_terminal` unbinds ctrl-c in text field contexts so it
reaches the platform and quits
- `isTerminal()` exported from @gpuix/native, @gpuix/react and @gpuix/solid
as a plain function (not a hook); browser.mjs returns false
- examples/chat.tsx, mail.tsx drop the traffic-light gap in a terminal
- AGENTS.md: terminal mode swaps the platform; no hooks for non-reactive values
- README: "Run in a terminal" section and the `terminal` option row. The same
README hunk also carries a note on `windowBackground: "blurred"` tinting
- docs/terminal-renderer.md: architecture, traps, research notes
- zed submodule: gpuix-terminal commits (gpui_terminal crate, embedded mode,
SIGWINCH-only forwarding)
- changeset (minor for native, react, solid)
Discord: https://discord.com/channels/1422625037164351591/1562055543466496043
Session: ses_ea7105a47ffeNLpB02sEyo5t9g
Wake JavaScript through a weak, coalesced N-API callback after GPUI queues its display frame. Keep tick non-blocking and replace the independent 8 ms input timer with a 100 ms no-wake fallback. Explicit frameMs retains timer polling; Windows and Linux keep their existing loop.
Clear callbacks and timers on stop, make stop idempotent, and release the hook when the renderer drops. Exercise real display wakes through the shared runtime on Bun and Node, plus hidden-window fallback and replacement-loop cleanup. Document the latency and idle-CPU limits for both adapters.
Fixes#91
Session: ses_ec38cf159ffe0vKPJHE10v0gl6
Discord: https://discord.com/channels/1422625037164351591/1560050965288194129
Windows ARM64 timer scheduling can take more than 40ms for three frames. Wait for the existing termination callback and retain exact tick and termination counts instead of asserting before the loop completes.
Session: ses_ed48f90d4ffeqym03XRjqsA5xw
Discord: https://discord.com/channels/1422625037164351591/1558853977012174980
Derive logical screenshot coordinates from the native viewport instead of a fixed 1280px width. Assert Dialog portal and backdrop coverage against that viewport and preserve exact popup size and centering assertions. Select ARM64 installer assets in the updater fixture on ARM64 runtimes.
Validation: React build and all 536 tests pass on macOS. Keep all screenshot geometry assertions and all platform tests enabled.
Session: ses_ed48f90d4ffeqym03XRjqsA5xw
Discord: https://discord.com/channels/1422625037164351591/1558853977012174980
The list ignored left and right padding, so rows spanned its full width. Rows now sit inside all four sides, and a scrollbarColor thumb stays at the list's own edge, past the padding. Moves the zed submodule to the commit that lays rows out inside horizontal padding.
A padded box around a list is better served by padding on the list itself,
with the thumb at the list's own edge. The list now takes padding on all
four sides.
Each render() call wrapped the tree in a Provider whose value was a new
{ renderer } object, so every call was a context change. In the chat example
that re-rendered ChatApp, whose inline onRetry closure then busted the memo on
Transcript: 305 SafeMdxContent renders and about 9 ms per find-bar keystroke.
The example also keeps onRetry stable now, so typing in the composer no longer
re-renders every row.
Allocate initial child lists only for parents and drop them after materialization. A real 5k-turn chat retains about 2 MiB less JavaScript heap. React lifecycle and integration suites pass.
Session: ses_f1be8a987ffeslJ1dooQ4pgDnx
Before, any `focusVisible` style (even `{}`) returned early from
apply_focus_visible, so one key controlled two states: the look of the
element while it has keyboard focus, and whether it dims while another
control has it. A custom focus ring therefore could not dim.
