Repository navigation
feat(ui): glass search field floating over the session and crew lists - #14896
Conversation
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Verify the [DESIGN-REVIEWED] 8be2dc3 Description read[DESCRIPTION-READ] ebc526c8769f71b171e8c8e65d0f250934f61b27b237458bec2ca1afa26f86a6 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) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The restyle's only checkable support is symmetry with the composer; the maintainer's design-session request exists nowhere but this description. Not justified as shipped
What this change shipsInventory (10 items) — 9 justifiedIntent: make the two list search fields wear the composer's Liquid Glass and let the rows scroll under them — an ADDITION (a restyle; behavior, test ids and names preserved).
Counted: [FIRST-PRINCIPLES-REVIEWED] 8be2dc3 Description read[DESCRIPTION-READ] e5418546ec843f80da278fe3c305e9155fd66b111934671990cc95d57d53dcbe That sha256 names the PR description this verdict read. Recompute it with |
UX Review (Fable 5) — ✅ PASSUX-level review of UX-Verdict: PASS Typing/focused state was captured only in dark (shot-03); verify the opaque-tint focus cue and clear button stay legible on kiro-light. [UX-REVIEWED] 8be2dc3 Description read[DESCRIPTION-READ] f4791954451ca146f1aaa3f993ddfc05da3c9266c8c2916f1ffd0f36755edff3 That sha256 names everything this verdict read: the PR description, plus the 18 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 |
Opus 5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
aadfcb7 to
d84c17d
Compare
d84c17d to
0b3da02
Compare
0b3da02 to
d93dda4
Compare
a345b9f to
778ce98
Compare
778ce98 to
bebaacc
Compare
The Sessions sidebar and the Crew Members roster share one search row (SearchFilterBar). It was a solid bordered input sitting above the list; it is now a Liquid Glass pane (the composer's material, chip recipe, 10px radius) that floats over the list inside a new ListDock, and the rows scroll under it the way the transcript scrolls under the composer dock. At rest nothing is under the glass: ListDock measures its own height (search field, filter chips, list-level notices) with a ResizeObserver and writes it to --list-dock-h, and LIST_BODY_CLS pads every list body's top by that much, so the first row starts 4px below the pane and only slides under it once the user scrolls. The dock wrapper is click-through between its children so the row beside a short chip row stays reachable. The input inside the pane is transparent and carries no ring of its own: .glass-shadow:focus-within (the composer's and the Settings capsule's cue) steps the tint opaque and darkens the side lines on focus. Known limit: in the board (columns) lane the column strip pads its top by the dock height instead, so column headers start below the glass and do not scroll under it.
bebaacc to
8be2dc3
Compare
…#14995) The dock shipped in #14896 ended 4px under the pane: the search row kept its pb-1, and a bare field's shelf still painted a 4px scrim fade. A row scrolling under the dock therefore emerged through a strip of a third tone between the blurred glass and the sharp list, and that strip read as a leftover box behind the pane. The pane's bottom edge is now the dock's edge. The search row drops its bottom padding, the shelf hides when empty (the roster's star-error wrapper mounts only while there is an error, so it does not hold the shelf open), and the 4px rest gap between dock and first row moves into the measured variable: ListDock publishes --list-dock-h as its height plus the gap and --list-dock-pin-inset as the gap, so the lane pads and scroll-pads to dock + 4 while a pinned folder header still sits exactly on the dock's bottom edge. The keyboard-reveal band subtracts the inset too, so it ends where the header stack ends instead of 4px past it.
Problem / Motivation
Goal: The search field at the top of the Sessions sidebar and of the Crew Members roster wears the same Liquid Glass material as the composer, and the list scrolls under it the way the transcript scrolls under the composer dock.
This is the maintainer's own request from the design session that produced it: with the composer band and the mobile Settings capsule on glass since #14460 / #14568, he asked for the two list search fields to join, and for the list to scroll under them like the transcript under the composer — with the first row clear of the field at rest. Since #14460 / #14568 the composer band and the mobile Settings search capsule are glass; the two list panels still open with a solid, bordered input sitting on a strip of panel background above the rows. The two search fields are the one surface a user sees on every visit that has not joined the material.
