Skip to content

feat(ui): one glass dock for the composer band - #14568

Merged
CrysisDeu merged 2 commits into
mainfrom
feat/glass-followup-dock
Sep 28, 2026
Merged

CrysisDeu merged 2 commits into
mainfrom
feat/glass-followup-dock

Conversation

@CrysisDeu

@CrysisDeu CrysisDeu commented Sep 27, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Everything that floats over the transcript in the band above the input box — the follow-up chips, the tip and folder-suggestion cards, the question card, the memory-mode chip, the jump-to-bottom button, the sub-agent approval card and the tool approval bar — wears the same Liquid Glass material the composer got in #14460, the recipe lives in one place, and the transcript scrolls under that dock the way content scrolls under an iOS toolbar.

After #14460 the composer is a glass pane, but every other control in the same band is still a solid bg-bg-elevated box with its own border. The band reads as one glass object sitting among a scatter of unrelated cards, and a tool approval bar fused to the composer's top meets the pane at a visible seam (solid box above, glass below). The recipe (tint, edge, light band, halo) was also restated per call site.

And the glass had nothing to refract: the transcript scroller ended where the composer began, with an opaque fade band between them, so the only thing ever behind the pane was the page background. Measured on the 40-message transcript at the bottom: 0 text nodes under the glass.

Why it matters

The composer band is the one region every user looks at every turn. Half glass and half card looks unfinished, and the seam at the approval bar shows up exactly when the user is being asked to decide something. A glass dock over a blank strip is also a material with no job: the blur, tint and edge only read as glass once real content moves beneath them, which is why the layout change ships with the material rather than after it.

Not a goal

  • Glass on the queue cards (QueueStack). They overlap in a peek stack, and a translucent card shows the cards behind it, which is why they were hard-coded opaque before this PR. They stay opaque.
  • Glass on ErrorNotice. It is the standard error component and is left as is.
  • Refraction. The edge bend stays at the depth feat(ui): liquid glass composer and settings capsule, colored tiles #14460 landed; only tint, frost and the light bands are re-tuned (see "The recipe, re-tuned" below).
  • A fade band between transcript and dock. Variant C (glass plus a bottom fade) was mocked and rejected in favour of A: the content shows through, and the material carries legibility on its own.
  • The app-SDK ChatPane / split-grid composers. They keep their own layout; this PR is the main chat page only.

What changed (motivation → approach → change)

One material, one place. The recipe moves into components/Glass.tsx, a thin polymorphic wrapper over the ui/liquid-glass primitive: the pane IS the host element, there is no wrapper box. The pane draws no ring — its top and bottom are the light bands, its left and right a 1px --glass-edge line, nothing around the corners, as in the reference material (two lit edges set against thin dark sides) — so the host carries only the caller's shadow. Call sites say what they are (variant: panel for composer-sized boxes at frost 6 / light 25, chip for the small pills and cards at frost 4 / light 18), how round they are (radius) and which element they ARE (as: div, span or button — a follow-up chip is <Glass as="button" onClick …>, so the button is the flex item, carries the width cap, the entrance animation and its own handler, and the effect layers sit inside it under the label at z-index: -1 in the host's own isolated stacking context; a <button> host stays valid phrasing content because the layers are <span>s). className on the host carries layout plus the box-shadow state the caller owns: glass-shadow (the neutral rest shadow; its focus-within form deepens the shadow and swaps the tint for --glass-tint-focus, a brighter, more opaque step of the same hue, and the side lines for --glass-edge-focus, the same ink at about twice the alpha, instead of colouring), composer-halo (the session composer's own accent focus glow) or approval-glow while a decision is pending. Which shadow a pane wears at this instant is the caller's state, not the material's, so it is not a prop, and the accent glow is the composer's alone — the material never lights up in the theme colour, so the mobile Settings capsule now focuses by deepening its shadow, lighting its tint a step and darkening its two side lines, not by turning accent (measured on the built capsule, whole pane incl. shadow: +4.6 in light, +23.4 in dark after the focus tint was stepped up to dark rgba(84,84,92,.72) / light white .92 — the first .60 / .72 step read as no change in a still, and the tint step alone still went unread beside unchanged edges, so the side lines now step with it: light black .24 → .60, dark white .14 → .55; measured on the built capsule at mid-height, 2×: the side line goes 65 → 174 in dark while the interior goes 27 → 68, and 195 → 128 in light while the interior goes 240 → 254 — the largest pixel change between the rest and focused frames is now 110 in dark and 85 in light, against 42 and 68 on the previous head). Under the three solidifying fallbacks none of those layers paint, so there the same .glass-shadow:focus-within rule draws the app's standard 2px accent outline instead — a solid box holding focus must show it. The optics are not open at the call site. The composer and the capsule both render through it; SidePanelLayout.tsx drops its hand-built hairline + halo box.

One dock pane. The composer's approval bar, the notice row, the composer and the collapsed bar now share ONE pane (composer-dock). The input-wrapper stays transparent in every state — the previous solid fallback while an approval was fused above it is gone, since there is no longer a seam to hide — and the dock swaps its resting halo for the approval glow while a decision is pending, so the whole pane says "waiting on you" rather than one bar inside it. The pane stays mounted across the transition, so an approval never remounts the editor. The sub-agent approval card (spawn) is a separate Glass pane in the plain material, on purpose: the first round's UX read saw a warn-tinted spawn card and the warn tool-approval bar in one band as ONE request ("I'd fear double-approving"), so the card keeps its Bot framing and pulse and drops the shared warn tint, and the two asks stop looking alike. The third round's blind read still hesitated over the two panes ("is this a second, separate question or the same one … one click only answers half of it"), so while a tool approval is pending the spawn card withholds its own Approve/Reject and glow and keeps only its count and "Review in panel", and its line reads "N sub-agents pending — answer the request below first" (new plural key components.chatInput.spawn_pending, all catalogs) rather than "awaiting your approval", so the line points at the surface that holds the decision instead of at itself. The blind read of the previous head then tripped on the vocabulary — the line sent it to a bar that says "Allow once / Reject" while the card's live state says "Approve / Reject" for the same act, and it could not tell whether those are one decision — so the line now names the ORDER ("answer … first") rather than a synonym for the buttons, and the card's live-state sibling line, which was an English literal on main, is the plural key components.chatInput.spawn_awaiting in every catalog. The per-agent rows collapse out (AnimatePresence, height + opacity) when a tool approval lands, the same way the action area cross-fades, instead of unmounting on one frame under a fading header. One set of decision buttons on screen at a time; the dock's glow marks the one that is live. A spawn approval frame also mints the bar's own permission row (useWebSocket approval → sseChatMessage), so the bar's decision on a spawn IS the spawn decision, and the card takes the buttons back only once no permission row is pending after the last user message — in practice when the user sends a message while the spawn still waits (a new user turn retires the stale rows). The action area's three forms (resolving / panel link / Approve + Reject) cross-fade under AnimatePresence mode="wait" instead of swapping, so the slot reads as one control changing state.

