Skip to content

1.17.0

Choose a tag to compare

@github-actions github-actions released this 28 Sep 06:15
· 181 commits to main since this release

The release the second audit of every public component produced. It went through all of them again, 187 findings in all, and looked hardest at what had landed since the first audit: the agent components, labelPlacement and the notch, the chart brush, the second value axis, references and CSV export, and the A2UI catalog and its adapter.

Most of what it found is accessibility, and most of that is the keyboard and the screen reader. Controls lost the focus when a press disabled them; a DataTable's column widths, column order, counts and multiple sort were out of the keyboard's or the screen reader's reach; links that open a new tab did not say so; live regions read their buttons out or said nothing at all; and small targets were pressed below a finger's size. The pickers are fields of a Form now, a resting float label stops short of what a field draws at its end, a two-axis chart reads each axis in its own units and puts its far ticks on the gridlines, and a server-rendered Calendar hydrates cleanly.

The A2UI half is mostly the catalog telling the truth. The functions declare the arguments the protocol's implementations take, a check may leave out its message, a field that is not bound can still be changed, openUrl waits for a press, a component's accessibility is read, the published catalog names the minor version it belongs to, and the adapter carries a copy of the catalog without the prose, which is about 6.5 kB less on every page that draws a surface.

Eight changes are breaking and come first. A Switch's labelPlacement is deprecated in favour of labelSide, and engines.node is >=18.20.0. Each change below says what it was and what it does now.

Breaking changes

  • ScatterChart and TimelineChart take no brush or secondaryAxis. Both type-checked and neither could be honoured: a brush windows a series by index, which on a cloud of points is an arbitrary cut rather than a range of x, and on a timeline is a cut by row while the spans on it run past the plot; a scatter's second axis wrote its numbers in the first axis' format everywhere but its ticks. They are out of both types, and ignored when passed from JavaScript. Remove them from a ScatterChart or a TimelineChart that sets them.

  • A Rating a reader sets draws its empty stars at 3:1 against the page. They were the muted ink at 40%, about 1.7:1 on white, which is under what WCAG 1.4.11 asks of a control's state. They are at 80% now, which clears 3:1 in both themes, so an interactive Rating's empty stars are visibly darker. A readOnly Rating keeps the fainter ghost.

  • AnimateCounter, AnimateScramble and AnimateSplit render a <span> by default, as AnimateTyping, AnimateHeadline and AnimateMarquee do. A <div> is not allowed inside the <span> a Statistic's value is, which is the pairing the AnimateCounter page recommends, and inside a <p> it broke hydration. The root is inline now rather than a block; pass render={<div />}, or a block class, where the block was wanted. Their ref is typed HTMLElement.

  • An Alert's live region is its title and message rather than the whole alert. role="alert" or role="status" was on the root, and a live region is read out whole, so every announcement ended with the action's and the dismiss button's names. The role, and a role you pass, go on the column that holds the title and the message now; the ref, className and every other attribute stay on the root. A test that found the alert with getByRole('alert') or getByRole('status') to read a class or a slot should find the root instead.

  • A HowToSteps with no title writes its steps' titles at headingLevel. They stayed a level below a title that was not drawn, so the default guide under a page's <h2> put its steps at <h4> and skipped a level in the outline. A guide with a title is unchanged.

  • A ContextWindow's root is a plain <div>, and role="meter" is on the ring, the label and the count. The whole gauge was the meter, and a meter's children are presentational, so the split was not read as a list and a link passed as children was not a link. The ref, className, style and every other attribute still go to the root. A test that found the gauge with getByRole('meter') to read a class or an --n-* slot should find the root instead.

  • A Gallery viewer's Previous and Next keep the focus at the ends. Pressing Next onto the last picture, or Previous onto the first, made the button disabled, and the focus went with it. Both stay in place with aria-disabled, as a Carousel's arrows do; a test that asserted toBeDisabled() on them should assert aria-disabled="true".

  • A Transfer's move buttons keep the focus when the press runs them out. Moving every ticked row made the button disabled, and the focus fell to the document, so a keyboard reader moving rows one at a time started again from the top after each. Both stay in place with aria-disabled, as a Pagination stepper does; a disabled Transfer still disables them outright. A test that asserted toBeDisabled() on either should assert aria-disabled="true".

Added

  • The A2UI adapter reads a component's accessibility. Every component accepted it and none used it, so an agent that named a group of controls named nothing. On a Flex and a Card, label makes the root a group by that name, and on a Button with no text it names the button; description describes a Flex, a Card, a Button and an Image. The components that already name themselves from a prop they draw ignore it.

  • A ScrollArea takes a label, and its viewport always has a name. The viewport is a tab stop whenever it overflows, and it had no name, so the focus landed on something that said nothing; an aria-label passed to the component named the root instead. With label the viewport is a region by that name, and without one it is a group named "Scrollable content" in the locale's words, as a ScrollZone's strip is.

  • A TreeView moves to a row by its first letter. Typing a character moves the focus to the next visible row whose label starts with it, wrapping at the end, so pressing it again steps through the rows it starts. Accents and case are ignored, as a DataTable search ignores them.

  • A DataTable's columns resize and move from the keyboard. A resize handle was pointer-only and hidden, so a column cut short was text a keyboard reader could not get back. Each handle is now a focusable separator named "Resize" and the column's label (table.resizeColumn, in every registered language), and ← → move it 16px at a time. On a reorderable table, Alt + ← → on a heading moves its column one place, and a heading with no sort button takes the focus for it. Under the pointer, a cut cell without render shows its whole text as a tooltip.