Now the two are separate:
- `focusVisible` only styles the element while it has keyboard focus
- every other focusable element off the focus path dims to 40%, with or
without a `focusVisible`
- new style key `keyboardFocusDim: false` keeps one element at full opacity
(StyleDesc.keyboard_focus_dim, not inherited, excluded from state styles)
- `Select.Content` and `Dialog.Popup` (React and Solid) switch from
`focusVisible: {}` to `keyboardFocusDim: false`
- the window-wide `render({ keyboardFocusDim: false })` option is unchanged
Test: the focus scene now checks that `keyboardFocusDim: false` opts an
element out and that an element with a focusVisible ring still dims.
README Focus styles and the pending focus changeset describe the new rule.
Session: ses_f231682b0ffe2fr5xaNLPPWQre
GPUI stores color images as BGRA (RenderImage, Metal and DirectX atlases).
setImagePixels only took RGBA, so every upload swapped bytes 0 and 2 of
every pixel on the UI thread. BGRA sources (ffmpeg -pix_fmt bgra,
VideoToolbox, node-canvas toBuffer('raw')) now skip that pass.
- native: PixelFormat enum in img.rs; napi, test renderer, UiCommand and
wasm bindings take an optional format. Unknown formats throw
- host types: ImagePixelFormat, ImagePixelsOptions
- react/solid: setImagePixels(w, h, pixels, { format })
- waveform example writes BGRA by hand
- README: format docs, node-canvas raw is BGRA premultiplied (opaque only)
- test: rgba and bgra of the same color paint byte-identical screenshots,
misread bytes swap red/blue, unknown format throws
Session: ses_f230bc6fdffe5CqhxBXqCeacz9
`render(<App />, { keyboardFocusDim: false })` (React and Solid) and
`createTestRoot({ keyboardFocusDim: false })` (React) stop GPUIX from dimming
every other focusable element while a control has keyboard focus. Apps then
draw their own focus look with `focusVisible`.
How it reaches native:
- new batch op `["setKeyboardFocusDim", bool]`, parsed in apply_batch_to_tree,
so macOS, Windows/Linux, web and the test renderer share one path with no
per-backend method and no pending state while the web window boots
- stored as RetainedTree.keyboard_focus_dim (default true)
- keyboard_focus_path() returns None when it is off, so apply_focus_visible
never dims
- each root enqueues the op on creation (React createRoot, Solid attach), so
a bun --hot remount without the option restores the default
Docs: README Focus styles section documents the option and the 8px pointer
slop from the GPUI bump; the pending focus changeset gets both bullets.
Test: the styles.test.tsx focus scene with keyboardFocusDim: false paints the
same frame as idle after Tab.
Session: ses_f231682b0ffe2fr5xaNLPPWQre
Instead of marking the focused control, dim the rest: while a control has
keyboard focus (last input keyboard, focused handle is not an input or
textarea), every other focusable element renders at 40% of its own
opacity. The focused one stands out, and every Tab target is visible.
- GpuixView::keyboard_focus_path computes the focused control and its
ancestors once per render; BuildCtx and CustomRenderContext carry it.
Ancestors stay undimmed because gpui opacity covers the subtree.
- A focused text field dims nothing: typing is keyboard input too.
- An element with its own focusVisible (even {}) is never dimmed.
- GPUI refreshes the window on focus and modality change, so the path is
recomputed without any extra wiring.
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Remove the default focusVisible ring and the inherited focusRingColor
style. Without its own focusVisible:
- a focusable control (Button, tabIndex div) renders at 70% of its own
opacity while keyboard-focused, which shows where Tab went without a
ring (FOCUS_VISIBLE_DIM, paint only)
- <input> / <textarea> draw nothing; the caret already shows focus
apply_focus_visible takes a FocusVisibleDefault (Dim | None) instead of
an inherited ring. outlineWidth/Color/Offset stay for apps that want a
ring via focusVisible.
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
<input>/<textarea> track the same focus handle on the editor wrapper and
on the inner text element. GPUI's TabStopMap inserted a second order node
for the repeat, so prev() from the field returned the field itself and
Shift+Tab did nothing (Tab worked, next() skips both).