The primitive on every surface, no CSS twin. The first push gave the small surfaces a CSS copy of the material (.glass-pane): same tint and edges, no refraction, because at chip size the bend measured as zero pixels and the primitive's extra DOM would have broken the chip tree that the entrance animation and FollowUpBar.test.tsx address by structure. The copy drifted from the primitive on every re-tune (its top light was a hard 1px line with no bevel, and the capsule and the chips visibly read as two materials), so it is gone: FollowUpBar (each chip, and the split wrapper when instant-send is on), TipCard, FolderSuggestionCard, QuestionCard, MemoryModeChip and the jump-to-bottom button in app-sdk/ChatScrollChrome.tsx render Glass itself in the chip recipe, hosted AS the button or card element (as), so the chip tree is unchanged — one element is the control, the flex item and the pane — and every pane carries the primitive's bands, bevel, side lines, hairlines and refraction map (about 120×30 device px per chip; the map is rasterised once per settled size). The border border-transparent the chips kept so their box does not move by a pixel stays; nothing is drawn. Picked, pending and hover states mix their colour INTO the tint on the host instead of carrying a second recipe: glass-accent (a picked chip, an attached card), glass-warn (the incognito memory chip) and glass-hover:hover each swap --glass-tint for a step derived once on :root (--glass-tint-accent = 14% --accent into the tint, --glass-tint-warn = 12% --warn, --glass-tint-hover = 8% --text into the tint, a step toward the foreground in either polarity). The three existing glass fallbacks (reduced transparency, more contrast, no backdrop-filter) address .liquid-glass alone again; under all three the -accent / -warn hosts re-assert their mixed fill on the solid --bg-elevated (the !important fill would otherwise outrank the tint swap), more-contrast draws a 1px outline in --text (an outline, so no layout shift), with -accent / -warn outlining in their own colour so a picked chip and an incognito chip still read apart, and .liquid-glass:focus-visible re-asserts the 2px accent focus ring inside the same block. The layer hide addresses the primitive's own data-liquid-glass-layer attribute (every effect layer and the filter's zero-size <svg> carry it), NOT aria-hidden: the children render directly, so a decorative icon among them — TipCard's <Lightbulb aria-hidden>, FolderSuggestionCard's <FolderInput> — is a direct child too, and an aria-hidden selector deleted it with the layers under every fallback. It covers the frost clip wrapper, so the far-edge blur fix below solidifies with the rest. Measured on the built chip's top edge in dark (device px, 2×): the CSS copy [9, 43, 43, 78, 75, 71, 70, 67, 66, 44, 43] — a hairline, a 1px gap, a 3px flat-topped band, then a hard step; the primitive [9, 68, 61, 52, 47, 44, 43, 43, 43] — the hairline meets the lit edge and it fades inward smoothly, the same profile as the composer's [9, 66, 63, 59, 55, 51, 48, 43, 39].

The recipe, re-tuned (index.css, Glass.tsx, ui/liquid-glass.tsx). With content scrolling under the pane the #14460 values read too opaque and too flat against the reference material, so: --glass-tint dark goes rgba(30,30,34,.55) → .40; light goes white .72 → a cool light grey rgba(238,238,243,.45) — not white, because on a white page a white fill leaves the top/bottom light bands nothing to stand out from (measured: pane interior 228 vs page 254, band +11 over the interior; with the white tint the band was +6). Frost drops 24 → 6 on the composer and 18 → 4 on the chips: the text scrolling under the pane has to stay readable as text, and at 16, 12 and again at 8 it still read as mush. The light bands take their colour from a new polarity-fixed --glass-band (dark rgba(255,255,255,.22), light rgba(255,255,255,.92)) scaled by the light setting (25 = the full token, the chip's 18 ≈ 70%): on the light grey tint the old white-alpha peak topped out at ~22% white and read grey, so the light-mode band is now near-white. The bands fade to CLEAR past the corner arc, and the sides carry a 1px --glass-edge line (light .16 → .24 black; dark white .14), drawn on the bevel layer as two horizontal inset shadows: a horizontal offset paints nothing on the top and bottom edges and thins out through the corner arcs by itself, so the line never closes into the ring #14460 drew around the pane. Each light band sits against a half-pixel --glass-hairline (light black .20, dark black .50) just outside the top and bottom edges: two outer box-shadows offset ±0.5px, which show only along the edge they moved away from and taper through the arcs — a very thin dark line under each lit edge, as the reference's, one device pixel on a 2× display. Measured over white in light: page 248 → hairline row 198 → band 251 → interior 246. That is the reference material — two lit edges on thin dark hairlines, set against thin dark sides; measured on an iOS 27 capsule over white, the side line reads ~185/255 at mid-height, ~215 through the arc, interior ~248. Measured on the built dock at rest: top edge +37 over the interior in dark (65.9 vs 29.1), +8 in light (251.6 vs 243.2, on a 251 page); side line 186 vs interior 246 in light (218 at the start of the arc → 186 on the straight flank), 58 vs 25 in dark; no line on the top or bottom edge.

The blur reaches the far edges (ui/liquid-glass.tsx). Measured on the built component over 8px horizontal stripes: the bottom 17px of the 24px-blurred pane showed the stripes at a third of their contrast while the top edge was flat, and the same over vertical stripes on the right edge but not the left. A bare div with backdrop-filter: blur(24px) did the same, at radius 0 too, so this is Chromium's backdrop blur under-blurring the far edges of the element that carries it, not this composition. The frost layer is therefore rendered two blur radii larger than the pane on every side inside a clipping box, so the under-blurred band lies outside what is shown. After: top and bottom rows of the pane mirror each other (dark 92→80 / 80→93, the rim light), stripe swing at the bottom edge 47 → 9, the same as the top. Reduced-transparency / more-contrast / no-backdrop-filter still hide the layer, since the clip box carries data-liquid-glass-layer like the layers before it. Because every small surface is now the primitive too, the chips and cards get the same clipped overhang and no surface keeps the under-blurred band.

