From `expanded` a standard side sheet sits in the layout beside the
content, and the view took it out of the layout the moment it closed:
`data-md-drawer-collapsed` (`display: none` on the root) followed `! open`
at once, so its fade was never seen and the content beside it took the
sheet's width, and the gap, in one jump — in every engine. M3's side
sheets: opening a standard sheet shrinks the body beside it, and closing
gives the room back.
drawer.css: the standard root is a clipping flex box whose `inline-size`
springs between none and the sheet's width, with a negative margin as
wide as its flex parent's gap (`--md-drawer-gap`), on the sheet's own
exit and entry timings; the sheet keeps its width at the root's far edge,
so it is uncovered from its inner edge while it fades. drawer.blade.php:
`settle()`, from `x-effect`, measures the gap on each change and holds
the collapse (`closing`) for the root's and the sheet's closing
durations, read a frame on; reopening lets a pending end go by. The
binding reads `collapsed`, which only `settle()` and the window's width
write: a binding reading `open` ran before `settle()` in the same flush
and, already queued, did not run again once `closing` changed.
ContainmentTest closes the showcase's sheet at 1000px and samples, in the
page, for the root part-way to none while the column beside it has grown
part of the way, then checks the collapse, the column at the row's full
width, and a reopen part-way through the exit ending fully open; it fails
on the previous code in Chrome, Firefox and Safari. OverlayTest pins the
root's declarations and the view's bindings.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`fullScreen` followed `open`, so the moment a full-screen search closed,
`data-md-full-screen` went with it: the fixed header bar went back to its
resting pill — or to `display: none`, behind the search icon — and the
view, still fading out under Alpine's hold, dropped to the docked layout
for its exit. In every engine the bar vanished on the first frame while
the view went on fading somewhere else.
search.js: a close from full screen now `hold()`s the layout — `leaving`
keeps `fullScreen`, and with it `data-md-full-screen`, the back arrow and
the fixed bar, for the view's own closing duration, read from its
computed style a frame on (zero under reduced motion); reopening lets a
pending end go by. search.css fades the header bar out with the view on
the view's spring and keeps the root above the page while it leaves,
and the icon trigger's bar stays displayed until the layout settles. The
focus trap now binds to `open && fullScreen`, so it still lets go at the
close and its return of focus lands inside RETURN_GUARD_MS rather than
reopening the view at the end of the exit.
PickingTest samples both exits in the page — the icon trigger, and the
bar trigger on a compact window — for a moment where the root is still
full screen with a fixed bar and both bar and view part-way through their
fade, then checks the settled layout and focus back on the icon or the
field; both fail on the previous code in Chrome, Firefox and Safari. The
tests wait for the entry's own transitions first: Firefox reads a
transition's end value until its next refresh tick.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
collapse.css animated `<details>`' `::details-content` from `block-size: 0`
to `auto`, which needs `interpolate-size: allow-keywords` (Chrome 129 and
nothing else) and a `content-visibility` held through the close with
`allow-discrete` (not Firefox), so in Firefox and Safari the section
snapped open and shut. No stylesheet can do it there.
resources/js/collapse.js animates the `<details>` itself, the same way
in every engine: a Web Animation of its `block-size` from the height it
is drawn at to the height it is going to, on the fast spatial spring's
tokens, with `overflow: clip` and a `min-block-size` that keeps the
summary whole through the spring's overshoot, all inside the keyframes,
so nothing is left in `style`. Content below moves with it. Opening sets
`open` at once and grows the height; closing keeps `open` while the
height goes, marks `data-md-collapse-closing` so the chevron turns back
at the start, and drops `open` (and fires `toggle`) when the height has.
A press mid-way turns it round from the height it has reached. The CSS
animation is gone, so Chrome takes the same path and never animates
twice.
Routes: a press on the summary (pointer, Enter, Space) is taken over;
the Alpine and Livewire bindings call `materialCollapse()` from the
view's `x-effect` (its first call, as Alpine starts, sets the state at
once); a `name` group's open member closes on the same spring, its
`name` lifted for the close so the browser's exclusivity does not shut
it first and put back once closed. Anything else that sets `open`
(find-in-page, an application's own script) stays instant, as the
browser draws it, and is followed. Under reduced motion every route is
the browser's own.
Browser tests in Chrome, Firefox and Safari catch the height part-way
open and shut (and what is under it moving) from a press and from the
Alpine binding, the chevron turning at the start of a close, the clip
while it moves, Enter on the summary, a reversal mid-way, a name group,
and reduced motion. The four motion tests fail on main in Firefox and
Safari.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The side sheet, the bottom sheet's panel and the docked search's scrim
and view kept `display` alive through their exit with
`transition-behavior: allow-discrete`. Firefox does not transition
`display` (Chrome 117 and Safari 18 do), and `x-show` sets
`display: none` in the frame the exit starts, so in Firefox the sheets
vanished instead of sliding out and the search view and scrim vanished
instead of fading. The docked search had a second problem in every
engine: neither the view nor the scrim had a closed state to transition
to, so where `display` was held (Chrome, Safari) the view stood at full
opacity for its duration and then disappeared.