  • A Transfer says what a move did. A polite status reads out how many rows moved and to which list, "2 moved to Selected", in the locale's words. It is transfer.moved, in every registered language.

  • createNebaCatalog takes a locale. The A2UI functions that write numbers — formatNumber, formatCurrency and pluralize — followed the runtime's language whatever the components around them were set to, so a Korean surface could say "1,234.5" in a Korean sentence and pick the English plural. createNebaCatalog({ locale }) builds them in the language given, which is the one a NebaProvider should be given too.

Changed

  • A Switch takes labelSide for the side its label sits on, and labelPlacement is deprecated there. On every field labelPlacement says where the label is drawn — above the field, on its edge, inside it — and a NebaProvider fills it in for all of them; on a Switch the same name meant which side, start or end. labelSide takes those two values now. labelPlacement is still read when labelSide is left out, and is removed in the next minor release; SwitchLabelPlacement is an alias of the new SwitchLabelSide for as long.

  • engines.node is >=18.20.0. The package said >=18.0.0, and neba/a2ui imports its catalog with an import attribute that Node has only from 18.20 and 20.10. Nothing else in the package needs more.

  • The A2UI catalog names the minor version it belongs to. Its catalogId was the bare https://neba.cdget.com/a2ui/catalog.json, served from the docs, which deploy from main — so the file behind it could describe components the adapter a host had installed did not have yet. A published catalog's catalogId is …/a2ui/<major>.<minor>/catalog.json now, served at that address, and the adapter registers the same id; the bare URL is always the latest. A host that hard-coded the old id in a createSurface should use nebaCatalogId.

  • A pressable Pill lights and presses its button rather than its whole shell. The pointer light, the hover fill and the press were on the shell, so pressing a control in endIcon or inside the open details pressed the whole Pill, and the light washed across the details. They are on the pressable middle now; an outline Pill's edge no longer changes under the pointer, since the edge belongs to the shell.

  • useShortcut runs once for a held key, and leaves a key a control already answered. Every auto-repeat of a held combination ran the handler again, and a key a focused field had handled with preventDefault ran it as well, so a field's own Mod+K also opened a CommandPalette bound to the same key. Repeats are skipped unless repeat: true is passed, and still have their default prevented; a key another handler prevented is left to it. A CommandPalette that is already open no longer asks to open again.

  • An outline Toggle that is on keeps its plate under the pointer. Hover and press swapped the dyed plate for a paler wash, so a toggle that was on faded towards off as the pointer arrived. The plate stays now, and hover moves the edge to the accent and the label to the full ink.

  • neba/a2ui carries a copy of its catalog without the prose. The adapter imported the whole catalog, and about three quarters of it gzipped is descriptions written for a model, which the renderer never reads. It reads a copy with them taken out now, which takes about 6.5 kB off every page that draws an A2UI surface. neba/a2ui/catalog.json is unchanged and is still the file to hand an agent.

  • A DataTable sorted by several keys puts aria-sort on the first key only. Every sorted heading carried it, which says nothing about which key decides, and the number beside each arrow was hidden from a screen reader. The later keys' buttons end with their place and direction now, "sort 2, descending", in the locale's words (table.sortAscending and table.sortDescending, in every registered language), and each heading is still named by its label alone.

  • Home, End and the Page keys move a bounded DataTable's active row. They scrolled and left the active row behind, so the next arrow jumped back to where the reader had been. The active row goes along now without choosing anything, as it does under Ctrl and an arrow, and with Shift the four extend the run.

  • A DataTable row's tick is named after its row and is out of the tab order. Every tick was a tab stop named "Select row", so a table of a hundred rows was a hundred stops a keyboard crossed for a choice Space already makes on the active row, and a hundred controls a screen reader could not tell apart. Each is tabIndex={-1} now and is named "Select row" followed by its row's first cell. A test that found the ticks by the exact name "Select row" should match the start of the name.

  • A Chip's delete button is named after its label. Every one was named "Remove", so a row of chips was a row of buttons a screen reader could not tell apart. A chip whose label is a string names it "Remove Draft", in the locale's words (action.removeLabel, in every registered language); one whose label is a node keeps "Remove", and deleteLabel still wins over both.

  • An Avatar's picture is in the markup from the first render. Base UI fetched it with a detached probe after hydration and mounted the <img> only once it had loaded, so the picture was not in the server HTML and a loading="lazy" in imageProps did nothing. The <img> is there from the start now, laid over the stand-in and kept out of sight and out of the accessibility tree while it carries data-loading or data-error, and it fades up once it has loaded.

  • A DataList's label column stops at half the list's width. Left without a labelWidth, it was as wide as the widest label and could not shrink, so a long or translated label made the list wider than a phone's screen and squeezed the values. It is still as wide as the widest label, so the values start at the same place, up to half the list; past that a label wraps. A labelWidth is taken as it is.

Fixed

  • An Anchor with a container lands a heading exactly below its offset in Safari. The position it scrolled to could fall a fraction of a pixel off the grid, and WebKit floors a fractional scrollTop where Chromium rounds it, so the heading came to rest a pixel under the offset. The position is rounded before it is scrolled to.

  • An A2UI surface opens an address only when the reader presses something. The catalog declares openUrl as needing a user activation, and nothing enforced it: the protocol's implementation ran the call wherever an agent wrote it, and a call written into a label ran when the label was drawn and again on every change to the data it read. The adapter runs it only while a press is being handled now. A browser without navigator.userActivation (Firefox before 120) leaves it to its popup blocker, as before.