The transcript scrolls under the dock (ChatPage.tsx). The dock — status stack, jump-to-bottom pill, FlyingQuote, the composer with everything above it — leaves the pane's flex column and becomes one absolute inset-x-0 bottom-0 box (composer-dock-root). The transcript scroller therefore fills the pane to its bottom edge, and pays for the covered strip with paddingBottom: dockH + DOCK_CLEARANCE_PX (16px), where dockH is the dock's live offsetHeight: a callback ref attaches a ResizeObserver to the dock root — a callback ref, not a [] layout effect, because the dock sits inside the pane's conditional branch and a mount-once effect ran before it existed and never measured. The ref fires in the commit phase, so the first painted frame already carries the right padding. The welcome hero pads by the same dockH so it stays centred in the visible strip. Growth and shrink of the dock (chips appearing, an approval bar, a taller draft) change the scroller's scrollHeight through its padding, which the existing follow controller already treats as content settling: at the bottom it re-pins, above the bottom it leaves the reader alone — the scroller was already observed for exactly that. The memory-chip row loses its bg-bg backdrop, since the content scrolling under it is now the point.

The opaque bottom mask (TRANSCRIPT_MASK_ABOVE_PX / COMPOSER_MASK_OVERSHOOT_PX, the from-bg from-[62%] gradient) is removed outright: it existed to feather the transcript into an opaque composer box, and over glass it would paint a solid strip across exactly the content the layout exists to show. The status bars' z-[2] layers stay (they now state the dock's paint order rather than clear a mask, and their comments say so); QueueStack stays below the composer's z-10 so its fuse still slides under the input box. The dock root has NO z-index on purpose — positioned with z-index: auto it forms no stacking context, so SubagentProgressBar's z-[46] still clears the theme-experience overlays — and paints over the scroller because it follows it in DOM order. It is pointer-events-none; its two wrapper boxes (the status stack and the input area) are dock-inert — inert themselves, with input restored on their CHILDREN, each of which is its own 900px column — so wheel and touch in the empty width beside the column reach the transcript beneath, as the strip beside an iOS toolbar does. The status stack is also a scroll box (max-h-[50svh] overflow-y-auto), and a scrollbar belongs to the box, not to a child, so an inert stack could not be scrolled by its own thumb once it overflowed: the band's existing ResizeObserver (useChatPageTranscriptController) marks it data-overflowing — it crosses the cap only by changing height, so every crossing is a resize it already sees — and .dock-inert[data-overflowing="true"] lets the box catch input as a whole for exactly that stretch. The root is also inset on the right by the scroller's measured scrollbar-gutter (6px in Chromium, read as offsetWidth - clientWidth in the same ResizeObserver that measures the dock height, which also observes the scroller so an engine with overlay scrollbars keeps the inset correct), which both keeps the scrollbar thumb uncovered down to the pane's bottom edge and lines the dock's column up with the transcript's, which is centred in the content box that excludes the gutter.

scripts/capture-liquid-glass.mjs grows from four composer scenes to fifteen — the collapsed composer bar, follow-up chips at rest and picked, a pending tool approval, the sub-agent approval card LIVE (the spawn request, then a user message, so the bar is gone and the card carries Approve / Reject and the glow), both approvals pending together (asserts the card says "1 sub-agent pending", offers no Approve and does not glow while the tool bar is up), the tip card, the folder-suggestion card, the question card, incognito on an empty session (asserts the warn memory chip), the welcome state with the memory chip, the 40-message transcript both pinned to the bottom and scrolled up, and a spawn-flow scene that records the approval band's three transitions to a GIF (tool bar + withheld card → the tool decision lands and the bar shows the spawn's own request, card still withheld → the user speaks, the bar hides and the card's buttons fade in), asserting each state before pushing the next frame, trimmed to open on the settled themed band rather than the blank load — in both polarities, and asserts the floating-dock geometry on the long scene (scroller bottom == dock bottom, padding == dock height + 16, last line clears the dock, text sits under the glass once scrolled, the jump-to-bottom pill is up). Cards are raised by pushing websocket frames into the stubbed socket, 900ms apart when a scene needs several.

Backwards compatibility

Compatible: no route, config key or API change; two i18n additions (components.chatInput.spawn_pending and spawn_awaiting, both plural, in all 13 catalogs + the pseudolocale — i18n/locales/*.json, plural shape in i18n/pluralKeys.json); the ui/liquid-glass primitive gains one optional prop, as (div default, span, button), and its className / style now land on the pane's host (the element carrying liquid-glass) rather than on a wrapper, so a consumer that addressed .liquid-glass's parent for layout would now address the pane itself; no consumer outside this PR does. Glass is new; .glass-pane and its -accent / -warn / -hover variants exist only between the first push of this PR and this head and are removed. Two polarity-fixed CSS variables change shape: --glass-band is new, and --glass-edge is no longer drawn around any pane — it is the SIDE line only (left and right flanks, thinning through the corners) and the in-pane divider, so every glass surface loses the top and bottom of its 1px rim and its light-mode value darkens .16 → .24; the mobile Settings capsule no longer turns accent on focus (its shadow deepens, its tint and side lines step instead; under the solidifying fallbacks it takes the standard accent outline). Three things a user can see differ: the composer no longer turns solid while an approval bar is attached (that solid state existed only to hide the seam this PR removes); the last line of the transcript no longer fades into the composer — it stops 16px clear of the glass at the bottom and shows through it once scrolled; and the empty width either side of the composer column now scrolls the transcript instead of swallowing the wheel. One shadow rule is new in index.css: .glass-shadow, the neutral offset-free rest shadow for every glass pane that is not the session composer, whose focus-within form deepens the shadow and swaps the tint for --glass-tint-focus and the side lines for --glass-edge-focus; the bands and side lines live on the primitive's own layers, so they never compete with the host's shadow for one box-shadow. Dark themes keep the 1px accent ring under the composer's .composer-halo focus glow.