Fixed in the fork (upstream tab_stop.rs has the same code): a repeated
handle is recorded for replay but adds no second node, with a unit test.
- zed: c0d597f -> 672c259 (gpui: count a focus handle tracked twice as one tab stop)
- regression test: Shift+Tab / Tab around a textarea
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
A ring around the composer textarea after a click looked bad. Browsers
ring a clicked <input>, but GPUIX now shows focusVisible and the default
ring only after keyboard input, for every element. FocusVisibleRule is
gone; apply_focus_visible always uses gpui's focus_visible.
Test: a clicked input matches a clicked input with the ring off (the
caret differs from idle); Tab into the input still shows the ring.
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Drop the `focus` style key. `focusVisible` now matches CSS
:focus-visible exactly, including the rule browsers apply to text fields:
- controls (div with tabIndex, Button): GPUI .focus_visible(), so only
after keyboard input
- <input> / <textarea>: GPUI .focus(), so also after a mouse press,
because a field in use must look active
apply_focus_visible(el, style, ring, FocusVisibleRule) applies either the
declared style or the inherited default ring, right after track_focus,
and replaces apply_default_focus_ring. apply_interactive_styles is back to
hover + active only. The refinement borrows the style (gpui runs it
eagerly), so nothing is cloned per frame.
Test: a clicked input now shows the ring; a clicked div still does not.
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
- examples/chat.tsx and examples/solid/chat.tsx: every div + onClick
control (IconButton, SidebarAction, ConversationRow, thread filter,
WorkedFor, ModeToggle, GhostButton, search hits, cycle-overlay) is a
Button, so the chat is fully keyboard reachable. OverlayCard is a Dialog:
Escape, backdrop press, Tab trap, and focus in (search input via
initialFocus) and back out to the opener. Overlays stay mounted with
open={...} instead of conditional rendering.
- example-app/app.tsx: IconButton, SidebarRow, Checkbox, Add are Buttons;
Add is disabled while the draft is empty.
- Select.Content and Dialog.Popup set focusVisible: {} so the default ring
does not outline the whole menu or card; they are focused only to take
keys.
- @gpuix/solid exports the HostElement type for refs.
- chat.test.tsx: Escape closes search and focus returns to its button.
Solid chat port done by a subagent; verified with chat.automation.ts
(6/6) and a throwaway test of open, Escape, close, and focus return.
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Focus styles work like hover: nested style objects that map straight to
GPUI's .focus() and .focus_visible() refinements (focus_visible already
means keyboard-only, via window.last_input_was_keyboard()).
A ring needs to appear without moving layout. A border changes the box
and a zero-blur box-shadow fills the whole rect, so the zed fork gains
Style::outline (width, color, offset), painted after the border as a
border-only quad grown by offset + width that follows corner radii.
GPUIX exposes outlineWidth / outlineColor / outlineOffset.
Default ring, like a browser: every element that tracks a focus handle
(div, input, textarea) gets a 2px accent focusVisible outline unless its
style sets focusVisible. focusRingColor is an inherited style prop
(transparent turns it off), resolved through Inherited like selectionColor.
Button (React + Solid), Base UI shape: tab stop, role=button, onClick on
press, Enter keydown (not isHeld), Space keyup; disabled leaves the Tab
order, focusableWhenDisabled keeps it; state style function. buttonProps()
is shared; Dialog.Trigger / Dialog.Close now use it. GPUI has no
aria_disabled, so none is emitted.
- zed: c6531c8 -> c0d597f (gpui: add Style::outline)
- tests: default ring on keyboard only, no layout change, transparent off,
custom focusVisible; Button activation in both adapters
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Second oracle review of 6382a80 found five issues:
1. A pending focusElement request survived Tab / clicks and stole focus
when its element appeared later. PendingFocus now records the focus at
request time and is dropped if focus moved any other way first. The
check runs before the frame's autoFocus, so an explicit request still
beats same-frame autoFocus (live and test renderer now agree).
