Repository navigation
Conversation
Design Review (Fable 5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS The shelf scrim's geometry hangs on an implicit containing-block chain across ChatPage/ChatInput, asserted only by a script CI never runs. Watch
[DESIGN-REVIEWED] 7fdc2b9 Description read[DESCRIPTION-READ] c6f303698bd4d8ad6b3e4aa662da4c382b6b1cba3ad195e16928508a9812b856 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
First Principles Review (Fable 5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Every measured claim (tint composite, blur smear, bevel profile) is asserted only by What this change shipsInventory (10 items) — 10 justifiedIntent: make the composer read as one piece of the shared glass material — neutral focus, matching edges, working fallbacks — and retune the dark glass against the reference. This is a FIX.
List capped at 10: the 11th item is the new capture-harness scenes ( Evidence notes: the deleted dark-ring pin ("the glow alone did not separate a focused capsule from a resting one") is reversed with declared, measured replacement evidence (edge step; side line 143→213, interior 56→68), so lens 8 is satisfied. Grepped [FIRST-PRINCIPLES-REVIEWED] 7fdc2b9 Description read[DESCRIPTION-READ] bce0e4d243142054fe1b5f6dfced7980e689286cca9451af2bba872c70cf9454 That sha256 names the PR description this verdict read. Recompute it with |
Opus 5 Review — ✅ no blocking findingsReviewed Review detailsDark-polarity composer focus indicator drops below 3:1 with nothing else in its place. FINDING — website/src/components/ChatInput.tsx:4274 — the wrapper's focus cue is now unconditional [OPUS-REVIEWED] 7fdc2b9 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
UX Review (Fable 5) — 🔴 BLOCKUX-level review of UX-Verdict: BLOCK cannot evaluate: missing shots of the repainted overflow edge cues (file strip / control row) and the reduced-motion steady approval ring. Watch
Evidence gaps
[UX-REVIEWED] 7fdc2b9 Description read[DESCRIPTION-READ] e7ef8846308b4dc23b21f49b22449fabc30af82f80220769e7ebcf2b90f8815f That sha256 names everything this verdict read: the PR description, plus the 22 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
01d9cb6 to
54e5538
Compare
|
|
|
|
|
|
|
54e5538 to
9661330
Compare
|
|
|
9661330 to
c35e7ce
Compare
|
|
c35e7ce to
bbbde82
Compare
|
|
|
|
|
|
|
|
…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.
bbbde82 to
7fdc2b9
Compare
|
|
|
|
Problem / Motivation
Goal: The chat composer reads as one piece of Liquid Glass: focus is the glass getting clearer, nothing turns the theme colour, and every glass pane shows the same edges.
Follow-ups on the glass dock that shipped in #14568, and one re-tune of the material itself against the reference:
composer-halo: an--accent-glowshadow on focus plus a 1px accent ring in dark themes, and the input wrapper turned its borderaccent/50. Every other glass pane (Settings capsule, cards, chips) already used the neutralglass-shadow.@supports not backdrop-filter,prefers-reduced-transparency,prefers-contrast: more) a hovered chip did not change at all, and the jump-to-bottom pill sat 8px above the dock, glass on glass.--text, not the white the reference types in.Why it matters
The composer is the most-focused element in the product. A theme-coloured glow there, and none anywhere else, made the glass look like two materials. A missing hairline on chips is the same tell in miniature. A hover that does nothing in the reduced-transparency modes is a lost affordance for exactly the users who asked for a plainer surface.