Each element now carries `x-transition:enter`/`:leave="md-transition"`,
the approach the two sheet scrims already took (renamed from
`md-scrim-transition` to one name for all of them). The class only
switches Alpine to CSS-transition mode, so `x-show` holds `display` for
the element's computed transition-duration before hiding it, in every
engine, and a reopen during the exit cancels the pending hide; nothing
styles it. `display` and `allow-discrete` leave the transitions so
Chrome and Safari do not hold a second time. Alpine reads the first
`transition-duration` listed, which is the closing slide or fade in each
list (the preset panel lists translate before height). The search view
now closes back into the bar (opacity 0, `scale: 1 0.9`, the reverse of
its `@starting-style` entry) and its scrim fades out on close and when
the search turns full screen. Under reduced motion the durations are
zero and every one of them closes at once.
Four browser tests sample each exit mid-way in the page, in the same
round trip as the close: the sheets part of the way to their closed
offset, the search scrim and view part of the way faded, each still
displayed, then `display: none`. All four fail on main in Firefox (the
two search tests in Chrome too) and pass in Chrome, Firefox and Safari.
OverlayTest pins the drawer's new transition and the view's markup.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A rail that was open over the page left it on `transition: … display …
allow-discrete`: the compact adaptive rail's slide-out, the slide-out of
a rail that hides when collapsed, and every rail scrim's fade. Firefox
does not transition `display`, even with `allow-discrete` (Chrome 117
and Safari 18 do), so there the panel and the scrim vanished on the
first frame. Unlike the sheets' scrims these are drawn by attribute
rules on `data-md-open`, not `x-show`, so Alpine had nothing to hold.
navigation.js now marks a rail that closes `data-md-closing` — `sheet`
when the panel leaves the window, `scrim` when it stands in the layout
again — and drops it once the panel's `translate` and the scrim's
`opacity` transitions have finished (at once under reduced motion, or
when the rail opens again). navigation-rail.css keeps what is leaving
displayed, and a sliding panel in its open geometry, while the
attribute is set, and no longer transitions `display` anywhere, so
Chrome and Safari do not hold a second time. The view sets it from
`x-effect`, beside the `x-bind` that drops `data-md-open`, so both
attributes change in one flush: a `$watch` a microtask later let a
style read in between settle the scrim as already hidden. The collapsed
branches and `--md-navigation-rail-value` are untouched.
NavigationTest samples each exit part-way in the page (the compact
slide, the hide-when-collapsed slide, a modal rail's scrim) and checks
it ends hidden; all three fail on main in Firefox and pass in Chrome,
Firefox and Safari. They wait for the entry to finish first: Firefox
creates a transition on its next refresh tick, so straight after a
change its computed style already reads the end value. NavigationRailTest
pins the holds and that no rule transitions `display`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Menus, submenus, tooltips, rich tooltips and the FAB menu held their
exit with `transition-behavior: allow-discrete` on `display` and
`overlay`. Firefox transitions neither (MDN browser-compat-data,
`display.is_transitionable`: Chrome 117, Safari 18, Firefox none), so
every one of them vanished on its first frame there. A script cannot
hold a popover open instead: `beforetoggle` is not cancellable on the
way out, and the browser's own light dismiss (Escape, a press outside)
never asks.
resources/js/popover-exit.js: a popover marked `data-md-popover-exit`
closes for real at once — focus, aria-expanded and toggle stay the
browser's — and a copy taken in `beforetoggle`, while it is still drawn,
stands in for the exit. The copy is decoration: a manual popover in the
top layer (closing no other popover), inert, aria-hidden, without ids or
nested popovers, `x-ignore`d so Alpine starts nothing in it, pinned to
the popover's box with its resolved colours. It is shown with its
transitions off, so `@starting-style` does not replay the entry, then
marked `data-md-popover-closing`, which each stylesheet turns into its
closed values (`:popover-open:not([data-md-popover-closing])`, and the
FAB menu's items' sink), so it moves on the component's own tokens. It
is removed once the longest of them has run, and opening the popover
again takes it away. Under reduced motion every duration is zero and no
copy is made. `display`, `overlay` and `allow-discrete` leave the
transitions, so Chrome and Safari take the same path.