Why it matters
The dashboard's chrome now reads as two generations side by side: a glass composer at the bottom of the chat and a flat bordered field at the top of the list next to it. Aligning the list panels closes that gap on the two surfaces every user looks at first.
Not a goal
Glass.tsx,ui/liquid-glass.tsx, the--glass-*tokens).index.cssalready address every.liquid-glasshost.What changed (motivation → approach → change)
One field, one component. Both lists already render their search row through
components/SearchFilterBar.tsx(pinned bysearchFilterBar.parity.test.tsx), so the change to the field is made once. The row'sSearchInputis replaced by a<Glass variant="chip" radius={10}>pane hosting the lucideSearchglyph, a transparent<input>, the clear button and the trailing controls at their existing offsets (the right-padding / clear-offset table is unchanged). The input carries no ring of its own (focus-cue-ok, saying where the cue lives):.glass-shadow:focus-within— the cue the composer and the Settings capsule already give — steps the tint opaque and darkens the side lines on focus, and under the three solidifying fallbacks the same rule draws the 2px accent outline.Rounded, not a capsule. Both were built and photographed on the real app. The composer's 16px turns a 32px field into a pill; the rows under it are
rounded-mdboxes, and the pill read as a foreign shape among rectangles. 10px keeps the corner family of the rows while the material stays the composer's.The list scrolls under it. A new
components/ListDock.tsxwraps each list: the dock (search field, filter chips, the list-level notices that already sat between them and the rows) isabsolute top-0over the lanes, and the lanes fill the panel to its top edge. At rest nothing is under the glass: the dock's height comes from the existinghooks/useMeasuredHeight.ts(the composer's own strip measurement — a callback ref plusResizeObserver, so ListDock adds no measurement code of its own) and is written to--list-dock-hon the wrapper;LIST_BODY_CLS(components/listShell.ts, the class both lists' scrollers wear) pads its top by that variable —pt-[var(--list-dock-h,0.5rem)], falling back to the recipe's ownp-2for a body mounted with no dock — and sets the same value asscroll-pt-[…], so a row brought in byscrollIntoView(the roster's deep-linked?member=row,block: 'nearest') lands below the glass instead of under it. Measured on the built sidebar: pane bottom 125px, first row top 129px, so the first row starts 4px below the pane and only slides under it once the user scrolls. A chip row appearing or an error notice mounting grows the dock and the padding follows in the same frame. The dock wrapper ispointer-events-nonewith each direct child re-enabled, so the empty space beside a short chip row still reaches the row underneath.Pinned folder headers pin on the dock's bottom edge. A sticky folder header offsets itself from the lane's padding edge by
--folder-row-sticky-inset(index.css), which was the lane's ownp-2. Under the dock the lane's padding is the dock's height, so with the old inset a pinned header sat 4px under the dock. The variable now readsvar(--list-dock-pin-inset, 0.5rem)and ListDock sets it to 0: the header sits exactly on the dock's bottom edge whatever the dock's last block is — the shelf's 4px fade under a bare field, a chip row, a notice with no bottom margin — so it never tucks under the dock's surface and never leaves a clear strip the rows show through (measured: dock bottom 44px from the lane top, pinned header top 44px).playwright/sidebar-sticky-folder-header.spec.tsmeasures headers against that pin edge (the dock's bottom, or 0 with no dock over the lane) instead of the lane's top.Nothing else starts under the dock. The two list-level notices that rendered above the lane switch (remote instances loading, remote read failed) moved into the dock with the other notices; the board lane's container pads its top by the dock height (the columns scroll on their own inside the strip), so its hidden-folders line and board error stay reachable.
The shelf stands on a scrim. The dock is two layers in two materials. The field is the glass pane — rows show through it blurred, which is the point. What hangs under it (the filter chips, the list-level notices) is translucent tint tuned for the solid panel it used to sit on (a chip fills at 10% of its colour, a notice at 10% danger), so rows sliding under them rendered text through text.