  • An A2UI field that is not bound to the data model can still be changed. The views are controlled, and the setter the renderer hands them writes only to a path, so a TextField, NumberField, Checkbox, Switch, RadioGroup, Select or Slider whose value was a literal or left out could not be typed in, ticked or moved at all. Each holds what the reader enters itself now, follows the resolved value when it changes, and still writes through when it is bound; the catalog tells the model to bind a field's value to a path, since that is where what the reader enters is sent back from.

  • An A2UI Statistic draws the delta its previousValue asks for. The catalog types value as a string, since most figures are written out, and a Statistic works out a delta only from a number — so previousValue, delta and betterWhen drew nothing unless the value was bound to a number in the data model. A value written as plain digits is taken as that number when there is a previousValue, and the catalog tells the model to write it that way; anything else is drawn as written.

  • The A2UI adapter accepts a check without a message. v1.0 lets a check leave it out, and the v0.9 schema the adapter registers with required it, so a check written exactly as the catalog allows was refused along with every component in the same batch. It is accepted now and the field shows "Validation failed"; the catalog's instructions ask the model to give every check a message all the same.

  • The A2UI catalog declares the arguments its functions take. formatDate was described with a dateStyle and a timeStyle, pluralize with a count, and formatNumber with two fraction-digit arguments, none of which the protocol's implementations read — so a model writing exactly what the catalog said got an ISO string, blank text and an unrounded number back. They are the specification's now: formatDate takes a TR35 format, pluralize takes value and every plural category, and formatNumber and formatCurrency take decimals and grouping. length and numeric ask for at least one bound, and and and or for at least two values, which is what the implementations accept.

  • A DataTable writes a boolean in a column with no render. React draws nothing for true or false, so the cell was empty while the sort, the search and the CSV all saw the value. It is written as true or false now; give the column render for words a reader would use.

  • A server-rendered Calendar hydrates without a mismatch in its tab stop. The tab stop started on today during render, and a server in another time zone has another today, so the cell the server made reachable and the one the browser hydrated could disagree. A calendar rendered on a server starts it on the 1st and moves it to today once hydration is over; one mounted in the browser starts on today as before. Pass locale and defaultMonth to one rendered on a server, as the page now says.

  • A resting float label stops short of what a field draws at its end. A TextField, NumberField or Combobox whose label rests inside it let the label run under the end icon, the steppers or the chevron, and a field only as wide as its label cut the label to "…" as it came down, since it had been sized by the label at the smaller size it has on the edge. The label now stops where the end adornments begin, and the field is laid out to hold it at both sizes.

  • The pickers are fields of a Form. A DatePicker, TimePicker, DateTimePicker, DateRangePicker, ColorPicker or TreeSelect inside a Form registered nothing with it, so its value was missing from onSubmit, a required one let the form submit empty, and an error handed to the Form under its name never showed. Each registers its value now: onSubmit gets the string it submits, or an array from a DateRangePicker and a TreeSelect with multiple; an empty required picker holds the submit and takes the focus; and the Form's error, or the field's own validation message, shows under it. A native submit sends the same fields as before, though the first input under the name is a visually hidden text input rather than type="hidden", so a test that found it by input[type="hidden"] should find it by its name.

  • A StreamingText fades in Japanese, Chinese and Thai word by word. It cut its text at whitespace, so a paragraph in a script written without spaces was one token: it faded in once, and everything after arrived with nothing moving. A run in one of those scripts is cut again with Intl.Segmenter now, and only that run, so a page in a spaced language builds no segmenter.

  • A failed ToolCall's error is headed "Failed". It sat under the heading "Result", which said the call had come back with something. The heading is the word a failed status is already read as, in the locale's words; a failed call with no error still shows its result under "Result".

  • A TimelineChart's time axis widens to hold a reference past its spans. The axis was solved from the spans alone, so a deadline after the last one was drawn off the plot, where nobody could see it. It takes the references in, as every other chart's value axis does.

  • A secondaryAxis's ticks sit on the first axis' gridlines. It was only asked for the same number of ticks, which rounding to clean numbers overruled, so a revenue axis of four intervals beside a rate axis of five drew labels between the rules, and the far edge's labels, thinned at the first axis' stride, stood beside nothing. Its range is now cut into exactly as many intervals as the first axis has, in 1-2-5 steps. A tickCount given to the second axis still rounds it on its own.

  • A CodeBlock colours a script's css` and gql` templates. highlight.js hands them to the CSS and GraphQL grammars, which nothing fetched for a JavaScript or TypeScript block, so they stayed plain unless a block in one of those languages had happened to load first on the same page. A script with one of the two tags fetches that grammar with its own now; one without either fetches nothing more.

  • A chart with a secondaryAxis describes each axis in its own units. The sentence a plot is described by took one range across every series and wrote it in the first axis' format, so revenue beside a rate read "$2 to $246,000". The first axis' series are summarised as before, and the second axis' follow in a sentence of their own, written through its tickFormat (chart.summarySecondary, in every registered language).

  • A PageLayout's skip link moves the focus into the page. The <main> it jumps to could not take the focus, so following the link scrolled to the content and left the focus behind, and the next Tab went back into the header. With scroll="content" the <main> is also the one thing that scrolls, which the keyboard could not reach. It is tabIndex={-1} now, with an inset ring that a click does not light; a tabIndex in mainProps still wins.