2. React read previous focus in a layout effect; in the test renderer
that read flushes and an autoFocus could already hold focus. The
DismissableLayer now reads it during render (React) / component body
(Solid), before its elements are committed, and passes it to
pushDismissLayer({ previousFocus }).
3+4. Each Popup moved focus on its own in effect order, so nested
dialogs opened together focused the outer one and closed together
lost focus. DismissLayer gains initialFocus() / finalFocus(previous);
the host calls them in stack order: only a layer landing on top takes
focus, a parent inserted under same-commit children rewrites their
return target to itself, and a child whose parent already closed skips
its restore.
5. FocusTarget function results now accept refs.
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Dialog.Popup takes Base UI's initialFocus and finalFocus: true (default),
false (stay), a ref or element, or a function returning one of these.
The shared resolveFocusTarget lives in @gpuix/native/host.
- initialFocus defaults to the popup itself. Base UI picks the first
tabbable element, but GPUI's tab map exists only after the popup paints,
so the first Tab walks that order instead.
- finalFocus defaults to the trigger, else the element focused when the
popup opened. The Popup records it on open, so a dialog opened from app
state (no Trigger, e.g. chat.tsx overlays) returns focus on close.
- Popup no longer uses autoFocus; it focuses imperatively after reading the
previous focus.
Native fix found by the Solid test: sync_focus_handles took
pending_focus_element even when the handle did not exist yet, so a frame
rendered before the element's batch dropped the request. That is the
normal order for Solid (microtask flush) and for Windows/Linux (UI thread).
The request now stays until the handle exists. The test renderer's
focusElement uses the same request_focus path as the live renderer.
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Oracle review found two bugs in the layer stack commit:
1. React runs child layout effects before parent ones, so a Dialog and a
nested Select opening in one commit registered Select first. The stack
put the Dialog on top and one Escape closed both. Layers now carry a
parent (DismissableLayer component + context in React and Solid), and
pushLayer inserts a parent below any child that registered first.
2. A prevented Tab still closed an open Combobox, because the input closed
on its own Tab keydown before an ancestor could prevent the default.
The Combobox now closes when focus leaves the input, like Base UI, so
it closes only when Tab really moved focus.
onBlur was untestable: the offscreen test window is never active and gpui
reports focus paths only for an active window. The zed fork gains
Window::set_active_for_tests (same body as the platform active-status
callback, extracted into active_status_changed), and the test renderer
calls it. Tooltip open-on-focus is now covered too.
- zed: 81c99f8 -> c6531c8 (gpui: let tests report a window as active)
- useDismissLayer hook replaced by DismissableLayer component
- regression tests: nested defaultOpen layers, ancestor-prevented Tab
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Keyboard defaults run in the shared JS host, after every element onKeyDown
and the window onKeyDown, so any handler can cancel them:
- Tab / Shift+Tab move focus (GPUI focus_next / focus_prev)
- Escape closes the top layer of a per-window dismiss layer stack
- KeyEvent gets preventDefault / stopPropagation / defaultPrevented
- render({ tabNavigation: false }) turns the Tab default off
Why JS and not a GPUI KeyBinding: the event callback is a non-blocking
ThreadsafeFunction, so JS runs after GPUI finished dispatching and could
never veto a native binding.
Native:
- keystrokeId on key events, bumped in a root capture listener, so JS can
group one keystroke's element events and its trailing window event
- root listener always installed; stops Tab after all element listeners so
the platform never types "\t" into an input
- anchored: explicit transparent background paints no fallback fill
- bounds_tracker pinned at inset 0; bounds were shifted by padding
Primitives (React + Solid):
- Select / Combobox / Tooltip triggers are tab stops, also with asChild
- Select popup traps Tab; outside press no longer steals focus back
- Escape handling moved to the layer stack
- new Dialog: Root, Trigger, Portal, Backdrop, Popup, Title, Description,
Close (Base UI shape), modal Tab trap and focus restore
- Solid asChild keeps the child's handlers and children
Session: ses_f2717eddcffeXQkI5JXN8eAnFD
Consume the pending changesets into a minor release. Users get the official Solid 1 renderer, AnimatePresence exit animations, live image uploads and scrollIntoView on host refs, native window controls, and an OS file picker.
