Demos that open a second window, take files, move the window, read the
clipboard or set native menus show a note on the web in the blocked style.
IBM Plex Sans Regular joins the assets, unmodified, where gpui's svg
renderer loads it: gpui_web answers .SystemUIFont with it. The gallery's
web entry registers a Hebrew cut of Noto Sans Hebrew. scripts/web.sh builds
the bundle with wasm-bindgen and wasm-opt -Oz, 20.0 MB, and fails past 25 MiB.
The gallery's entry runs natively or, on wasm32, from a wasm_bindgen start
through gpui_web, single-threaded with WebGL2 fallback. ?page, ?story and
?theme choose what shows; an unknown one fails in the page and the log.
set_theme switches the mode later. Pictures load from ELY_GALLERY_ASSETS.
scripts/stories.py writes examples/gallery/stories.json from the pages.
Every page and scripted state was shot with gpui 0.2.2 and with the
port, light and dark, full width and at 280px, judged pair by pair by
agents with skeptics, and every page's log diffed. What differed is
fixed, each with a test that fails without it where the test platform
can show it:
- The gallery's harness takes posted mouse moves (gpui_macos listens
to a tracking area), posts each drag move twice (it no longer
replays a held drag), and waits three frames before each shot.
- Rolling digits: a block column, so taffy 0.13 lends the cell no
baseline, and a line of whole device pixels, since gpui snaps each
length on its own. DescriptionList's row aligns ends.
- Margins that taffy 0.13 collapses through a plain block: the
drawer's pull and snap, the stagger and transition rise, the dialog
and palette rises, the fullscreen dialog's slide.
- An Input keeps room to type in a squeezed row; the field rows drop
w_full, whose reason was taffy 0.9's.
AGENTS.md's rules that named gpui 0.2.2 or taffy 0.9 were re-tested
against both sources and rewritten, and new ones say what the port
changed. TASKS.md's tags follow: what the new gpui makes possible is
prove. Captions, doc comments and README say what holds now.
The calendar's grid clears its preview on leave, and ignores a leave
that follows a key. gpui records the exit a key makes before the
handler runs, so the guard swallowed it and a real exit afterwards
had no change left to report: the stale day stayed the preview. The
grid now listens independently of input, while its days keep the
default, so a key is no leave and a real exit is. A test leaves the
grid after an arrow, a letter and a month key.
shaped()'s doc goes back to its one line; AGENTS.md keeps the why.
gpui and gpui_platform now come from Zed's repo by git at 1a28cff,
the code gpui-pre 0.3.7 republishes, so the gallery can later build
for the browser. gpui takes stacker, as 0.2.2 always did, and
gpui_platform takes font-kit, without which macOS draws no text;
runtime_shaders moves to gpui_platform.
The API moved: focus, blur and Tab steps take cx, Corner is Anchor,
shaped lines paint with an alignment, defer_draw takes a content
mask, the inspector registers a factory, async updates return their
value, and wrapped lines are shared, not cloned. Updates through a
window context go through the window, so a closed window still ends
them as it did.
Where the new gpui behaves differently, Ely keeps its old intent.
Holds and reveals under a resting pointer survive typing through
gpui's input-modality-independent hover; previews the keyboard also
moves keep the default, and the calendar's leave ignores a key press.
Enter reaches a button inside the code editor, since gpui now arms a
key click on the press. The event logger skips a lone modifier. A
video and an image grid keep their shape in a flex column under
taffy 0.13, as a picture does.
AGENTS.md records the new stack and the rules this port changes; the
rest of the old gpui and taffy rules are re-tested in T44c.
The site shows every component running, as gpui-kit.com does, never
a screenshot. gpui 0.2.2 cannot build for the browser, so T44b moves
gpui to Zed's repo at 1a28cff, T44c re-checks every page and rule on
it, and T44d builds the gallery for wasm32 as one story per
component. The screenshot items give way, and the site deploys to
Dokploy.