  • A DataTable says how many rows a search left and how many are chosen. Only a virtual table with no footer had a live count, so a search that left three rows of two hundred said nothing, and the footer's counts were text that changed with nothing reading it. The footer's range and selection count are one status now, and a table with no footer keeps a visually hidden one.

  • Every link that opens a new tab says so. Only TextLink read out "(opens in a new tab)". A ChatBubble's link preview, a MenuItem, a NavigationMenu link, an AppLogo, a BottomNavigationItem, a BreadcrumbItem, a Sources row and an InlineCitation given a target that leaves the tab now end their accessible name with the same visually hidden words, in the component's locale or the provider's. A test that found one of those links by its exact name should add the words.

  • A queued useConfirm question opens as a question of its own. The second of two questions took over the open sheet rather than opening it, so its initialFocus never ran — a danger question after a plain one came up with the focus on the destructive button — and its title was not announced. Each question is a sheet of its own now, and the focus goes back to what asked once the last one is answered.

  • The chosen option of a set is marked in a forced palette. A Radio's dot, the chosen Segment, a pressed Toggle, a chosen day or time in the pickers and the current BottomNavigationItem said so only through a fill, which a forced palette such as Windows High Contrast repaints the colour of everything around it. Each is painted in the system Highlight there now, as a checked Switch already was.

  • An InlineCitation with no link is named and reachable. Without a usable href the mark was an <a> with no href, which takes no focus and whose aria-label a screen reader may not read, so its preview was out of the keyboard's reach. It is a <span> now that writes its name as visually hidden text and takes the focus when it has a preview to open. Its ref is typed HTMLElement.

  • A ScrollZone drag, a DataTable column resize and a Mockup measure in layout pixels inside a scaled ancestor. A ScrollZone wrote the pointer's travel to its scroll position as it came, so inside a Mockup's screen or a zoom entrance the cards ran ahead of the hand. A DataTable wrote down every column's drawn width on the first press of a resize handle, which shrank them all to the ancestor's scale, and then added the pointer's travel as it came. A Mockup took its scale from its box as drawn, so one inside another scaled ancestor was scaled twice. Each reads the layout box now, and a drag divides the pointer's travel by the scale, as a Sidebar does.

  • A stacked chart's item tooltip names the segment under the pointer. Each series was measured from its own value rather than from where its segment is drawn, so two segments of 10 on one BarChart column were both "at 10" and the lower one was named wherever the pointer was. A segment is measured from the end of the one below it now, on an AreaChart as well, and a chart on its side hangs the panel at the end of the segment it names.

  • Every rule on an element in neba/styles.css's reset weighs nothing. The reset promises that a single type selector of your own beats it, and the rule that takes Firefox's red glow off an invalid field counted as a class, so an input { box-shadow: … } of your own disappeared from an invalid field in Firefox. It is wrapped in :where() like the others.

  • Escape closes a CommandPalette. Its list of commands is held open while the palette is, so the list is what heard the key, and it kept Escape from the dialog around it without closing anything — the palette stayed up with or without anything typed, and only a click outside or a command took it down. Escape closes it now, and the next open starts with an empty field, as the other ways out already did.

  • A treemap HeatmapChart's CSV has a cell in every column of every row. The file's columns are every name any group uses, and a group's row stopped at the last of those it used — so a group that did not use the names further right had a row shorter than the header, which a strict reader of the file takes as a broken line. The empty cells are written out now, as the chart's hidden table already had them.

  • An A2UI surface keeps a field when the agent puts something in front of it. A container's children were keyed by their place in the list, so a child the agent added before the others made React draw every child after it again from nothing — and a field among them lost the focus, and anything typed in it that had not reached the data model yet. Children are keyed by their id now, and a template's rows by their id and the path they are bound to.

  • The A2UI catalog says what a component does when a prop is left out. It told an agent that a Chip is sm unless it says otherwise, and it draws at md; and that a Typography is primary and aligned to the start, when with neither it takes the colour and the alignment of the text around it — which its own description of color says. A model reads a default as a fact about the renderer, so the Chip's is md now and the Typography's two are gone.

  • useElementSize, a Panes and a resizable Sidebar measure the room they have inside a scaled ancestor. All three read the box as drawn, and inside a Mockup's screen or a zoom entrance that is a fraction of the room the element is laid out in. A chart sized with useElementSize drew at that fraction and stayed there, since a transform resizes nothing and no observer said otherwise; a Panes split a defaultSize length against it and put the pane at the wrong share; and a Sidebar's handle announced the drawn width and jumped to it on the first drag. Each reads its layout box now, and a drag divides the pointer's travel by the scale, as a WindowPane already did. useElementSize reports whole pixels as a result.

  • A Stack given a transition arrives inside an Animate* that is waiting for its trigger. Each item took whether its animation was running from the wrapper around it, so a pile inside an AnimateFade with trigger="hover" sat on its first frame — invisible, for a fade — until the pointer happened to pass over it. Its items run their entrance on mount, as a Card's transition already did.

  • A Sidebar and a SidebarTrigger speak the language of the PageLayout they are in. Both take the layout's locale when they are given none, but a NebaProvider with a locale of its own filled theirs in first — so a layout set to Korean inside a provider set to English had an English trigger and an English sidebar name. The layout is nearer, and it wins now; a locale on the slot itself still wins over both.

  • A ScrollZone press let go outside the strip is over. A mouse press on the strip that was released somewhere else before it had moved far enough to be a drag left the strip listening for it, so the next time the pointer passed over the strip — with no button held — the strip followed it as if it were being dragged. The first move with no button held ends the press now.