Browser tests in Chrome, Firefox and Safari slow the motion tokens so a
round trip still finds the exit on screen: a menu after Escape and after
a press outside (the real menu closed and focus back on its button, the
copy inert, fading, with no Alpine state, and gone after), a reopen
part-way through, reduced motion, a submenu while its menu stays open, a
tooltip, the FAB menu's items part-way down their sink, and a persistent
rich tooltip.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
With nothing focusable inside, showModal() focuses the <dialog> itself.
Chrome's scroll containers are keyboard-focusable, so there the scrolling
body became the focus delegate and modal.css drew its inset ring; Firefox
and WebKit focused the dialog, where the arrow keys scroll nothing, and
WebKit never lets Tab reach a scroll container, so a keyboard could not
read a long text-only dialog there at all.
<x-modal> now opens through materialShowModal() (resources/js/dialog.js):
showModal(), and when the dialog took the focus itself and its body
overflows, the body gets tabindex="0" and the focus, as in Chrome. The
tabindex goes when the dialog closes. The body is wire:ignore.self, like
the dialog: a Livewire render would otherwise morph the tabindex away and
WebKit dropped the focus out of the dialog. ContainmentTest now also
checks the focus stays on the body through a render.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two gaps found rewriting the Boost skill for 2.0's plain CSS, each
closed the way M3 describes it.
`<x-loading size="96">`: M3 gives the loading indicator a responsive
size, 48dp by default and flexible from 24 to 240dp ("never exceed that
range"), with the container and the active shape in a fixed ratio. The
size had no prop, so an application wrote a width and height of its own;
`size` now takes a whole number of px in that range, written as
`--md-loading-size`, and the SVG keeps the 48:38 ratio as it scales. A
value outside the range is ignored, as `<x-icon>` and `<x-shape>` ignore
theirs, and an application's own width and height still win.
`--md-navigation-rail-value`: M3 Expressive's rail has two values,
collapsed and expanded (Compose's WideNavigationRailValue), and content
in a rail follows it. Without a hook an application copied the rail's
seven conditions — mode, `data-rail`, `data-rail-auto`, open, and the
window band — out of navigation-rail.css. The rail now publishes the
answer from the same branches that narrow it: `expanded` by default,
`collapsed` wherever it is drawn collapsed, so a style query in the
application's CSS switches at the first paint and at the same moment
as the rail's own items. A rail open over a scrim reads `expanded`, and
outside a rail the property is unset. A JS attribute would have missed
the first paint; a width container query would have lagged the collapse
animation.
Style queries on a custom property need Firefox 151, so the browser
floor moves from Firefox 147 to 151 (README, the CI note, UPGRADE's new
2.1.0 section); Chrome 125 and Safari 18.4 are unchanged.
Browser tests pin both in Chrome, Firefox and Safari: the indicator's
drawn box at 96 and 32px, and the rail's value — with a style query
acting on it — at the first paint for fixed modes, across the window
classes and the menu button for an adaptive rail, and open and closed
for a modal one. NavigationRailTest pins the value in each of the five
collapsed branches.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2.0.0 took Tailwind out of the stack, but the texts Boost copies into
every application still taught its utilities: the core guideline said
"built on Tailwind CSS", the always-on material-3 guideline wrote every
rule as `bg-primary`, `type-*`, `rounded-corner-*`, `state-layer` and
`medium:`, and both skills' tables and examples did the same. None of
those classes exists in 2.0's stylesheets, so an agent following the
guideline wrote markup that compiled to nothing. Found moving ReStride
onto 2.0: its CLAUDE.md, and SealShare's, carry these texts.
The guidelines and skills now teach what 2.0 has: component and layout
component props (`gap="space200"`, `hide-from="medium"`), the `md-type-*`,
`md-ink-*` and interaction classes, and `--md-sys-*` tokens in the
application's own CSS, with breakpoints as range media queries.
UPGRADE.md §1, §2 and §4 describe the finished move instead of the
in-between state, and header comments that pointed at the removed
tokens/utilities.css and tailwind.css, or called a component "still
Tailwind", say what is true now.
BoostVocabularyTest runs DesignGuard over the code in every shipped
guideline and skill (fenced Blade and CSS, and each inline class list),
so a text that teaches a class the stylesheets do not define fails the
suite; it finds 268 violations in the texts as 2.0.0 shipped them. The
guard's own table of what it reports is exempt. BoostResourcesTest now
asks the design skill for each breakpoint's prop value and media query
instead of the removed `medium:` variants.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`ItemSegmentedContainerColor get() = ColorSchemeKeyTokens.Surface`
(ListTokens.kt, C-21) was taken literally, but foundation/base.css
paints the page `surface` too — so a segmented list was invisible on any
default page. The token reads against a Compose scaffold, not against
this package's own page.