ListDocknow takesfieldandshelfseparately and stands the shelf on.list-dock-shelf(index.css): the panel's own surface faded in over the first 8px below the pane — the same move as the composer's context shelf onto the page colour (.glass-shelf, #14887) — in the same surface tokens as the pinned folder header, so scrim, pinned header and card are one colour on every surface (kiro-light's--panel, the/embed/sessionshost's--bg). With nothing on the shelf it is 0px tall and paints nothing. The two sidebar notices also takeactionPlacement="below", ErrorNotice's own escape hatch for a ~300px host, so their message no longer wraps to a word or two per line beside the action column.Pointer events follow what the eye sees. The dock wrapper is
pointer-events-none, so the transparent margin around the field passes a click to whatever is under it; the glass field and every focusable in the band (button, link, input, select, textarea, a tabindex'd orrole=buttonelement) re-arm. The shelf re-arms as a whole: it is opaque, and a click on what reads as blank panel beside a chip must land inert rather than open a session the scrim hides — the first cut let it through, and the blind read named that as an invisible target. Both facts are hit-tested in CI:playwright/sidebar-glass-dock.spec.ts(the E2E gate) asserts on the Sessions sidebar and the roster that every focusable inside the dock is whatelementFromPointreturns at its centre, and that the space beside the lone filter chip lands on the dock, not on a row — so a future shelf control that forgot to re-arm goes red in CI instead of dead-clicking in production. The capture harness runs the same two assertions for its screenshots. The dock sits atz-30, above a pinned folder header'sFOLDER_ROW_STICKY_Z(20): a header being pushed off its pin by the next folder travels up through the band, and atz-10its opaque surface painted over the glass.Keyboard reveal counts from the padding edge.
clearPinnedHeaders(chat/sessionRowNav.ts) re-aligns an arrow-key-focused row thatblock: 'nearest'left under the pinned header stack, by testing whether the row lies within the stack's height of the lane's top. With the dock the stack starts at the lane's scroll-padding edge, so the band isscroll-padding-top + margin; the test now adds the scroller'sscroll-padding-top(0 with no dock). This is what the Playwright sticky spec's arrow-key claim caught on the previous head (row 4 landed 11px under the pinned header).useMeasuredHeightsurvives StrictMode. The hook's only disconnect lived in a cleanup-only effect; under StrictMode (dev) React replays cleanup + effect on mount while the callback ref runs once, so the observer was dead after the first render and the height frozen at its mount value — the dock stayed at its search-row height while an error notice grew it to 211px (caught by the harness's padding assertion). The effect now re-attaches to the node it still holds;useMeasuredHeight.test.tsxrenders the hook under<StrictMode>and requires exactly one live observer that still updates the height. Production was not affected (no replay), but every dev-server read of the composer's strip heights was.Where it is mounted.
ChatSidebar.tsx: the search row, the tag chip, the filter chip row and the seed / folders / columns / folder-sort / rename notices move into the dock; the four lanes (tree, conductor, flat, board) are the children.MembersPage.tsx: the search row, the aggregate filter chip, the star-error notice and the gone-member notice move into the dock; the roster<ul>is the child. Neither page changes what it renders, only where the dock sits.Known limit. In the board lane the columns scroll individually inside a horizontal strip, so the lane pads its top by
--list-dock-hinstead and column headers start below the glass rather than scrolling under it. Making the columns scroll under the dock means restructuring the column strip, which is out of this PR's scope.Backwards compatibility
Compatible: no API, storage or i18n key changes; every test id, accessible name and placeholder on the field is unchanged;
LIST_BODY_CLSgains one utility with a fallback equal to its previous padding.Tests
src/test/ListDock.test.tsx(new): the measured dock height lands in--list-dock-h, a later ResizeObserver callback updates it, the dock wrapper is click-through with its children re-enabled, and children render as siblings of the dock.src/test/listShellParity.test.ts: theLIST_BODY_CLSpin updated to the new string.src/test/useMeasuredHeight.test.tsx(new): under<StrictMode>exactly one observer survives the effect replay and a later resize still updates the height.src/test/ChatSidebar.arrowNav.test.tsx: a row 60px into a lane with 44pxscroll-padding-topand a 56px header stack is re-aligned; one at 104px is left alone.src/test/ListDock.test.tsx: also pins the per-control pointer re-arming (no[&>*]), thez-30band, and that the shelf slot carries the scrim class and holds what hangs under the field.playwright/sidebar-glass-dock.spec.ts(new, browser gate): every focusable inside the dock hit-tests to itself on both lists; the space beside the filter chip lands on the inert shelf, not on a hidden row.playwright/sidebar-sticky-folder-header.spec.ts(browser gate): pinned headers are measured against the pin edge — the glass field's bottom — on all three surfaces; the nested-offset, hand-off, opacity and arrow-key claims are unchanged.searchFilterBar.parity.test.tsxand the two pages' existing tests are unchanged and still address the same ids and names.Manual verification