  • An AnimateCounter, an AnimateTyping and an AnimateScramble play again when told to. A play number that goes up, or a second hover on one with a finite repeat, is meant to start the effect over, as it does on every other Animate* — but these three run in JavaScript, and their loops were keyed on having started, which they still had. The count stayed where it had landed and the text stayed typed. All three start over now.

  • A labelled Skeleton of several lines draws without a warning. Its label went into the same array as its bars with no key of its own, so every page with one logged React's warning about keys in development. The label and the bars are drawn as siblings now; the markup does not change.

  • A Card's headerAction sits at the end of the header on a right-to-left page. It was pushed along the row by a left margin, which on a right-to-left page pushes it to the start instead; next to a title the title's width hid that, and on a header with no title the action sat on the wrong side. It takes the logical margin now.

  • A vertical Tabs bar with more tabs than room scrolls. Only a horizontal bar was made a scroll container, and a vertical one with the default overflow="scroll" still faded its foot when its tabs ran past it — so the tabs down there were faded out and could not be scrolled to. A vertical bar scrolls down its length now.

  • A DataTable cell being edited keeps what was typed when its row scrolls out of view. A virtual body draws only the rows near the screen, and the row being edited went out of the document with the rest — its editor too, with no blur to commit what was in it — so scrolling away and back reopened the editor on the old value and the edit was gone. That row stays drawn, in its own place, until the edit is committed or thrown away.

  • A DataTable column being carried is marked on the edge it will land on. The mark was always on the left of the heading under the pointer, and a column carried forward lands after that heading, not before it — so the mark was on the wrong side of every forward move, and on a right-to-left page it was mirrored from both. It is on the heading's start edge for a column carried back and its end edge for one carried forward, in either direction of text.

  • Page Up and Page Down in a bounded DataTable move by a screen of it. Only a virtual body measured its own height, so a table with height or maxHeight and virtual={false} paged by nine rows whether it was a hundred pixels tall or a thousand. The box is measured on the press now, less the header that a sticky one lays over it.

  • Shift-click and Shift+arrow in a DataTable answer when the anchor is not on screen. They take the run from the row last chosen, and when that row was on another page or hidden by a search there was no run to take, so the press did nothing at all. The pressed row is chosen on its own now and becomes the anchor, as a plain press would; with Ctrl held too it is added to what was chosen.

  • Ticking a DataTable row, or clicking inside a cell being edited, is not a press on the row. The tick's click and the editor's went on to the row, so ticking a row also called onRowClick, and a double-click to select a word in the editor called onRowActivate and opened whatever the row opens. And a finger that went down on a row and then scrolled left the row marked as touched, so the next tap on that row's tick chose it alone and dropped the rest of the selection.

  • A Gallery caption is valid inside the tile it is drawn in. The caption was a <div>, and on a tile that can be pressed it is inside a <button> — or inside the <span> the picture is framed in, when it is laid over it — neither of which may hold one, which a server-rendered page reports as a hydration error. It is a <span> now; what is drawn does not change.

  • A caption or overline clamped to two lines or more stays clamped when it is aligned or given a gutter. Either prop makes those two levels a block, and the stylesheet puts block after the clamp, so it took the clamp's own display away and the text ran on past its last line. A clamp of two lines or more is a box already and is left without the block.

  • An Image given only a srcSet shows up when it loaded before hydration. A picture that finishes before React is listening is asked afterwards whether it did, and only a picture with a src was asked — so a server-rendered image given a srcSet alone, which a fast connection or a cache finishes early, stayed at nothing behind its placeholder. A changed srcSet also starts the picture over now, as a changed src already did.

  • A Drawer keeps clear of a notch at the side of the screen. safeArea held a panel's contents clear of the top and bottom insets only, so on a phone turned on its side — notch on the left or the right — a side panel's contents ran under it, and so did the ends of a top or bottom sheet. Each panel takes the insets of all three edges it runs to now, as its props row already said.

  • A Tooltip around a ContextMenu opens on React 18. The page puts the tooltip outside the menu, and the tooltip hands its trigger a ref — which React 18 drops on the way into a function component, so the tooltip had nothing to open against and never appeared. ContextMenu forwards its ref to the area it is given, as every other component already did.

  • A read-only Select with its label on the edge keeps its frosted blur. The desaturation that says "read-only" was a filter on the box around the trigger and its edge, and a filter makes an element the root of every backdrop blur under it — so the trigger blurred only that box, which has nothing in it, and read as opaque. The trigger and the edge are desaturated each instead.

  • A Fieldset inside a disabled Fieldset disables its fields too. The inner group told its fields they were enabled, whatever the group around it said, so a field inside it was disabled by the browser and still drawn available.

  • A read-only SegmentedButton does not light up under the pointer. Its segments could not be changed and still drew the pointer light, the press flash and a darker ink on hover, which is how the library says a press will do something. A read-only set keeps its chosen segment and draws none of the three.

  • A TreeSelect takes a provider's variant and density, and a ProgressBox its locale. Both pages promise it, and neither did: the TreeSelect passed the two to its shell without asking the provider for them, and the ProgressBox never asked for a locale, so a formatted value was written in the runtime's language. The check that holds every component to the provider's axes now reads the shared props an interface extends as well as the ones a component names, which is how these two got past it.

  • An inline ColorPicker behaves in a form as the popup one does. A disabled inline picker still submitted its value, and a required one that was empty let the form go, because the popup's shell is what did both and an inline picker draws none. It does both itself now: its value is left out of the form while it is disabled, and an empty required picker holds the submit back and takes the focus to its square.