The tile is `surface-container` now, the step M3 names for a container
that has to read against the page, and what 1.x drew. Found in
SealShare: the files chosen for an upload list under the drop zone with
nothing to set them apart from it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
M3's colour list for a standard toolbar ends "Standard button
(Primary)" and the vibrant list ends "Standard button (On primary
container)" (N-13) — both name the *standard* button. The two rules
painted every icon button in the toolbar instead, so a button that
brings its own container got a primary label on a primary fill and its
icon disappeared: SealShare's floating navigation marks the current page
with a filled button, and that button came out a blank circle.
A button that fills its own container is now left alone — filled, tonal,
or either of them selected — and keeps the label colour that goes with
the fill. A disabled button still falls back to button.css's disabled
ink in a toolbar of either colour, as before. AppBarTest pins the
exclusion on both rules.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Plan step 46, M3 § Tabs: "Labels: single row by default; may wrap to a
max second line if needed with truncation, or use scrollable tabs to
give longer titles more room." In fixed tabs (<x-tabs>, <x-section-nav>
up to four sections) the label's text, now its own data-md-tab-text in
both views, wraps and truncates at two lines. The tab's 48px (64px
stacked) is a minimum, the bar stretches its tabs to one height, and the
content stretches with them, so the indicator stays on the divider.
Scrollable tabs keep one row. SealShare's "Two-Factor Auth" and
"Appearance" showed an ellipsis at 600px.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Found while doing plan step 46: withoutFontFace() searched the raw
bundle, so the `@font-face` that foundation/base.css's and
error-page.css's header comments mention read as the rule. Each cut from
inside its comment to the end of the next block and left the comment
open, so the no-build fallback lost base.css's html rule and icon.css's
rules. The search and the brace count now read Stylesheets::mask(), and
only tokens/font.css's real block leaves.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Plan step 46, from SealShare's breakpoint walk: an application imports
the stylesheets its own views render, and none renders the error layout,
so under a real build its 403/404/500 pages drew unstyled. With a build
the page now keeps the app's Vite tags (foundation, tokens, scheme, font)
and inlines Stylesheets::bundle() of components/error-page.css and its
imports beside them (ErrorPage::layoutStyles(), cached). Without a build
the fallback is unchanged. An application imports nothing for its error
pages; one that imports all.css gets the same layered rules twice, which
is harmless.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
missingStylesheets() names an import the views need, but nothing named
one they had stopped needing: when SealShare's last card left its views,
card.css stayed in every page until a reviewer noticed by hand.
unusedStylesheets($cssEntry) reads the entry's direct package imports
and reports each one no scanned view needs, counting what a needed
stylesheet imports itself and foundation.css always; an entry that
imports all.css is left alone. Opt-in, beside missingStylesheets()
(plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
<x-card> always titled itself with an <h3>, so a card straight under a
page's <h1> skipped a level, which SealShare's download page did with
no way round it from the application. `heading` takes h2 to h6, or div
or p for a title that is not a heading, as <x-pane> and <x-app-bar>
already do; h3 stays the default. The skill also says a card holds one
subject, as M3 does (plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A transition-duration or -timing-function listing one token per property
(a spatial spring for scale, an effects spring for opacity) was reported
as literal: blanking each var() left a bare comma, which is not one of
the safe values. Every item of a list is now checked on its own. A
message also quoted the blanked shell, `var( ), 200ms`; it now
quotes the declaration as written. SealShare's drop zone found it
(plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both applications show values a person reads or types character by
character: SealShare's share tokens and recovery codes, ReStride's file
and folder names and its keyboard shortcuts. The package wrote the
platform's monospace stack twice (the reset's <code>, <kbd>, <samp> and
the mono field) and gave everything else nothing, so an app copied the
stack by hand. --md-ref-typeface-mono now holds it, read by the reset,
the field and a new md-mono text class; the guard points font-mono at
<code> or md-mono. Not an M3 token: M3 names a brand and a plain
typeface and nothing for code. Decided with the user (plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
<x-stack> stretches its children across by default and <x-form> is a
one-track grid, so a button written directly in either drew as a
full-width pill. M3 keeps a button's width "dynamic to fit label" and
says not to "stretch buttons into long flat shapes on large windows".
A button there now sits at the start edge at its own width; a stack
aligned start, centre or end is untouched, and an application's own
width still wins (a cap on the button itself was rejected for exactly
that: it would have limited an app's inline-size, which CascadeTest
guards). Chromium test in both directions, failing without the rules.
Decided with the user (plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The showcase and the error pages set grayscale font smoothing on
themselves, but the foundation did not, so an application that dropped
Tailwind's antialiased class rendered heavier text on macOS than the
package's own pages, silently. base.css now sets it on the page; the
showcase's and error page's copies go, and the guard tells a leftover
antialiased class to go too. Decided with the user (plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The guard said a directive does not compile inside a component tag but
not what to write instead, and the skill's note said only "use {{ }}".