website/scripts/capture-glass-search-dock.mjs(new; nothing in CI runs it) photographs the real SPA over the shared stubbed API on/and/members, dark and light, at rest, scrolled and while typing. It asserts a livebackdrop-filter: blur()layer inside the field and that each scroller's top padding equals the dock's height. The roster fixture holds 24 crewmates so the Members scrolled frame is a real scroll, not a rest frame. Extra scenes photograph the dock's GROWN states over scrolling rows: the running-only filter chip (rest + scrolled; the space beside the chip is hit-tested to reach the row), a pinned folder header measured on the pane's bottom edge (glassBottom === headerTop), the remote-instances error notice in the Sessions dock, a refused star write's notice in the roster dock (the dock's height is asserted equal to the scroller's padding in every scene), and the board lane stepping below the field.text-muted/50to/70(it now sits over scrolled rows), and a genuinely scrolled Members frame. Its "the Members pane is inset relative to the cards" read did not reproduce: measured on the built roster the pane and the row buttons share the same edges (x 245–491 at 1280 wide).tsc,eslinton the changed files,i18n:check,scripts/docs-lint.sh.Screenshots / video
Sessions sidebar, dark — at rest (first row below the pane), scrolled (rows under the pane), typing:
Sessions sidebar, dark — the dock grown by an active filter chip, at rest and with rows scrolled under field and chip:
Sessions sidebar — a pinned folder header sitting on the glass field's bottom edge while its folder scrolls (dark, light):
List-level notices in the dock over scrolled rows — the Sessions remote-instances error, the roster's refused star write:
Board lane, dark — columns step below the field (declared limit):
Crew Members roster, dark — at rest and scrolled (24-crewmate roster, so the scroll is real):
Light, and the full Members page
Review dispositions (round 1 → this head)
<svg>: fixed — lucideSearch.focus-cue-okon the input naming.glass-shadow:focus-withinas the cue.scrollIntoViewlands under the dock: fixed —scroll-pt-[var(--list-dock-h,0.5rem)]onLIST_BODY_CLS.--list-dock-pin-inset, spec updated to the pin edge.useMeasuredHeight: fixed — it uses the hook;className/testIdprops and the default export dropped.Review dispositions (round 2 → this head)
SearchFilterBar+ListDockand the floating geometry.z-10under a pushed pinned header: fixed —z-30.scroll-padding-top + margin; unit case added; this is also the E2E red on the previous head (the sticky spec's arrow-key claim; the two "skipped" specs were its serial siblings, main is green).[&>*]:pointer-events-autore-arms the whole band: fixed — per-control re-arming; harness hit-tests the gap beside a chip.Review dispositions (round 3 → this head)
--list-dock-pin-insethardcoded to the search row's gap so a notice as last block put the header 4px under the dock: fixed — the inset is 0 and the header pins on the dock's bottom edge for every last block; the shelf always paints its 4px fade under a bare field so that edge is soft, not a clear slot.padding + marginovershot the header stack by the inset: fixed by the same change — with inset 0 the stack ends exactly atpadding + margin.sidebar-glass-dock.spec.ts) as well as the harness.actionPlacement="below"on the two sidebar notices: taken.Review dispositions (round 4 → this head)
backdrop-filter, reduced transparency): already true on the base after fix(ui): fade the context shelf onto page colour; drop the composer's accent focus #14887, whose solidify blocks draw.liquid-glass { outline: 1px solid var(--border) }; this branch is rebased onto it and the field inherits the edge.playwright/sidebar-glass-dock.spec.tsruns them in the E2E gate.Review dispositions (round 5 → this head)
z-30) sat above the sidebar / roster resize grip (z-10) and took the inner half of its 6px strip: fixed — both grips move toz-40, above the dock.bebaaccf71:test_diag_procs::test_a_read_refused_as_too_soon_does_not_restart_the_window, a 0.8s timing floor on a shared runner; this PR touches no backend file.Related Issues
Follows #14460 and #14568 (the composer and the mobile Settings capsule).
Checklist