  • A FilePicker and an inline ColorPicker inside a disabled Fieldset are disabled. Base UI's Fieldset stops the fields it knows, and neither is one of them: a file dropped on the picker still reached onFilesChange, the inline picker's square and rails still answered the keys and the pointer, and both were drawn available. Both read the Fieldset now, as the other fields already did.

  • A Reasoning, a ToolCall or an AgentStep that timed itself reports the time it took. The clock ticks once a second, and when a run ended without a duration the total was whichever tick had come last — 4.9 seconds of thinking read "Thought for 4s", and anything under a second had no figure at all and read "Finished thinking". The clock is read once more when the run ends, so the total says what a duration of the same length says: "4.9s", or "900ms".

  • An AgentSteps chain's running marker counts the step under way, not the whole chain. The marker running draws stands for whichever step is happening now, and it stayed mounted as finished steps were appended in front of it — so its clock went on counting from the first step, and the fourth step of a chain that had been going a minute said it had been running a minute. It starts again with each step appended before it.

  • A PromptInput fits its height to its text when its width changes. The field was measured only when its value changed, so the same words wrapped onto more lines in a narrower window and were cut off, and a field that first rendered hidden — in a closed panel or an inactive tab — measured a box of no width and kept a height of nothing once shown. It measures again whenever its width changes, and when minRows or maxRows does.

  • A PromptInput's onKeyDown can keep a key from sending. The field sent first and called the caller's handler after, so a mention list or a slash-command menu open over the field could not take Enter to choose an entry — the message went with the half-typed name in it. The caller's onKeyDown runs first now, and a preventDefault there means the key does not send.

  • A disabled or readOnly PromptInput takes no dropped files. Its shell stayed a drop target whatever state the field was in, so a file dragged onto a field that took no typing still lit the edge and reached onFiles. It is off while the field takes no input, as a FilePicker's zone already was.

  • A PromptInput that takes files leaves dragged text to its field. Its shell answered every drag as a drag of files, so text selected somewhere and dragged into the message lit the drop edge, was never inserted, and called onFiles with an empty list. The shell of a PromptInput and a FilePicker answers only a drag that carries files now, and a drop of nothing but folders is not reported at all.

  • The chevron on an Accordion, a Collapsible, a ToolCall, a Reasoning and a Sources header turns when it opens. Each read the open state off itself, and Base UI writes it on the trigger and on nothing inside it — so the chevron pointed down whether the panel was open or shut, and only the header's colour said which. It reads the trigger's state now, and turns over as the panel opens.

  • A PieChart with one slice left draws a whole ring. The gap that parts two slices was taken off the ends of every slice, including one with nothing beside it — so a pie whose other slices were zero, missing or hidden from the legend had a notch cut into it where it started. A slice that is alone is drawn whole.

  • A flat series is drawn when gradient is on. The fade along a line was measured against the line's own bounding box, and a series that holds one value all the way along has a box with no height, which SVG answers by ignoring the gradient — so the stroke was painted with nothing and the line vanished. The fade is measured across the plot now, which is the same thing for every line that spans it.

  • A chart's value axis no longer opens an empty step past its data. The axis rounds its ends out to the step its ticks are on, and dividing by a decimal step is rarely exact: 2.4 / 0.2 is 11.999…, which rounded down to a step below the data, and 0.28 / 0.02 is 14.000…2, which rounded up to one above it. A chart whose lowest value was 2.4 drew an axis from 2.2. The ends are rounded with the same guard the ticks already had.

  • A chart's CSV leaves an element xAxis.label out of its header. The label heads the first column of the file, and one written as an element went in as the element's JSON — {"type":"em","key":null,…} in the top-left cell. Only a string names the column now, as a ScatterChart, a TimelineChart and a DataTable already had it; anything else leaves the cell empty.

  • A chart brush draws a series of any length. The strip found the lowest and highest values of the series it outlines by spreading them into Math.min and Math.max, which passes every value as an argument, and past somewhere between a hundred and two hundred thousand of them the browser throws — so the chart failed to render on exactly the long series a brush is for. It walks the series instead.

  • A brushed chart keeps its strip clear of an axis given a thickness. The strip is laid over a band the chart reserves at its foot, and a bottom axis' thickness replaced that whole band rather than the axis' own share of it, so the strip was drawn over the axis labels. thickness is the axis' room and nothing else, as its description says, and the strip's band is added under it.

  • A series on a secondaryAxis is labelled and pointed at on its own scale. Its value labels were written in the first axis's format, so a rate beside a revenue read $4; tooltip={{ mode: 'item' }} measured the pointer against every series on the first axis's scale, so a rate near the top of the plot lost to a revenue near the floor; and a horizontal chart placed its tooltip, and chose which side it hangs on, on that scale too. All three read the series' own axis now, and a value label goes through secondaryAxis.tickFormat as the table and the tooltip already did.

  • A secondaryAxis label is drawn above the plot rather than over it. It is written along the top of the chart, at the far end of the line a yAxis label starts, and that line was only made room for when the yAxis had a label too — so on its own the name sat over the top of the plot and the highest points under it. The top is taken for either name now. On a vertical chart the right-hand band no longer widens for a name that was never drawn in it, so the plot gains that width back.

  • A chart reference on the category axis stays on its own column under a brush. Its value is an index into the whole series, and the plot placed it as an index into the window — so with the window on days 10 to 19, a rule at day 10 was drawn off the end of the plot, and the hidden list read it out under the wrong category's name. It is placed and named against the whole series now. A category reference outside what the plot draws, windowed or not, is left out rather than drawn past the axis, and a band is cut at the plot's edge.