SealShare's two-factor challenge had @js in an x-data on <x-stack>, and
its implementer had to work out {{ Js::from() }} from Laravel's source.
Each message now names the form: Js::from for @js, json_encode for
@json, Arr::toCssClasses and toCssStyles for @class and @style,
$wire.entangle for @entangle, a bound attribute for @disabled and its
kin, and a prop, an echo or a plain element for control structures
(plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
SealShare's only navigation is a floating toolbar at place="bottom". In
1.x it set --material-bottom-bar to lift the snackbar over it; in 2.0.0
the toolbar reads that variable to place itself, so an application can
no longer set it, and the snackbar landed on the toolbar. M3 nudges a
snackbar "upward to avoid overlapping FABs/docked toolbars" and never
puts one in front of navigation.
A bottom-placed toolbar now publishes --material-bottom-toolbar, the
distance from the window's bottom edge to its top: floating 64px (as
tall as an md or lg FAB beside it), docked 64px plus the safe area it
pads, rounded docked floating from 840px. It is declared on :root and
again on the scaffold's root, the only place its formula sees the bar.
The snackbar clears whichever of the bar and the toolbar reaches higher,
and a page pads its end with it. Two Chromium tests, both failing on the
old formula: a docked toolbar on a scaffold's navigation bar, and a
floating toolbar without a scaffold (plan step 46).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Action/Input/Containment/NavigationStylesheetsTest.php each carried the same
"imports the stylesheet of every component its view renders" check on their
own group's dataset. StylesheetsTest.php now runs it once, over every
resources/views/components/*.blade.php backed by a stylesheet of its own
(icon and shape included, no longer needing a group to run in at all;
tabs/tab's shared stylesheet, N-16, still handled) — a new component gets
the check by existing, not by being added to the right group's array too.
Each group file keeps its own dataset for the checks that stay group-shaped
(shape, tokens, breakpoints, the block it imports from).
Plan step 42 (Phase F), Part A.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Measured the way the showcase and the error page's fallback actually serve
it — Stylesheets::bundle() in PHP, then PHP's own gzencode() — so CI holds
the line with no Node build. Currently 106,838 bytes gzipped; the budget is
120,000, about 10% of headroom for organic growth before the failure
message points a maintainer at what grew.
Plan step 42 (Phase F), Part A.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Audits step 42's "every component view has a stylesheet" against the whole
of resources/views/components/: everything already has one, apart from two
that need none, now documented in the test itself — theme-script renders an
inline <script> only, and tab is drawn by tabs.css, its parent <x-tabs>'s
own stylesheet.
Plan step 42 (Phase F), Part A.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
StylesheetsTest.php's new test runs DesignGuard's own family table (check
(i)) over resources/views, resources/js, src, tests/Browser and
tests/Feature — the guard turned on the package's own source, so the two
never drift apart. tests/Feature/DesignGuardTest.php and
tests/Fixtures/design-guard/ carry Tailwind on purpose and are left out.
The scan found Tailwind-shaped class strings used as caller classes in 21
test files (a component test proving a caller's class lands on the root or
a caller's size wins, mostly), left over from before Tailwind cleared the
whole stack: replaced with neutral application-style names (`app-hero-icon`)
or, where one already carries the right meaning, an `md-*` class
(`md-text-end`, `md-ink-warning`). BreakpointsTest.php's own heredoc probe —
which needs a genuine Tailwind-shaped breakpoint prefix to prove its
detection works — now builds that one string from a placeholder at runtime,
so its own source stays clean for this same scan.
Plan step 42 (Phase F), Part A.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Check (iii) flagged two shapes the package's own stylesheets (and their
stylesheet tests) already treat as fine: `linear`, M3's own easing for
continuous motion (an indeterminate progress indicator's rotation), and a
box-shadow that is only an inset or outline-style ring — `[inset] 0 0 0
<n>px` in a `var(--md-sys-color-*)` role, or one `color-mix()`-dimmed for a
disabled ring — the shape a focused field's edge or a day's "current"
outline draws. A blurred shadow or one with a literal colour is still
reported.
Plan step 42 (Phase F), Part B.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
outsideTheScale() read ease-in/ease-out/ease-in-out/duration-* line by line
across every file, the way it still reads its other scale steps — but a
plain CSS keyword read out of an API into a JS string (matchMedia(…) ?