Not a goal
Glass/LiquidGlass: the retune is token and recipe values.glass-shadow.approval-glow): a pending decision still lights the dock warm. Underprefers-reduced-motionit now holds a steady ring instead of no ring (see below).What changed (motivation → approach → change)
The composer's focus is the glass, not a theme color
Symptom: focusing the composer painted an accent halo and an accent border. Cause: the dock carried
composer-haloand the wrapperfocus-within:border-accent/50, both predating the shared material. Change: the dock wearsglass-shadowlike every other pane, and the wrapper's border stays transparent in the default memory mode (incognito / temporary keep their warn / aim border, which is a mode signal, not a focus cue). Focus is the material getting a step clearer:--glass-tintsteps to--glass-tint-focusand the side lines to--glass-edge-focus, the same cue as the Settings capsule and the cards.composer-haloand its-warn/-aimforms are deleted fromindex.css.While a tool approval is pending the dock is
glass-shadow approval-glow: the glow takes the box-shadow slot (--approval-shadow, animated as before) and the tint + edge focus step still applies, so the textarea keeps a visible focus cue in that state too, in every theme and in the three fallbacks (their.glass-shadow:focus-withinoutline now matches).Measured on this head's build (harness asserts it on every run):
rgba(0,0,0,.30) 0 0 18pxrgba(0,0,0,.55) 0 0 18pxrgba(127,134,147,.30)->rgba(127,134,147,.42)rgba(0,0,0,.06) 0 0 18pxrgba(0,0,0,.14) 0 0 18px#eeeef373->#ffffffebComposited interior through the input: dark
#383742->#444451(luminance 56 -> 68), light 247 -> 255; dark side line 143 -> 213. With an approval pending: shadow stays the warm glow, tint and edge step the same way.The dark material is the reference's colour, not a darkening
Symptom: in dark mode the dock read as a flat grey slab beside the reference, and the blur hid what scrolled under it. Cause:
--glass-tintwasrgba(30,30,34,.40), a black fill, so the glass interior could only land below the page; the side line was 14% white (+20 over the page, the reference's rim is about +100); frost was 6 / 4. Change (tokens and the two recipe numbers only):--glass-tintrgba(30,30,34,.40)(black)rgba(127,134,147,.30): the reference's composited colour#37393Dsolved back into a translucent fill over the app's own dark page (#181818): 24 + 0.30 x (127 - 24) = 55. Measured on the build:#383742at rest. (Solving it over pure black instead -- a 24% tint ofrgb(222,234,255)-- matched the reference on black but composited to#6A6B79on the real page while focused, far too bright.)--glass-tint-focusrgba(84,84,92,.72)rgba(127,134,147,.42):#444451measured--glass-edge/-focusGlassrecipe)liquid-glass.tsx)0.2 x lightalpha, 1x reachBEVEL_GLOW = 0.48,BEVEL_REACH = 2.5: the reference's rim carries a white glow fading inward over ~5pt (+25-35 over the page); ours had almost none. Dark top edge inward, 2x pixels:66 64 60 56 52 49 44->113 112 109 107 104 101 98. This layer paints in light mode too, where white on a near-white pane is invisible: every light edge profile is within 1/255 of the previous head.--text/--muted--glass-text/--glass-placeholder:#ffffff/#9c9c9con smoked glass, the theme's--text/--mutedon a light pane (white would vanish)Light-mode tint, bands and hairline are untouched. The tokens stay polarity-fixed (
website/docs/theming-contract.md). The retune applies to everyGlasspane, the mobile Settings capsule included (frames below); the capsule has no shelf under it, so the fade is the composer's alone.The context shelf stands on a short fade to page colour
Symptom: the chips under the pane (agent, project, branch, model) read against transcript text scrolling under the floating dock. Change: the shelf is
glass-shelf, a::beforethat fades from nothing at the pane's bottom edge to solid--bgby 60% of the shelf's height. It is positioned against the dock'srelative z-10 dock-inertwrapper (the shelf itself is deliberately not positioned), so it spans exactly the dock root -- which already stops short of the scroller's reserved scrollbar column, so the thumb stays uncovered -- and itsz-index: -1resolves in that wrapper's stacking context, BEHIND the pane and its.glass-shadow: the shadow falls onto the solid strip the way it fell onto the page. (A first cut painted the fade over the shadow, which whitened the shadow out part-way down and left a bright band under the pane that read as a reflection.) ChatInput hands it the shelf's border-box height as--glass-shelf-hso it starts at the pane's edge. Nothing paints inside the pane: the glass above the shelf is untouched (the harness asserts the pane's own::beforecarries no gradient). Measured with a message body scrolled under the whole dock (2x pixels): text under the glass peaks at 152 over the 108 interior; under the shelf it is 37 at the fade's top and gone (page 24) by its lower half; light 255 throughout.Secondary ink and the strip edge cues follow the material
Cause (Opus, round 3): the retuned dark tint composites to about
rgb(67,71,81)over the default page, where the theme's--muted#7f7f88keeps barely 2:1 -- half of what it had on the near-black fill. Change:.liquid-glass { --muted: var(--glass-muted) }, a polarity-fixed#b4b4bcon smoked glass and the theme's own--mutedon a light pane, so everytext-mutedinside a pane (chip labels, the memory chip, card hints, the control row) brightens with the glass. Measured on the follow-up chips, dark: muted ink 129 -> 154 (max 181); light frames byte-identical.Cause (Opus, round 3): the file-preview strip's and the control row's overflow cues faded