Tests

  • ChatInput.liquidGlass.test.tsx (rewritten for the dock structure): the wrapper is bg-transparent with no bg-bg-elevated in the no-approval state; with a pending approval the wrapper stays on the shared pane (still transparent) and the dock carries the approval glow instead of the resting halo; exactly one dock pane is mounted around the wrapper and the dock element IS the primitive's host (carries liquid-glass, no ring class, composer-halo, 16px glass, --glass-tint layer); the accent focus glow exists only as .composer-halo (no .liquid-glass … .composer-halo rule; .glass-shadow:focus-within is the neutral deepen and swaps the tint); index.css contains no glass-pane; the state steps are --glass-tint swaps derived on :root (glass-accent, glass-warn, glass-hover:hover); and every one of the three fallback rules solidifies .liquid-glass, re-asserts the -accent / -warn fills, hides the [data-liquid-glass-layer] layers (never an [aria-hidden] selector) and gives a focused .glass-shadow pane the 2px accent outline.
  • FollowUpBar.test.tsx: a chip button carries liquid-glass (it IS the pane) and a picked one glass-accent, an unpicked one not; the chip tree, entrance animation and click behaviour assertions are unchanged.
  • ChatPage.dockClearance.test.tsx (new, replaces ChatPage.fadeClearance.test.tsx): the scroller's padding is dockH + DOCK_CLEARANCE_PX, never a constant; the dock root is absolute bottom-0 and is measured from a callback ref through a ResizeObserver (and NOT from a [] layout effect); the clearance and tail spacer are px, never viewport units; the welcome hero pads by dockH; no fade gradient and no mask constants remain; the memory-chip row has no bg-bg.
  • ChatPage.statusStackLayering.test.tsx (new, replaces ChatPage.statusStackAboveMask.test.tsx): keeps the child-list guard, QueueStack below the composer, and the stack wrapper context-free; adds that the dock root is positioned with no z-index (so it forms no stacking context) and that the root is pointer-events-none, both wrappers are dock-inert (never pointer-events-auto themselves) and index.css defines .dock-inert as inert-with-children-live.
  • ChatPage.scrollShell.recipe.test.tsx / .render.test.tsx / ChatPage.queueBandReanchor.test.tsx: the pinned paddingBottom literal follows the new expression, and the mask recipe pin becomes a DOM-order pin (the dock root follows the scroller) plus a no-gradient assertion.
  • ChatInput.approval.test.tsx: with a tool approval and two spawns pending together, the spawn card reads "2 sub-agents pending — answer the request below first" (and nothing says "awaiting your approval"), shows "Review in panel", offers no Approve all / per-agent Approve, carries no glow, and the dock carries the glow.
  • LiquidGlass.test.tsx: the composer fixture follows the recipe (frost 6, light 25) — blur(6px), a 12px frost overhang; the bevel layer's box-shadow leads with the two --glass-edge side insets (inset 1px 0 0, inset -1px 0 0), then the two outer --glass-hairline shadows (0 ±0.5px 0 0), before the vertical bevel insets; the specular layer's gradient is read from its --liquid-glass-bands custom property (jsdom drops a color-mix() background) and starts at the full --glass-band.
  • ChatInput.liquidGlass.test.tsx also pins the glass tokens per polarity (tint, focus tint, band, edge, focus edge, hairline), the focus rule (.glass-shadow:focus-within:not(.glass-accent, .glass-warn) swaps the tint and the edge) and the -accent / -warn re-assertion under the solidifying fallbacks; MemoryModeChip.test.tsx expects the chip button to carry liquid-glass + glass-warn + text-warn for the incognito chip (was border-warn); es/fr/it/pt gain spawn_pending_many, which their plural rules select; LiquidGlass.test.tsx checks every layer and only the layers carry data-liquid-glass-layer (an aria-hidden icon child does not); ChatPage.statusStackLayering.test.tsx pins the data-overflowing rule and the observer write.
  • scripts/mutation-check-scroll-shell.mjs: the bottom-mask region is re-anchored on the dock root ({/* Composer dock → <JumpToBottomButton) and the suite list names ChatPage.dockClearance.test.tsx, so npm run mutation:scroll-shell walks the new shell instead of exiting on a missing anchor.
  • ChatPage.composerChromeOcclusion.test.tsx is unchanged and still passes on this head: the status-stack cap, the title-band lift and the drawer scrim ordering are untouched.

LiquidGlass.test.tsx also pins the polymorphic host: <LiquidGlass as="button" …> renders a <button> that carries liquid-glass plus the caller's className, forwards ref, type, onClick, aria-* and style to that button, and holds the effect layers inside it. No new test for Glass.tsx itself: it is a thin composition over the primitive (recipe table + cornerRadius from radius).

Manual verification

node scripts/capture-liquid-glass.mjs against the built dist of this head, dark and light: composer at rest / focused / collapsed / chips / picked chip / tool approval / sub-agent approval / both approvals together / tip / folder suggestion / question card / incognito / welcome, the 40-message transcript at the bottom and scrolled up, the mobile Settings capsule at rest and focused, the desktop Settings rail. The harness asserts a blur() backdrop layer renders on the composer (url(#lg-…) + blur(8px) saturate(1.55)) and the capsule, that the capsule's box-shadow changes on focus and stays neutral (dark rgba(0,0,0,.30) → .55, light .06 → .14; no accent) and its --glass-edge steps to the focus value (dark #ffffff24 → #ffffff8c, light #0000003d → #0009), that the live spawn card offers Approve / Reject with the glow and no bar, that the withheld card says "1 sub-agent pending — see the bar below" with no Approve and no glow, that the incognito scene renders the warn memory chip, and on the spawn-flow recording each of the three states before the next frame is pushed, and on the long scene: scroller bottom == dock bottom (the transcript runs under the dock), scroller padding == dock height (143px) + 16, the last message clears the dock by 70px at the bottom, 3 text nodes sit under the glass once scrolled up 180px, the jump-to-bottom pill is present, the dock's right edge stops exactly the reserved scrollbar gutter (6px) short of the scroller's, elementFromPoint in the 91px gutter left of the column is not a dock box, and a wheel delivered there scrolls the transcript 120px. Static gates run locally on this head: tsc -b --force, eslint (0 warnings), i18n:check. Tests run in CI.

Screenshots / video

All frames from this head's build (node scripts/capture-liquid-glass.mjs), dark and light.