'linear' : 'ease-out') names a real value, not a Tailwind class, and the
same line scan read the words back out of a stylesheet test's own regex
literals (`ease-in`, `ease-out`) as if they were Tailwind too. Tested both
ways per the plan's instruction: Blade/PHP class lists have the identical
false-positive shape, so the two families move to FAMILIES, matched only
through a class list (literalClasses()/tailwindFamilyHint()) as the newer
families already are, in Blade and PHP as much as in JavaScript.
Plan step 42 (Phase F), Part B.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Different checks land violations for the same file in their own pass order
(colours before shadows before media queries, an application's CSS always
appended after its views) rather than the order a maintainer would read the
file in. DesignGuard::violations() now sorts everything it returns
deterministically before handing it back, and a fixture proves it crosses
both dimensions: a later-processed application-CSS file whose path sorts
first, and two different checks landing on the same line in the opposite
order from how they run.
Plan step 42 (Phase F), Part B.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review: "colour the theme does not declare `text-red-500`"
was the one family (i) finding without a 2.0.0 hint; it now says the
utility compiles to nothing and points at md-ink-* or var(--md-sys-color-*).
The foundation-only fixture view also carries an <x-dynamic-component> and
an <x-slot:title>, which the stylesheet check must neither report nor trip on.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review: the committed smoke test scanned the package's views
without the stylesheet check, so the claim that the package, showcase and
Workbench pass it through all.css was untested; it now runs it, and
checks the showcase's plain tags do need imports from a foundation-only
entry, so the pass is not vacuous. Doing so exposed that a component class
one test evals leaked into every later check in the process: the shadow
check now guesses the class the way Blade does (Blade::component()
aliases, component namespaces, the application's own root namespace), and
the test gives its class a namespace of its own.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review, check (iii): calc() hid every literal inside it and
now only one built on a var() does (min(), max() and clamp() too), so
calc(8px + 4px) and clamp(1rem, 2vw, 2rem) are reported. Declarations are
read whole, so a last one without a semicolon is seen and a custom
property named --card-border-radius is not mistaken for border-radius;
the font shorthand is checked; a 0s duration is fine. Media queries read
only width features: a height or prefers-* condition is not a breakpoint,
and (600px <= width < 840px) passes. The type hints no longer suggest
font-size: var(--md-sys-typescale-*), which is a font shorthand value.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review: an application's CSS entry may import what a Vite
build resolves but Stylesheets::bundle() refuses (tailwindcss mid-
migration, a font URL, a layer() import); resolvedFiles() threw on each,
failing the application's whole design test. It now follows relative
imports only and skips the rest. Each missing stylesheet, shadowed tag
and ->links() is reported once, at its first use, instead of at every
occurrence; a CSS entry that does not exist says so; the @import line goes
through vendor/ when Composer symlinks the package.
missingStylesheets() also switched on the literal-value check for the
entry's whole import graph, vendor stylesheets included, with no way to
check imports alone. Check (iii) now reads the .css files scan() is given,
like every other check; the entry's imports still feed the class
exemptions, now escaped selectors (.hover\:underline) included.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review: run over SealShare's and ReStride's views, family (i)
missed about a quarter of the class tokens, most of them 1.x's own
utilities (type-body-md, text-meta, rounded-corner-lg, focus-ring, link)
and common Tailwind families (relative, inset-0, z-10, overflow-hidden,
border, opacity-50, shrink-0, -mx-4, cursor-pointer, font-mono). Each now
has a hint, and the ones with a layout component name its prop: gap-4 is
gap="space200" (16px), a step between two of M3's names both, items-center
align="center", grid-cols-3 :columns="3"; variants say which selector or
media query replaces them. Shapes are tightened to real Tailwind values,
so an application class such as content-header or p-card is not matched.
Class lists are also read from wire:*.class, x-transition:*, unquoted
:class keys, double-quoted PHP strings, 'class' => '…' and plain PHP
files; a string compared in a condition (view === 'grid') is skipped,
and each token carries its own line in a multi-line list.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review (the user, 2026-09-15): an application without Tailwind
names a role through its --md-sys-color-* custom property (CSS, an inline
style, a script), an md-ink-* class, or a component's color/tone prop, not
a bg-tertiary utility. forbidColours() keeps its name and signature and
reports each of those, the role's on- and container roles included;
a Tailwind leftover stays family (i)'s single report. ReStride's
forbidColours(['tertiary', 'primary-container']) keeps its meaning.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review (the user, 2026-09-15): without Tailwind, bg-white and
text-on-surface/60 compile to nothing like any other utility, so family (i)
now reports every colour utility on an M3 role, a 1.x ink name
(text-meta, border-divider), white, black or current, with or without an
opacity, each with its 2.0.0 replacement (an md-ink-* class, <x-divider>,
<x-surface level>, md-state-layer or the role's var()). Check (iii) reads
named colours anywhere in a declaration's value (border: 1px solid white),
hex codes only in values (an id like #add no longer matches), and masks
strings and url() contents first.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 review (the user, 2026-09-15): SealShare and ReStride have
both left maryUI and daisyUI, and no application of 2.0.0 carries either,
so the <x-mary-*> tag check, the daisyUI class and colour lists and the
Tailwind look-alike exclusions they needed go, with their fixture lines
and the mentions in the header, README and development skill.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41(iii): fixtures/tests for every literal kind (colour, radius,
shadow, font size/weight, line height, letter spacing, easing, duration) and
an off-scale media query, plus material-scheme.css excluded by name and by
its generated header wherever it sits among the scanned paths. Also adds a
fixture proving Tailwind-shaped bare words in running prose are never read
as classes.