  • A chart brush reports a window only when it moved, and never one past the series. onRangeChange was called for every key and every pointer move, including a Home on a handle already at the start and a drag that stayed inside one category, with the window it already had. And a window one category wide at either end of the series reported a handle at -1 or one past the last category. Both ends are held inside the series now, and a gesture that changes nothing reports nothing.

  • An A2UI Typography's lines is held to what the catalog says. The catalog tells an agent lines is a whole number of at least one, and the renderer took any number at all — a 0 or a 1.5 was drawn rather than refused. The Zod the adapter derives from the catalog now carries minimum and integer, and a keyword it does not know throws when the schemas are built rather than being passed over.

  • A controlled CommandPalette binds its key once. Its opener was a window listener of its own that was taken off and put back whenever onOpenChange changed, and an onOpenChange written inline changes on every render of the page around it. It is bound through useShortcut now, which is what that hook's documentation already said, and which keeps the handler in a ref.

  • A chart's marks fade and grow under the crosshair, as they were meant to. The one transition every mark shares was written as three strings joined at run time, and Tailwind only generates a class it can read whole in the source — so the class was never in the stylesheet, and every mark snapped: the series a legend dims, the datum under the crosshair, the pixel it grows by. A PieChart and a HeatmapChart, which had moved before the marks were given one shorthand, snapped along with the rest. It is one literal now.

  • A chosen today is still marked as today. A day cell dropped aria-current="date" and its dot as soon as it was chosen, so a screen reader stopped saying which day was today exactly when it was the one picked. It keeps both now and is drawn as chosen, with the dot turned white on the fill, as the page and the month grid already had it.

  • A TreeView branch that is shutting is out of a screen reader's reach. Its rows stay in the document until the collapse finishes, which is what lets them fold away at a height, and nothing stopped a screen reader walking into them in the meantime — or for good, when the branch was shut while the tree was not displayed and no transition ever ended. The branch is inert from the moment it starts to shut.

  • A HowToSteps rail is a list to VoiceOver. It is an <ol> drawn with list-none, which is enough for Safari to stop calling it a list, and a row's place in it went with it. It says role="list" now, as a List, a Pagination and a Breadcrumb already do.

  • A HowToSteps row draws its whole focus ring. The rail is a scroll box whenever it is horizontal or given a maxHeight, and a scroll box clips on both axes, so the ring drawn just outside each row lost its top and bottom on a horizontal rail and its sides on a bounded one. It is drawn inside the row's edge now.

  • A pressable Pill draws its whole focus ring. The button inside inherited no radius from the row around it, so its ring was square, and it sat two pixels outside a button the lozenge clips — so the start of the ring, the end without an endIcon, and every corner were cut off. The ring is drawn inside the button's edge now, at the lozenge's own radius, as an Accordion header's is.

  • An Image or a Gallery tile that opens a preview says so. The button is named after the picture, which is right, and so it was read as "A ridge of hills, button" with nothing to say a press opens a larger view. It carries aria-haspopup="dialog" now; a Gallery tile that only calls onItemSelect does not.

  • Copying a CodeBlock a second time is announced a second time. A press inside the two seconds the button says "Copied" wrote the same word into the live region again, and a live region only speaks when its text changes, so a reader who pressed again to make sure heard nothing. The region is emptied and written again a moment later; the button's own label does not blink.

  • A CodeBlock that scrolls is named even when its title is a node. Only a string title named the focusable region, so title={<code>app.ts</code>} on a block tall enough to scroll left a tab stop with no name. The region is named by the title the toolbar draws, and by the language or "Code" when no toolbar draws it.

  • Moving through a Gallery viewer says what the picture shows. The live region under the picture announced only "Image 3 of 12", and the focus stays on the arrow, so the new picture's title was never read. The announcement starts with the title, or the alt when there is none, and is read as one sentence. What is drawn does not change.

  • A ScrollZone whose content fits is not a tab stop. The strip took the focus whether or not there was anywhere to scroll it, so a page of chip rows that all fitted was a tab stop per row that did nothing. It is focusable while it overflows, which is when the arrow keys have something to move, and keeps its role and its name either way.

  • A resizable Sidebar says its width again after it has been a drawer. The separator's aria-valuenow, aria-valuemin and aria-valuemax were written once, so a sidebar that collapsed into a drawer and widened back into a column — or first mounted as a drawer — came back with a separator that said none of the three.

  • A Tour jumps to each step's target under a reduced-motion preference. It scrolled the page to the target with behavior: 'smooth' whatever the reader had asked for, which is the page gliding under them. It jumps instead when the preference is set, as a ScrollZone already does.

  • Opening a ColorPicker or a TreeSelect takes the focus into its popup. The shell they share leaves the focus alone on opening, because a date picker's calendar moves it into the grid itself — but these two move it nowhere, so a reader who opened one from the keyboard was left on the trigger, and a searchable TreeSelect did not take what they typed next. The focus now lands on the first thing in the popup that takes it: the colour square, the search field or the tree.

  • A ColorPicker's sliders are read out as what they measure. The square was read as two bare percentages with nothing to say which was saturation, the hue rail as a bare number of degrees and the opacity rail as a number with no percent sign. The square now reads "Saturation 100%, brightness 90%" from a new areaValue label, in every registered language, and the two rails are written by the platform in the reader's own — "217 degrees" and "55%" in English.