from-bg-elevated, an opaque colour the glass no longer has, so at narrow widths they would have painted a dark wedge on the pane. Change: the cues are backdrop-blur wedges (backdrop-blur-[3px], masked to nothing toward the strip), which soften the clipped chip instead of covering it; same elements, sameuseScrollEdgesgate, same tests. Where the pane solidifies (the three fallback blocks, including@supports not backdrop-filter, where a blur wedge paints nothing)glass-strip-cue-l/-rswap the wedge for a fade from--bg-elevated, the pane's own colour there (Design, round 3).A pending approval keeps its glow under reduced motion
Cause (Opus, round 2): the global reduced-motion rule runs every animation once for 0.01ms, which parks
approval-glowat its 0% keyframe, where--approval-shadowis fully transparent; the dock that hands its shadow slot to that variable then showed no shadow and no glow. Change:@media (prefers-reduced-motion: reduce) { .approval-glow { animation: none; --glow-strength: .6 } }, a steady mid-pulse ring.Chips get their top and bottom hairline back
Symptom: no hairline on chips. Cause: the chip row is
overflow-x-auto; CSS makes the other axisautotoo, and the 0.5px--glass-hairlineis drawn just outside the box, so the row clipped it top and bottom. Change: the row getspy-px -my-px; the line paints, the row height does not change.Chip edge profile, outer -> inner (2x pixels): dark top
24 -> 12 -> 57 47 38(was24 -> 67 60 56, no line); light top255 -> 204 -> 250 249 247.Hover survives the solidifying fallbacks; jump pill floats clear
Cause:
.glass-hoversteps--glass-tint, which only the frost layer reads, and all three fallback blocks hide that layer. Change: each block moves the hover step onto the solid fill (.liquid-glass.glass-hover:hover), beside theglass-accent/glass-warnlines already there; theprefers-contrast: moreblock uses twice the mix (16%), so the users who asked for more contrast get the clearest hover, not the faintest. The jump-to-bottom pill floats 24px clear of the dock (-top-14); the layout comments inChatScrollChrome.tsxandChatPage.tsxname the new figure.Backwards compatibility
Compatible: CSS classes, tokens and two recipe numbers;
composer-halo*had no consumer outsideChatInput.tsx.--glass-text,--glass-placeholderand--glass-mutedare new polarity-fixed tokens, not theme roles;.liquid-glassscopes--mutedto the pane. No API, config or stored data changes.Tests
ChatInput.liquidGlass.test.tsx: the dock wearsglass-shadowat rest andglass-shadow approval-glowwhile a decision is pending; the wrapper never carriesfocus-within:border-accent;index.csshas nocomposer-halorule, the neutral.glass-shadowrules carry no theme colour, and the three fallback blocks each carry.liquid-glass.glass-hover:hoverand the.glass-shadow:focus-withinoutline.ChatInput.liquidGlass.test.tsxalso pins the new token values for both polarities, theglass-shelffade rule (containing block = the dock wrapper, no viewport-wide insets, no stacking context of its own), its class and--glass-shelf-hon the shelf (and that no in-pane scrim exists), the fallback rules that turn the strip cues back into--bg-elevatedfades, the scoped--glass-muted, the blur cues (nofrom-bg-elevatedleft in the file), and the reduced-motionapproval-glowrule;LiquidGlass.test.tsxpins the frost 3 recipe.Manual verification
tsc,eslint(touched files),i18n-check,check_focus_cue.py(diff-scoped, baseorigin/main): green.capture-liquid-glass.mjson the built SPA: 80 frames (this head,dock-r22). The long scene now also scrolls a message body under the whole dock and asserts the shelf fade is present, starts at the pane's edge and does not touch the pane. It asserts the focus deltas above (rest vs focused, with and without a pending approval), the chip hairline and the pill clearance, and fails on any unexpected dialog. Its newcontrastscene emulatesprefers-contrast: moreand asserts the fallback itself: the frost layers aredisplay: none, a hovered chip changes its solid fill (darkrgb(33,29,37)→rgb(48,44,52), light255→240), and the focused dock wears the2px solidaccent outline. No local test run; CI runs the suites.Screenshots / video
Rest / focus, dark and light (the retuned dark material: reference colour, brighter side line, frost 3):
Approval pending, focused: the glow stays, the glass still steps clearer.
Chips with their hairline:
Solidifying fallback (
prefers-contrast: more, emulated): every pane is a flat card, the hovered chip (first chip, "用 A") takes the 16% text mix, and the focused composer gets the standard accent outline. The other two fallback blocks paint the same frame with the 8% hover mix.Long transcript, pill floating clear of the dock:
Mobile Settings search capsule on the retuned material, REST (left) and FOCUSED (right) in one frame per polarity. The focus cue on glass is deliberately the material itself, never a theme colour or a ring (maintainer decision): the tint steps clearer and the side lines step darker/brighter, the shadow deepens. Measured: dark interior 57 -> 69 with the side line 14% -> 55% white; light interior 247 -> 255 with the side line 24% -> 60% black and the shadow .06 -> .14 (the edge under the capsule 249 -> 239). The caret is the third cue.
Message body scrolled under the whole dock: text shows through the glass, the shelf under it sits on the fade to page colour:
Related Issues
Follow-up to #14568.
Pattern harvest
Rule candidate: review-prompt
Pattern: a conditional className that swaps one shadow class for another drops every other cue the first class carried (here the focus step); prefer stacking (
base state) over swapping.Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)