The transcript under the dock — 40 messages, scrolled up so a paragraph and a code block sit beneath the glass, jump-to-bottom pill above it:

transcript scrolling under the dock, dark
transcript scrolling under the dock, light

Same transcript pinned to the bottom — the last line stops clear of the dock, nothing hides under it:

transcript at the bottom, dark
transcript at the bottom, light

Composer at rest — no ring: light bands top and bottom, a 1px dark line down each side that thins through the corners, the shadow for depth:

composer at rest, dark
composer at rest, light

Composer with follow-up chips — chips and pane are one material, same bands, same side lines:

composer with follow-up chips, dark
composer with follow-up chips, light

A picked chip (glass-accent):

picked follow-up chip, dark
picked follow-up chip, light

Composer with a pending tool approval fused above it — one pane, the approval glow on the whole dock, no seam:

composer with pending approval, dark
composer with pending approval, light

The sub-agent approval card live — the user sent a message while the spawn waited, so the bar is gone and the card alone carries Approve / Reject and the glow:

sub-agent approval card, live, dark
sub-agent approval card, live, light

Both approvals pending at once — the bar acts; the card reads "1 sub-agent pending — answer the request below first" with the panel link only, no buttons, no glow:

both approvals pending, dark
both approvals pending, light

The approval band's transitions, recorded: tool bar + withheld card → the tool decision lands (the bar now shows the spawn's own request; the card stays withheld) → the user speaks (the bar hides; the card's Approve / Reject fade in):

approval band transitions, dark
approval band transitions, light

Tip card, folder-suggestion card and question card on glass, over the dense transcript:

tip card, dark
tip card, light
folder suggestion card, dark
folder suggestion card, light
question card, dark
question card, light

Memory-mode chip: default state on the empty-session welcome screen, then the warn chip in incognito (same glass, the warn hue mixed into the tint) above the warn-bordered composer:

welcome state with memory chip, dark
welcome state with memory chip, light
incognito: warn memory chip and composer, dark
incognito: warn memory chip and composer, light

The collapsed composer bar — the whole dock is the "Show the message input" button, over the transcript:

collapsed composer bar, dark
collapsed composer bar, light

Mobile Settings capsule (390px) at rest and focused — focus deepens the shadow, lights the tint a step and darkens the side lines (neutral, no accent). Measured on these frames, interior / side line, 0–255: dark 27 → 68 / 62 → 168; light 246 → 255 / 199 → 170, page under the shadow 249 → 239. The light-mode step is deliberately quiet; a keyboard-only :focus-visible accent ring is an open maintainer question, not in this PR:

settings capsule at rest, dark
settings capsule focused, dark
settings capsule at rest, light
settings capsule focused, light

Related Issues

Follows #14460.

Pattern harvest

  • A material that several components share belongs in one wrapper that takes ROLE props (variant, radius) and never optics, and leaves per-instant STATE (which shadow, focus) to the caller's className rather than growing a boolean per consumer. A CSS copy of a component's material is a second recipe that drifts on every re-tune; when the component's DOM is the obstacle, make the component polymorphic on its host (as) rather than copying its output.
  • When a pending state used to be hidden by making a surface solid, ask whether the seam it hid still exists once the surfaces share one pane.
  • A translucent toolbar over a scroller wants the scroller to run underneath it and pay for the overlap with padding measured from the toolbar, not a fixed constant and not a fade band; measure with a callback ref when the toolbar mounts conditionally, because a [] effect only looks once.

Checklist

  • Two logical changes, two commits: the material, then the layout
  • Owning doc updated in the same change: website/docs/theming-contract.md records --glass-band, --glass-edge and --glass-hairline beside --glass-tint and that the pane draws no ring (lit top and bottom, a line down each side); no spec covers the composer band's layout
  • No new markdown files
  • Static gates green locally (tsc, eslint, i18n:check); tests run in CI
  • Screenshots attached for every user-visible change
  • No model id, no internal tool name, no committed evidence file

@CrysisDeu
CrysisDeu requested a review from a team September 27, 2026 23:56
@CrysisDeu
CrysisDeu requested a review from a team as a code owner September 27, 2026 23:56
@CrysisDeu

Copy link
Copy Markdown
Contributor Author

Intent. Every surface in the composer band (follow-up chips, tip / folder-suggestion / question cards, memory chip, jump-to-bottom button, sub-agent approval card, tool approval bar) wears the Liquid Glass material #14460 gave the composer, and the recipe lives in one wrapper (components/Glass.tsx) plus one CSS form (.glass-pane) for the small, DOM-pinned surfaces. The approval bar and the composer share ONE pane, so the seam between them is gone and the solid fallback that hid it is removed.

Not in scope. Transcript scrolling under the dock (ChatPage layout, follows this PR); glass on the peek-stacked queue cards (they would show through each other); ErrorNotice; any change to the optics landed in #14460.

Reviewers: a CONCERN that asks to make chips use the component form should weigh the measured zero-pixel bend at chip size and the pinned chip tree in FollowUpBar.test.tsx; a CONCERN about the composer no longer going solid under an approval should check that the seam it hid no longer exists (one pane).

@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Sep 27, 2026
@github-actions

github-actions Bot commented Sep 27, 2026 •

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of a120137398eff2abbad62fcbf6abb9320fb2bd3a and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] a120137

False positive or not applicable? A repository writer can comment:
/ai-review override gpt a120137398eff2abbad62fcbf6abb9320fb2bd3a: <one-sentence reason>

@github-actions

github-actions Bot commented Sep 28, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Fable 5) — ✅ PASS

Design-level review of a120137398eff2abbad62fcbf6abb9320fb2bd3a — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Verify before merge: that minting the tool-bar's permission row from a spawn frame routes a spawn decision to the correct spawn and cannot double-approve when both asks are live — a screenshot can't confirm approval routing.

[DESIGN-REVIEWED] a120137

@github-actions

github-actions Bot commented Sep 28, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Fable 5) — 🟡 CONCERNS

Premise-level review of a120137398eff2abbad62fcbf6abb9320fb2bd3a — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

First-Principles-Verdict: CONCERNS

Item 8 rides along beyond the frozen Goal: the spawn/tool approval arbitration (button-withholding + new copy + new plural i18n keys in every catalog) is an interaction-behavior change, not "wear the same material."

Not justified as shipped

  • Item 8 — rides along (non-trivial surface): the Goal is "same Liquid Glass material … transcript scrolls under the dock"; dropping the spawn card's warn tint to plain glass already removes the named "two warn panes read as ONE request" harm (ChatInput.tsx comment, "I'd fear double-approving"). The further step — withholding Approve/Reject, the spawn_pending/spawn_awaiting copy, and new plural keys added to ~13 locale catalogs plus pluralKeys.json — is a separate interaction rule from later UX rounds, riding a material-unification PR.
    Subtraction: keep the plain-glass spawn card (which meets the Goal); defer the button-withholding + spawn_pending/spawn_awaiting keys and the AnimatePresence action-swap to their own PR.
    Clears when: that logic is split out, or a maintainer confirms the arbitration change belongs in this Goal.