A sweep that read every render found 28 roots still on w_full,
reached by 64 components: the field button under Select and every
picker, the checked and mention fields, the slider track,
NumberInput, Steps, the state views, the agent previews, and every
chart and finance frame. All drop it. LivePreview's size_full becomes
h_full, since its card no longer fills.
Every host of those roots was listed and read, 231 in all. Five
tests that mounted a picker as the window root sit in a size_full
box, the index manager gives its method select a flex_1 box, and the
notification center and the gallery's state frames stack their state
views in a column so they stretch.
Each dropped root carries a debug selector, and six chapter tests
mount 43 components in a padded card whose column is measured by
content, through layout::tests::narrow_width: each root must take the
full 240px. With w_full back on the roots, all six fail.
AGENTS.md names the roots that keep w_full and why: picture frames
that keep a ratio, and rows that hand a text field its share.
Eleven more roots drop w_full: the docked ActionBar, MessageBubble's
three arms, PromptInput, TitleBar, Tabs, the horizontal Divider with
and without a label, DropZone and Playlist. Their hosts in the crate
are blocks or flex columns. The gallery's docked ActionBar sits in an
items_center column, so the demo gives it the width.
AspectRatio keeps w_full, and AGENTS.md says why: taffy 0.9 follows a
ratio from a set width, not from a stretch, and without one the image
gallery's tiles and the ratio demo collapsed.
Every page and scripted state was shot again at full width and at
280px and compared with the shots from before T44a. Nothing collapsed
or shrank. The mail reader's attachments now wrap side by side, as
their row asks.
145 component roots drop w_full, the early returns included. Three
field rows keep it, Input, InputGroup and TagInput: stretched without
a width, taffy 0.9 laid the field's flex_1 share at no width in the
template picker at 280px, and the placeholder drew nothing. Container
keeps it, since mx_auto needs a width to fill.
Hosts that relied on a root's width now give it: MessageList boxes
each row at the list's width, the splash gives its bar the column's,
and test stages that mount a component as the window's root sit in a
size_full box. Tests hold a narrow row across the list and the splash
bar across its column; each goes red without its fix.
The gallery gains --narrow <px>, which lays each page in a padded
card at that width and shoots it without its script. The page
header wraps: the title takes the room and the theme switcher drops
below, where a long summary used to push it past the edge. A file
panel step waits for the panel to become key instead of a fixed
pause, which lost a race in a capture.
The audit now holds fg_subtle on active, a selected row's fill, and
each syntax color over the editor's surface, a code block and the
current line. What it found is fixed at the source: fg_subtle is
#696764 light and #94928f dark, syntax.comment follows it, and the
light number, constant and attribute darken to #a3591b, #aa5529 and
#9f5d12. The token viewer adds avatar and container sizes, draws a
container at its share of the widest so it stays in its row, and
names what it shows: the shared colors and scales. Tests pin every
variant, the hue each chart row reads and the code rows.
DesignTokenViewer shows every token of the theme in the mode shown,
named as code reads it: colors with their hex, syntax, chart and
terminal colors, the type scale, radii, control and icon sizes,
elevations and spacing. A11yChecker audits each text color over the
surfaces it sits on at 4.5:1 and the focus ring and chart hues at
3:1, failures first. The audit found fg_subtle under 4.5:1 on hover
and overlay; it is #6f6d6a light and #8c8a87 dark now, and a test
holds all four Ely palettes to the audit. IconBrowser points to
forms::IconPicker and the ScreenshotTest Harness to the gallery's
capture, each with a section on the tooling page.
FpsMeter keeps the stamp before the last second, so a frame longer
than the span shows as the worst and a bar. The inspector's panel, a
second root out of the app's scope, sits in a FocusScope of its own,
so Tab goes on from Pick. EventLogger takes key downs through an
interceptor, which runs before bindings, and marks its box with a key
context to know focus is inside, so a field's Backspace and Enter log.
Tests hold the stall, Tab both ways through Pick, bound keys in a
field, a key while focus is outside, newest first and a bare border.
ZacharyZhang-NY