Bump the native Cargo crate, native npm package, React package, Solid package, generated napi loader version checks, and Bun workspace lock to 0.10.0. Keep published adapter dependencies on @gpuix/native as workspace:^. CI is the only publish path for native platform binaries.
Fixes#68
Session: ses_f3b7fae0dffeZjEBH48TZSQVBq
Drop the previous GPU image when src or a live upload replaces it, and when the img node is pruned. Without drop_image, every setImagePixels frame leaked a RenderImage.
scrollIntoView now uses the virtual-list logical index, not the windowed child index. A windowStart of 50 plus child 5 now scrolls to item 55.
Attach setImage, setImagePixels, and scrollIntoView in getPublicInstance instead of createInstance, so concurrent React can abandon a host node before the methods exist. Pass the original buffer through instead of copying it with Buffer.from.
The waveform test compares two live phases. Encoded setImage and windowed scrollIntoView have screenshot and list tests.
Session: ses_f3c07efdbffeHwbaQhsE7fIz5z
When every exiting child reported complete, the map still held those
keys as true. A later exit of the same keyed child then hit the
early return and never called onExitComplete or swapped rendered
children back to the present set.
Clear the map after a completed batch so the next exit can run.
Session: ses_f3d2f1310ffeo5K8hcRmsuYLWO
`<img>` refs can push image bytes without a data URL or JSON mutation.
`setImage` takes encoded PNG, JPEG, WebP, GIF, SVG, BMP, TIFF, ICO, or
Netpbm. `setImagePixels` takes packed RGBA. Prefer pixels for a live
waveform. There is no density argument. Upload a 2x bitmap into a 1x
layout box on retina.
Every host ref now has `scrollIntoView()`, which scrolls the nearest
overflow parent or `<virtual-list>` until that node is visible.
The waveform example paints a generated 2x RGBA buffer through
`setImagePixels`.
Session: ses_f3c07efdbffeHwbaQhsE7fIz5z
Stop copying napi wire types in `@gpuix/native/host`. Re-export
ElementBounds, HighlightMatch, PathPromptOptions, EdgeInsets, and
overlay stats from the generated addon types.
Helpers now ask only for the methods they use:
- createMutationQueue takes MutationHost (applyBatch)
- window observers take WindowSizeHost / WindowInsetsHost
- selection observation takes SelectionHost
NativeRenderer stays the adapter-facing type: required applyBatch,
plus live renderer methods as Partial<Pick<GpuixRenderer, ...>>.
Fakes can still implement applyBatch alone. Do not subclass the
napi GpuixRenderer.
Session: ses_f3c88f82affeXgsQdmQW4wpizp
Resolve the native renderer through the package browser condition and expose CPU throttling from a browser-safe subpath so both public adapters bundle for WebGPU without importing node:module.
Add browser bundle coverage to the native package suite and run those package checks in CI after both adapters are built.
Make Solid motion completions generation-aware, reject stale targets, and keep presence removal idempotent. Replace the native motion invariant panic with a safe target fallback.
Session: ses_f4102f732ffesSfYYRRHlXA2j9
Tie motion completion to the generation that started each target so queued events cannot complete a replacement animation. Preserve partial target values and finish no-op, invalid, and offscreen exits.
Harden AnimatePresence participant registration, cancellation, and keyed sibling handling with focused native-backed regressions.
Session: ses_f3d2f1310ffeo5K8hcRmsuYLWO
Tommy D. Rossi