Two fixes found by writing the fixtures rather than only the implementation:
a directly-scanned .css file was also running through the Blade/PHP/JS checks
(the same file's `ease-in-out` in a `transition` value tripped the old scale
check a second time), so a CSS file is now check (iii)'s alone; and
`withoutTokenFunctions()` leaves a `var()`/`calc()` call's own name and
parentheses standing once its arguments are blanked, which
`isSafeLiteralValue()` did not yet recognise as the same "nothing to see
here" as an empty value.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41(ii): exercises the tag-to-stylesheet mapping (plain, prefixed and
namespaced spellings), a package stylesheet's own imports counting toward its
dependents (split-button.css satisfies button.css and menu.css too),
->links() needing pagination.css, the always-required foundation.css, and a
package tag shadowed by the application's own anonymous or class component.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41(i): every Tailwind utility or variant in a Tailwind-free
application's views now gets a 2.0.0 hint instead of the old Tailwind-utility
replacement (which no longer compiles either) -- a layout component and prop,
an md-* class, or a var(--md-sys-*) token. New families cover flex/grid/gap,
spacing, sizing, display, text-layout and the M3 role utilities
(text-on-surface-variant, border-outline-variant, bg-surface-*, bg-primary);
the breakpoint and scale checks fold in with their hints rewritten the same
way. A class the application's own stylesheets declare is exempt.
literalClasses() now returns each class token unstripped of its variant
prefix (`sm:`, `hover:`) so the new checks can see it; the daisyUI check
strips it itself via the new withoutVariantPrefix() helper, as it always did
internally before.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 41 (ii): the design guard needs the set of files an application's
CSS entry pulls in, transitively, to check whether a package tag's
stylesheet is among them — reusing bundle()'s own import resolution and
cache rather than re-implementing @import parsing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 40. ErrorPage::fallbackStyles() now inlines
Stylesheets::bundle() of the foundation and error-page.css (which
imports button.css and shape.css) rather than a hand-copied
stylesheet, so the fallback can never drift from the built version.
Every @font-face block is dropped structurally (withoutFontFace(),
brace-balanced, not a text search) since there is no build to serve
the font file; --md-ref-typeface-brand already lists ui-sans-serif,
system-ui and sans-serif after the brand name, so the page still gets
a sensible system stack. The scheme half comes from
Scheme::forStylesheet() drawn through SchemeStylesheet::levels(), in
material-scheme.css's own selector shape: standard, the medium and
high contrast levels, and a block per colour profile keyed on
[data-scheme] (the theme script has already resolved and written the
active one to <html> before this stylesheet is read, so nothing here
picks one in PHP). Both halves are cached per worker, the scheme half
by the scheme file's path and mtime.
Tests: both render paths, the inlined CSS's shape (no @import, no
relative url(), no @font-face, the button/shape/error-page rules, the
scheme's roles, [data-contrast='high'] and a profile block), and a new
browser test that forces the fallback and checks the button, the
shape and a dark-mode repaint.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 39 (part 5). Tailwind no longer clears its own breakpoint
scale anywhere in the stack, so the guard's rationale changes from "a
leftover sm: compiles to nothing" to "a browser drops the invalid
class or wrong-width rule just as silently" — the scans themselves
already worked over plain files and needed no change.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 39 (part 4). With Tailwind gone the Workbench's Vite build
still turns every `:dir(rtl)` into a long `:lang()` list (its own CSS
minifier, not Tailwind's doing — confirmed by rebuilding and grepping
the output). Every RTL mirror rule across the eleven stylesheets that
had one now selects `:is([dir='rtl'], [dir='rtl'] *)` instead, matching
`:dir(rtl)`'s own specificity (one pseudo-class) and its inherited
"this element or a descendant of one carrying the attribute" reach; the
built CSS now keeps the selector as written (no `:lang(` or `:dir(` left).