What this change ships

Inventory (10 items) — 9 justified

Intent: Make every control in the composer band wear one shared Liquid Glass material and let the transcript scroll under the dock like an iOS toolbar. ADDITION / refactor.

  1. Follow-up chips now render as glass instead of solid bg-bg-elevated boxes — justified
  2. Tip, folder-suggestion and question cards now glass — justified
  3. Memory-mode chip now glass — justified
  4. Jump-to-bottom button now glass — justified
  5. Tool approval bar and composer share one dock pane, removing the seam — justified
  6. Transcript scrolls under the dock; opaque fade band + bottom-mask geometry (and its two pinning suites) removed, replaced by measured dock clearance + layering suites — justified
  7. components/Glass.tsx holds the recipe once; per-call-site frost/light/edge restatement removed — justified
  8. Sub-agent card withholds decision while a tool approval is pending, with new copy and new spawn_pending/spawn_awaiting plural keys across all catalogs — rides along (non-trivial surface)
  9. Settings capsule focus cue is now neutral (shadow + tint + edge step) rather than the accent glow — justified
  10. New glass CSS tokens (--glass-tint-focus, --glass-band, --glass-edge-focus, --glass-hairline) — justified

Deleted pins (fadeClearance, statusStackAboveMask) are honestly replaced by dockClearance/statusStackLayering, which carry the same px-not-vh and no-stacking-context invariants forward — not a reversed decision.

[FIRST-PRINCIPLES-REVIEWED] a120137

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Sep 28, 2026
@github-actions

github-actions Bot commented Sep 28, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Fable 5) — 🟡 CONCERNS

UX-level review of a120137398eff2abbad62fcbf6abb9320fb2bd3a — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

UX-Verdict: CONCERNS

Settings capsule's new neutral focus cue is invisible: the blind reader called the rest and focused frames pixel-identical in both themes.

Watch

  • Capsule focus unreadable. This PR swaps the capsule's accent hairline + halo for .glass-shadow:focus-within tint/edge steps, and the input suppresses its own outline (focus:outline-hidden). The reader: "Shots 31 vs 32, and 33 vs 34, look pixel-identical to me." A keyboard user on mobile Settings loses focus every time — persistent, moderate impact. The dark composer got a 1px accent ring for exactly this ("a reader could not tell the two apart"); give the capsule the same ring the fallback blocks already draw.
  • The new spawn_pending line points at a code-speaking strip with a different verb. "answer the request below first" sends the user to a bar reading [agent] spawn_run(review the dock… with "Allow once", while the card says "Approve" for the same act. Reader: "looks like computer code and I'm not certain what I'd be allowing" / "possibly the same action, possibly not." Hits every spawn-while-tool-pending decision. Fix: render the spawn row as a task name ("Start sub-agent: …") and use one verb across card and bar.
  • Jump-to-bottom pill reads as tucked under the composer (shot-01/02; reader: "half hidden behind the typing box… looks odd") — glass-on-glass at ~8px clearance. Raise its offset above the dock pane.

Suggestions

  • Pad the settings list bottom by the capsule height (the transcript's dockH + clearance pattern) so Privacy/Security can scroll clear; reader: "I can't tell if those covered rows would still be tappable."

[UX-REVIEWED] a120137

@github-actions

github-actions Bot commented Sep 28, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5 Review — ✅ no blocking findings

Reviewed a120137398eff2abbad62fcbf6abb9320fb2bd3a — this comment is updated in place on each push.

Review details

Advisory only: the new .glass-hover cue is the only hover feedback left on the reskinned controls, and none of the three solidifying fallback blocks reinstates it.

FINDING — website/src/index.css:2946 — .glass-hover:hover { --glass-tint: var(--glass-tint-hover) } swaps a custom property read only by the [data-liquid-glass-layer] frost span, which @supports not ((backdrop-filter: blur(1px)) or (-webkit-backdrop-filter: blur(1px))), prefers-reduced-transparency: reduce and prefers-contrast: more each set display:none !important on, while the jump-to-bottom pill's rewritten class list (ChatScrollChrome.tsx:54) dropped hover:bg-bg-hover hover:border-accent for a bare glass-hover — so in all three modes that control, the unpicked follow-up chip and the memory chip have zero hover feedback, unlike .glass-accent/.glass-warn which those blocks do re-assert → Fix: add a .glass-hover:hover{ background:color-mix(in srgb, var(--text) 8%, var(--bg-elevated)) !important } line to each of the three blocks, beside the .glass-accent / .glass-warn lines already there.

[OPUS-REVIEWED] a120137

Verdict parsed from the review's SHA-scoped output markers for commit a120137398eff2abbad62fcbf6abb9320fb2bd3a.

False positive or not applicable? A repository writer can comment:
/ai-review override fable a120137398eff2abbad62fcbf6abb9320fb2bd3a: <one-sentence reason>

@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Sep 28, 2026
@CrysisDeu

Copy link
Copy Markdown
Contributor Author

Scope update on 76851c9068: the transcript-under-dock layout, listed under Not a goal on the first head, is now the second commit of this PR (the material and the layout are one user-visible change: glass over a blank strip is a material with no job). The PR body is rewritten for the combined scope, and the UX Review evidence gaps on b4f30113b3 (tip / folder / question cards, memory chip in both states, sub-agent approval card, jump-to-bottom pill, picked chip, the dense transcript under the glass) are covered by the 28 frames now attached, all captured from this head's build.

For the CI lanes: Fast Gate on b4f30113b3 sat queued for 35+ minutes (runner starvation), which is why Await Fast Gate timed out and Coverage Gate failed closed with frontend-test=skipped — a runner state, not a PR state. This push starts a fresh round.