  • A ColorPicker's square and rails answer the keys a slider answers. The hue and opacity rails took only the left and right arrows, and none of the three took Home, End, Page Up or Page Down — so a screen reader that adjusts a slider with the up and down arrows moved nothing. A rail now takes either arrow pair, pages by ten and goes to its ends; on the square Page Up and Page Down change brightness and Home and End take saturation to either end.

  • A chart's brush is easier to reach and to tell apart. Its two handles are drawn 8 pixels wide and are now pressed across 24. Page Up and Page Down move a handle a tenth of the series, where the arrow keys alone took hundreds of presses to cross a year of days, and a key never takes a handle past either end of the series. The strip is a group named after the chart, so two brushed charts on one page are no longer four sliders called Start and End.

  • A Rating's stars are pressed at a finger's height, and across the gap between them. Each star took presses only on its own glyph, 20 pixels at md, and the gap between two stars took none. The target now grows to 24 pixels tall and across half the gap on each side, so it never reaches the next star. With precision={0.5} the two halves of a star meet in the middle and only grow up and down, which the page now says.

  • The × that takes a file off a FilePicker's list is pressed at the size of a finger. It is drawn at 1.3em of the file's name, about 17 pixels, with no larger target around it. It carries the same invisible 24-pixel target as a Chip's × now, and nothing drawn moves.

  • The × in the corner of a Dialog, a Popover, a Drawer and a Tour step is pressed at the size of a finger. It is drawn at 1.6em of the sheet's text, about 20 pixels at the default size, and unlike the × on an Alert, a Toast or a Chip it had no larger target around it, which WCAG 2.5.8 asks to be 24 pixels. Nothing drawn moves.

  • The working ring slows down under a reduced-motion preference. A Button that is loading, a TextField that is checking and a running ToolCall, AgentStep or Reasoning turned their ring on Tailwind's animate-spin, which the preference did not reach, so a transcript with several running rows kept several rings at full speed. The ring turns at an indeterminate ProgressCircular's rate now, 0.9 seconds a turn, and slows to 2.4 seconds with it.

  • A PromptInput that sends on Enter says so to a phone's keyboard. The field sets enterKeyHint="send" when submitKey is Enter, so the key is labelled as the send it is rather than as a line break. A form gives an <input> that label on its own and never a <textarea>, which is what the component and its page had assumed. A caller's own enterKeyHint still wins.

  • A PromptInput given an id keeps its accessible name. The id reached the textarea, but the visually hidden label still pointed at the one the component had generated, so the field was announced with no name. The label follows the id the textarea ends up with.

  • Answering an Approval hands the focus to the answer. The buttons are replaced by the line that records what was chosen, so the one that was pressed left the document with the focus on it, and a keyboard reader started again from the top of the page without hearing what they had agreed to. The record line takes the focus when a press is what answered it, which also reads it out; a decision that arrives later, after the focus has gone somewhere else, leaves it there.

  • A DataTable searched, cleared and searched again folds its rows once. The folded text every search is matched against was dropped whenever the field was emptied, so the first character of the next search normalised every searchable cell of every row again — the cost the fold exists to pay once. It is kept until the rows or the columns change.

  • Moving a chart's brush no longer rewrites its hidden table. The table holds every point whatever the window shows, but its props were rebuilt from the windowed series on each frame of a drag, so every cell of a long table was written again for nothing. They are built from the whole series now and stay the same while the window moves.

  • A ScatterChart and a TimelineChart export what their tables hold. Both wrote the file every other chart writes, a grid of categories against series, which fits neither: a scatter's second series was written beside the first one's x values and its sizes were left out, and a timeline wrote each row's name beside the number one. A scatter's file is a row per point with its x, its y and its size, and a timeline's is a row per span with its label and its two ends as dates.

  • A chart stacked to full exports the caller's numbers. The CSV was written from the shares the bars are drawn at rather than from the data, and a share is worked out against the series that are shown — so the file changed with the legend, and a hidden series was written as its share of the others (20 shown and 60 hidden came out as 300). A window on the chart switched the file back to the raw numbers. The file and the brush's outline are written from the caller's numbers now, which is what the hidden table already said.

  • A chart's brush stays when its window lands on a stretch with no data. The plot drew its empty state, and the strip, the hidden table and the export button were drawn only when the plot had something in it — so a window dragged, or started with defaultRange, onto the week a sensor was offline took away the one control that could move it back. They follow the whole series now, and only the plot says it is empty.

  • A DataTable that opens rows or edits cells without choosing them tells a screen reader which row is active. Such a table is a grid with an active row the arrow keys move, but aria-activedescendant was only written when there was a selectionMode, so the ring moved and nothing was announced. It now points at the active row in every table the keyboard can move through.

  • An A2UI Card or Button takes an elevation. The catalog declares it as 0 to 3, and the adapter built its schema with z.enum, which takes strings and refused every one of them — the default 0 included. The processor then threw for the whole updateComponents message, so a surface with one card at any elevation drew nothing, and the data model sent after it was lost too. An enum of numbers is a union of literals now.

  • Pressing a PromptInput's stop button stops the answer and sends nothing. The button turned from type="button" into type="submit" inside its own click, as onStop set submitting back to false, and the browser read the new type when it came to activate it — so the form was submitted and whatever the reader had typed while the answer was being written went out as a new message, restarting the answer they had just stopped. The button is a submit button in both states now, and the form decides between stopping and sending. It also stops with a required field left empty, which a submit button would otherwise be held back by.


Every release is in the changelog on the docs site and in CHANGELOG.md at this tag.