RTL browser tests drop the `lang="ar"` workaround the old rewrite
needed (ContainmentTest's side-sheet probe, LayoutTest's shared
layoutPage() helper) and set only dir="rtl"; CarouselTest's probe
already did. Feature tests asserting the selector's literal text
(Progress, Overlay, Menu, ListDetail, Icon, Carousel) updated to match.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 39 (parts 1-3, folded into one commit: PHP tests read the
deleted files directly, so they cannot land apart from it). Tailwind
leaves the whole stack:
- Delete resources/css/tailwind.css, tokens/theme.css and
tokens/utilities.css. Neither token file declared an --md-sys-*
custom property of its own (both only referenced tokens declared
elsewhere), so nothing loses a value; the md-* interaction and text
classes already mirror utilities.css's declarations exactly
(foundation/interaction.css, text.css).
- npm uninstall tailwindcss @tailwindcss/vite; vite.config.js drops the
plugin and its import; composer.json drops the tailwindcss keyword
(no lock change — keywords are outside Composer's content hash).
- The Workbench now builds one CSS entry, workbench/resources/css/app.css
(all.css, showcase.css and the scheme; package.css is folded in and
removed) instead of two, used by ErrorPage::assets() and every
browser-test probe page's raw @vite() call; the showcase's own pages
still take their CSS from the bundle route.
- TokensTest and StylesheetsTest: the two facts theme.css and
utilities.css carried (every scheme role becomes a colour, resolved
on the element; md-type-* matches the type-* utilities) are asserted
directly against the scheme and text.css now that there is no second
copy to cross-check; StylesheetsTest gained a full-tree scan (every
.css file under resources/css/ is reached from all.css or
showcase.css, no exclusions left for Tailwind); the Workbench-entry
test and every "moved out of tailwind.css" assertion updated for the
single entry and its removal.
- DesignGuard.php's comments and the development skill's setup section
no longer name the deleted files or a second Tailwind entry.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 39 (part 0): the step 38 review found <x-scaffold> and
<x-tooltip> never referencing $attributes, so a caller's class and
style were silently dropped. The scaffold's root is its single
data-md-scaffold div; the tooltip's root, documented in its header, is
the standalone wrapper it draws around a trigger (data-md-tooltip-anchor,
merged with its own anchor-name) — passed an anchor instead, it renders
only a popover fragment beside another component's root and takes
nothing of the caller's.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 38 review: ShowcaseController::layout() passes a title
("List-detail · Layout"), but the frame overwrote it with the section's
own, so all three pages and the Layout overview shared one title. The
frame keeps a title it was given; the layout page test asserts each.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 38 review: the component sections' code samples, which an
application copies, carried 60 data-md-showcase-* hooks that only
showcase.css draws (a sized loading indicator, the plain badge's own
colours, a bordered bar frame, the navigation demos' frames, a table's
scroll wrapper...), so a copied sample did nothing in the application.
They now write what an application has: <x-surface outlined corner>
for a frame and a line, <x-row> for the FAB menu's corner, and an
inline style from the tokens for a size, a scroll box or a colour,
as the Containment batch already did. The side sheet demo takes
align="stretch", which holds in both of <x-row stack-below>'s modes,
instead of a hook that undid align="start" once stacked. The menu
icon colours go through icon-class with two unlayered application-like
classes, which the sample's comment names, and the fields' and date
picker's widths are inline styles now that style reaches their roots.
The radio hint said "from sm"; it is medium.
showcase.css keeps only the frame and the foundation specimens; its
header describes the file instead of the batches that wrote it. The
tests reject a showcase hook or a class outside md-* and the named
application classes in any example, a showcase.css selector outside
its own hooks, and a hook no showcase view renders.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 38 review: the showcase kept two unlayered width classes
because <x-input> and <x-datepicker> "don't forward style usefully".
They did not: the Phase F rule is that a caller's class and style land
on the component root, and the field family sent class to the field's
root but style to the inner control, where a width sizes nothing.
Input, password, textarea, select and file now pass style to the field
root with class; datepicker and timepicker put it on their root;
<x-group> dropped style entirely and <x-split-button> gave it to the
leading button; <x-search> rendered it twice, on the root and the
input. A textarea given a style also lost its autogrow rows, since
its own style attribute came second and the browser ignores it.
A dataset test renders each component with both and requires them on
the root, once.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
Plan step 38 review: nine sections' examples wrapped their components
in <x-livewire-material::stack>, ::row and ::grid (148 tags) beside an
unprefixed <x-button>, while bars, containment and the layout pages
wrote <x-stack>. The code sample is what an application copies, and
the example component rewrites only unprefixed tags to a configured
prefix, so the namespaced ones taught a second form and escaped it.
Every example now writes <x-stack>, <x-row> and <x-grid>; a test
rejects a namespaced tag inside an example heredoc.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qwx5USif3wFFmxtHg5U1g9