@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Sep 28, 2026
@CrysisDeu
CrysisDeu force-pushed the feat/glass-followup-dock branch from 76851c9 to 4317666 Compare September 28, 2026 01:46
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Sep 28, 2026
@CrysisDeu
CrysisDeu force-pushed the feat/glass-followup-dock branch from 4317666 to 21b8e79 Compare September 28, 2026 02:44
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Sep 28, 2026
@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • span=b3edd2a2295d dark-mode .composer-halo overrides .glass-pane inset bands: fixed

Fixed in 21b8e79: .glass-pane now puts its inset bands in --glass-bands and .glass-pane.composer-halo (rest, dark rest, focus, dark focus) composes var(--glass-bands) with the halo shadow in one box-shadow, so an element carrying both classes paints both; TipCard / FolderSuggestionCard / QuestionCard are the consumers.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • span=083e7a6856df full-width pointer-events-auto wrappers consume wheel in empty dock gutters: fixed

Fixed in 21b8e79: the status stack and the input area are dock-inert (pointer-events:none, > * {pointer-events:auto} in index.css); every child is its own 900px column, so the width beside the column hit-tests to the transcript. Verified on the built dist by the harness: elementFromPoint in the 91px gutter is not a dock box and a wheel there scrolls the transcript 120px. The memory-chip row became w-fit for the same reason.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • span=6fc9b933adef dock children re-arm hit-testing full-width and cover the scrollbarGutter: stable column: fixed

Fixed in 21b8e79: the dock root is left-0 bottom-0 with style={{ right: dockGutter }}, where dockGutter = scroller.offsetWidth - scroller.clientWidth is read in the same ResizeObserver callback that measures the dock height, so the scrollbar column stays uncovered and the dock's column lines up with the transcript's (centred in the content box that excludes the gutter). Harness asserts dock right == scroller right - 6px. Wrappers are dock-inert (see the GPT disposition for the same file).

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • span=26ba1dddb132 .glass-pane-warn border loses to .glass-pane's shorthand; .glass-pane box-shadow collides with .composer-halo: fixed

Fixed in 21b8e79: .glass-pane-warn now sets border-color (warn 50%) like -accent does and MemoryModeChip drops the dead border-warn utility; the shadow collision is closed by the .glass-pane.composer-halo combined rules (--glass-bands + halo) for rest / dark / focus / dark focus, so TipCard / QuestionCard / FolderSuggestionCard paint bands and halo in both polarities.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • span=b12632617dc8 Item 7 — edgeClassName + GLASS_EDGE_CLASS have zero consumers: fixed

Fixed in 21b8e79: both deleted from Glass.tsx; the hairline class is a constant inside the component and the doc no longer claims a picked-chip consumer.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • declined span=0bd6ffa1e598 -- Item 7: the call the lane asks a maintainer to make was made, by the product owner, for this PR -- the theme-colour focus highlight belongs to the session text input only, no other Liquid Glass surface lights up in the theme colour. That is a product decision, not a symmetry taste, and it is why the two focus tokens exist (the capsule still owes a visible cue; a neutral one needs a tint and a side-line step to swap in). Kept.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • acknowledged span=de90caa8ba69 -- Measured on the frames the reader compared (fddfcb8 build, 2x, 0-255): dark capsule interior 27 -> 68 and side line 62 -> 168 between rest and focus; light interior 246 -> 255, side line 199 -> 170, page under the shadow 249 -> 239 -- 134k / 180k pixels differ, so the pair is not identical, but the light step is quiet by design (no accent, no ring, per the owner's rule). The body now states these numbers under shots 31-34 (a120137398ef) and names the open maintainer question -- a keyboard-only :focus-visible accent ring -- rather than claiming the cue reads on the phone. Not tuned further in this PR.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • fixed span=50a0f8925135 -- Fixed in a120137398ef: the pointer line no longer paraphrases the buttons. components.chatInput.spawn_pending reads 'N sub-agents pending -- answer the request below first' (en + 11 catalogs + pseudolocale), naming the ORDER of the two panes instead of a synonym for either pane's verb, so a reader is not asked to decide whether 'Allow once' and 'Approve' are one act. Harness and ChatInput.approval.test.tsx match the new copy; the both-pending shots and both GIFs are re-captured from this head.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • acknowledged span=9529905d0215 -- The show-through is the material's intent for this PR (the product owner's direction: the transcript scrolls under the dock and reads through the glass, as on the reference platform), and the ghosted text is not interactive -- the dock is a pointer-events:none layer over the scroller except on its controls, so a tap on ghosted text scrolls the transcript. A human pass on dense real transcripts is fair and is what the owner is doing on the built branch; the frost was lowered three times on that feedback (24 -> 6 blur) and the tint has hairlines and side lines to hold the edge. No change in this head.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • fixed span=0acbe3cd9d69 -- Fixed in a120137398ef: the live-state sibling line ('N sub-agents are awaiting your approval to run', an English literal on main) is the plural key components.chatInput.spawn_awaiting in all 13 catalogs + the pseudolocale, registered in pluralKeys.json; the card shows a translated string in both states.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • fixed span=c28bcb35f910 -- Fixed in a120137398ef: the per-agent rows sit in their own AnimatePresence (initial={false}) as a motion.div that collapses height + opacity over 150ms, so when a tool approval lands the card shrinks to its one-line form while the action area cross-fades, instead of the rows unmounting on one frame.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • fixed span=5ea2deba6226 -- Fixed in a120137398ef: the three regexes in ChatInput.liquidGlass.test.tsx now read the blocks as this PR writes them -- --glass-edge-focus between --glass-edge and --glass-hairline in both polarity blocks, and the .glass-shadow:focus-within rule with its --glass-edge: var(--glass-edge-focus) line. This was the Frontend Tests (1) red on fddfcb8 (the anchors predate the previous push's token insert); Frontend Tests (4) was the sibling drift in ChatPage.queueBandReanchor.test.tsx (observer regex missing the data-overflowing line), fixed the same way.

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • fixed span=d5541e3ef7af -- Fixed in a120137398ef: both literals are gone; the line is i18nT('components.chatInput.spawn_awaiting', { count }), plural, in every catalog (spawn_awaiting_one/_other, plus _few/_many where the locale has them), registered in pluralKeys.json. i18n:check passes (19 checks) on this head.

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Sep 28, 2026
@CrysisDeu
CrysisDeu enabled auto-merge (squash) September 28, 2026 16:28
@CrysisDeu
CrysisDeu merged commit 5d1f095 into main Sep 28, 2026
138 of 141 checks passed
@CrysisDeu
CrysisDeu deleted the feat/glass-followup-dock branch September 28, 2026 16:32
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Sep 28, 2026
CrysisDeu added a commit that referenced this pull request Sep 28, 2026
…allback

Three follow-ups on the glass dock that shipped in #14568.

The composer's focus cue is the glass itself, no theme color. The dock
wore `composer-halo` (an accent glow, plus a 1px accent ring in dark
themes) and the input wrapper turned its border `accent/50` on focus;
every other glass pane wore the neutral `glass-shadow`. Now the dock
wears `glass-shadow` too and the wrapper's border stays transparent:
focus is the material getting a step clearer (`--glass-tint-focus` and
`--glass-edge-focus`), the same cue as the Settings capsule and the
cards. `composer-halo` and its memory-mode forms are gone.
While a tool approval is pending the dock stacks `approval-glow` on
`glass-shadow`: the glow takes the shadow slot, the focus step stays.

The follow-up chips lost their top and bottom hairline: the chip row is
`overflow-x-auto`, which also clips vertically, and the 0.5px line is
drawn just outside the box. The row gets `py-px -my-px` so the line
paints at the same row height.

Hover feedback on glass buttons survives the three solidifying
fallbacks: `.glass-hover` steps `--glass-tint`, which only the hidden
frost layer reads, so each fallback block moves the step onto the solid
fill (the contrast block at twice the mix). The jump-to-bottom pill floats 24px clear of the dock instead of
8px, where glass on glass read as one pane bleeding into the other.
CrysisDeu added a commit that referenced this pull request Sep 28, 2026
…allback

Three follow-ups on the glass dock that shipped in #14568.

The composer's focus cue is the glass itself, no theme color. The dock
wore `composer-halo` (an accent glow, plus a 1px accent ring in dark
themes) and the input wrapper turned its border `accent/50` on focus;
every other glass pane wore the neutral `glass-shadow`. Now the dock
wears `glass-shadow` too and the wrapper's border stays transparent:
focus is the material getting a step clearer (`--glass-tint-focus` and
`--glass-edge-focus`), the same cue as the Settings capsule and the
cards. `composer-halo` and its memory-mode forms are gone.
While a tool approval is pending the dock stacks `approval-glow` on
`glass-shadow`: the glow takes the shadow slot, the focus step stays.

The follow-up chips lost their top and bottom hairline: the chip row is
`overflow-x-auto`, which also clips vertically, and the 0.5px line is
drawn just outside the box. The row gets `py-px -my-px` so the line
paints at the same row height.

Hover feedback on glass buttons survives the three solidifying
fallbacks: `.glass-hover` steps `--glass-tint`, which only the hidden
frost layer reads, so each fallback block moves the step onto the solid
fill (the contrast block at twice the mix). The jump-to-bottom pill floats 24px clear of the dock instead of
8px, where glass on glass read as one pane bleeding into the other.
CrysisDeu added a commit that referenced this pull request Sep 28, 2026
…allback

Three follow-ups on the glass dock that shipped in #14568.

The composer's focus cue is the glass itself, no theme color. The dock
wore `composer-halo` (an accent glow, plus a 1px accent ring in dark
themes) and the input wrapper turned its border `accent/50` on focus;
every other glass pane wore the neutral `glass-shadow`. Now the dock
wears `glass-shadow` too and the wrapper's border stays transparent:
focus is the material getting a step clearer (`--glass-tint-focus` and
`--glass-edge-focus`), the same cue as the Settings capsule and the
cards. `composer-halo` and its memory-mode forms are gone.
While a tool approval is pending the dock stacks `approval-glow` on
`glass-shadow`: the glow takes the shadow slot, the focus step stays.

The follow-up chips lost their top and bottom hairline: the chip row is
`overflow-x-auto`, which also clips vertically, and the 0.5px line is
drawn just outside the box. The row gets `py-px -my-px` so the line
paints at the same row height.

Hover feedback on glass buttons survives the three solidifying
fallbacks: `.glass-hover` steps `--glass-tint`, which only the hidden
frost layer reads, so each fallback block moves the step onto the solid
fill (the contrast block at twice the mix). The jump-to-bottom pill floats 24px clear of the dock instead of
8px, where glass on glass read as one pane bleeding into the other.
CrysisDeu added a commit that referenced this pull request Sep 28, 2026
…allback

Three follow-ups on the glass dock that shipped in #14568.

The composer's focus cue is the glass itself, no theme color. The dock
wore `composer-halo` (an accent glow, plus a 1px accent ring in dark
themes) and the input wrapper turned its border `accent/50` on focus;
every other glass pane wore the neutral `glass-shadow`. Now the dock
wears `glass-shadow` too and the wrapper's border stays transparent:
focus is the material getting a step clearer (`--glass-tint-focus` and
`--glass-edge-focus`), the same cue as the Settings capsule and the
cards. `composer-halo` and its memory-mode forms are gone.
While a tool approval is pending the dock stacks `approval-glow` on
`glass-shadow`: the glow takes the shadow slot, the focus step stays.

The follow-up chips lost their top and bottom hairline: the chip row is
`overflow-x-auto`, which also clips vertically, and the 0.5px line is
drawn just outside the box. The row gets `py-px -my-px` so the line
paints at the same row height.

Hover feedback on glass buttons survives the three solidifying
fallbacks: `.glass-hover` steps `--glass-tint`, which only the hidden
frost layer reads, so each fallback block moves the step onto the solid
fill (the contrast block at twice the mix). The jump-to-bottom pill floats 24px clear of the dock instead of
8px, where glass on glass read as one pane bleeding into the other.
CrysisDeu added a commit that referenced this pull request Sep 29, 2026
…allback

Three follow-ups on the glass dock that shipped in #14568.

The composer's focus cue is the glass itself, no theme color. The dock
wore `composer-halo` (an accent glow, plus a 1px accent ring in dark
themes) and the input wrapper turned its border `accent/50` on focus;
every other glass pane wore the neutral `glass-shadow`. Now the dock
wears `glass-shadow` too and the wrapper's border stays transparent:
focus is the material getting a step clearer (`--glass-tint-focus` and
`--glass-edge-focus`), the same cue as the Settings capsule and the
cards. `composer-halo` and its memory-mode forms are gone.
While a tool approval is pending the dock stacks `approval-glow` on
`glass-shadow`: the glow takes the shadow slot, the focus step stays.

The follow-up chips lost their top and bottom hairline: the chip row is
`overflow-x-auto`, which also clips vertically, and the 0.5px line is
drawn just outside the box. The row gets `py-px -my-px` so the line
paints at the same row height.

Hover feedback on glass buttons survives the three solidifying
fallbacks: `.glass-hover` steps `--glass-tint`, which only the hidden
frost layer reads, so each fallback block moves the step onto the solid
fill (the contrast block at twice the mix). The jump-to-bottom pill floats 24px clear of the dock instead of
8px, where glass on glass read as one pane bleeding into the other.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants