perf(desktop): kill the layout-thrash cascade on session switch

Follow-up to #65890 (router transitions off) and #65898 (structural
compare + first-paint budget): profiling the switch path on real 1000+-
message sessions with a new CDP harness showed the remaining freeze is
NOT markdown rendering — it's a forced-reflow cascade from mount-time
layout reads interleaved with style writes across the transcript's
layout effects, plus the first-paint budget cut landing too late to
stop the full-budget commit.

Measured on the two largest local sessions (996 and 1363 messages),
main-thread longtask totals per switch: warm 2450ms -> 557ms and
1158ms -> 194ms; first paint 1690ms -> 444ms. Harness:
scripts/profile-session-switch.mjs (same CDP family as
profile-real-stream.mjs).

- use-resize-observer: drop the synchronous initial callback and ride
  the observer's spec-guaranteed first delivery instead (same frame,
  after layout, before paint). The sync call ran while the commit's
  layout was dirty, so every size read in a callback forced a full
  reflow — with one instance per user bubble (measureClamp read
  scrollHeight, then WROTE --human-msg-full, re-dirtying layout for the
  next bubble), the switch commit thrashed for over a second. Inside RO
  timing the same reads are free. Composer metrics (2x
  getBoundingClientRect + documentElement style writes) rides the same
  fix.
- Same class, same fix at the remaining call sites profiling surfaced:
  ExpandableBlock and TerminalOutput (dozens per tool-heavy transcript)
  now measure/pin via RO initial delivery; the tool-window and
  thinking-preview pins drop their sync pin() call; the thread
  timeline's initial active-tick compute joins its existing
  scroll-time rAF batching so back-to-back transcript updates coalesce.
- thread/list: cut the render budget in the RENDER phase (state-from-
  props adjustment) instead of the post-commit layout effect. The
  effect-time cut was too late — on a warm switch React first built and
  committed the full 300-part tree, then re-rendered at 60, then bumped
  back to 300, so the expensive commit still happened (and on a cold
  switch the bump rAF usually fired while the transcript was still
  empty, so the prefetched messages rendered at full budget anyway).
  The render-phase cut restarts the component before any child renders;
  a second trigger handles the cold path where messages land later
  under the same sessionKey.
- thread/list: backfill 60 -> 300 inside startTransition so the older
  turns' markdown+shiki render is interruptible background work instead
  of a synchronous freeze one frame after the switch paints. Functional
  Math.max so an urgent "Show earlier" click can't be rebased back down.
- composer focus: skip the rAF/timeout focus retries when the element
  is already focused — focus() runs the full focusing steps (forcing
  layout) even on the active element, ~585ms per switch on a large
  dirty DOM.
- Replace the tautological render-budget test (it re-declared the
  constants locally and asserted 60 < 300) with behavior tests of the
  now-exported buildGroups + firstVisibleGroupIndex.

Verification: apps/desktop `npx tsc --noEmit` clean; full
`npx vitest run` 210 files / 1763 passed; manual CDP check confirms the
deferred backfill commits the full transcript, stays pinned to bottom,
and "Show earlier" still pages.

apps/desktop/scripts/profile-session-switch.mjs

@@ -0,0 +1,163 @@
+// CPU-profile a session switch — outputs a .cpuprofile, a top-self ranking,
+// longtask timings, and paint milestones for cold + warm switches.
+//
+// Drives the real resume path by setting location.hash (same code path as a
+// sidebar click: use-route-resume → resumeSession → prefetch + resume RPC).
+//
+// Usage:
+//   node apps/desktop/scripts/profile-session-switch.mjs <sessionA> <sessionB> [rounds]
+//   OUT=/tmp/switch.cpuprofile node scripts/profile-session-switch.mjs 2026.. 2026..
+
+import { writeFileSync } from 'node:fs'
+
+const CDP_HTTP = 'http://127.0.0.1:9222'
+const A = process.argv[2]
+const B = process.argv[3]
+const ROUNDS = Number(process.argv[4] || 2)
+const OUT = process.env.OUT || `/tmp/session-switch-${Date.now()}.cpuprofile`
+const SETTLE_TIMEOUT = Number(process.env.SETTLE_TIMEOUT || 30000)
+
+if (!A || !B) {
+  console.error('usage: profile-session-switch.mjs <sessionA> <sessionB> [rounds]')
+  process.exit(1)
+}
+
+class CDP {
+  constructor(ws) { this.ws = ws; this.id = 0; this.pending = new Map() }
+  static async open(url) {
+    const ws = new WebSocket(url)
+    await new Promise((r) => ws.addEventListener('open', r, { once: true }))
+    const cdp = new CDP(ws)
+    ws.addEventListener('message', (ev) => {
+      const m = JSON.parse(ev.data.toString())
+      if (m.id != null && cdp.pending.has(m.id)) {
+        const { resolve, reject } = cdp.pending.get(m.id)
+        cdp.pending.delete(m.id)
+        if (m.error) reject(new Error(m.error.message))
+        else resolve(m.result)
+      }
+    })
+    return cdp
+  }
+  send(method, params) {
+    const id = ++this.id
+    return new Promise((res, rej) => {
+      this.pending.set(id, { resolve: res, reject: rej })
+      this.ws.send(JSON.stringify({ id, method, params }))
+    })
+  }
+  async eval(expr) {
+    const r = await this.send('Runtime.evaluate', { expression: expr, returnByValue: true, awaitPromise: true })
+    if (r.exceptionDetails) throw new Error(r.exceptionDetails.exception?.description || 'eval failed')
+    return r.result.value
+  }
+  close() { this.ws.close() }
+}
+
+async function main() {
+  const list = await (await fetch(`${CDP_HTTP}/json`)).json()
+  const target = list.find((t) => t.type === 'page' && /5174/.test(t.url))
+  if (!target) { console.error('renderer not found on 9222'); process.exit(1) }
+  const cdp = await CDP.open(target.webSocketDebuggerUrl)
+
+  // Install observers once: longtasks + rAF frame gaps, tagged per switch.
+  await cdp.eval(`(() => {
+    if (window.__SWITCH_OBS__) return 'already'
+    const obs = { longtasks: [], marks: [] }
+    new PerformanceObserver((l) => {
+      for (const e of l.getEntries()) obs.longtasks.push({ t: e.startTime, dur: e.duration })
+    }).observe({ entryTypes: ['longtask'] })
+    window.__SWITCH_OBS__ = obs
+    return 'installed'
+  })()`)
+
+  const switchTo = async (sid, label) => {
+    const t0 = await cdp.eval(`(() => {
+      const o = window.__SWITCH_OBS__
+      o.marks.push({ label: ${JSON.stringify(label)}, sid: ${JSON.stringify(sid)}, t: performance.now() })
+      location.hash = '#/' + ${JSON.stringify(sid)}
+      return performance.now()
+    })()`)
+
+    // Poll until the transcript for this session has painted and settled:
+    // route matches, >0 message roots, and message count stable for 3 polls.
+    const deadline = Date.now() + SETTLE_TIMEOUT
+    let stable = 0
+    let lastCount = -1
+    let firstPaintT = null
+    while (Date.now() < deadline) {
+      await new Promise((r) => setTimeout(r, 50))
+      const s = await cdp.eval(`({
+        t: performance.now(),
+        route: location.hash,
+        msgs: document.querySelectorAll('[data-slot="aui_message"], [data-slot="aui_assistant-message-root"], [data-slot="aui_user-message-root"]').length,
+        parts: document.querySelectorAll('[data-slot="aui_thread-content"] *').length
+      })`)
+      if (!s.route.includes(sid)) continue
+      if (s.msgs > 0 && firstPaintT === null) firstPaintT = s.t
+      stable = s.msgs === lastCount && s.msgs > 0 ? stable + 1 : 0
+      lastCount = s.msgs
+      if (stable >= 3) return { t0, firstPaintT, settledT: s.t, msgs: s.msgs, domNodes: s.parts }
+    }
+    return { t0, firstPaintT, settledT: null, msgs: lastCount, timedOut: true }
+  }
+
+  console.log('starting CPU profile')
+  await cdp.send('Profiler.enable')
+  await cdp.send('Profiler.setSamplingInterval', { interval: 100 })
+  await cdp.send('Profiler.start')
+
+  const results = []
+  for (let round = 0; round < ROUNDS; round++) {
+    for (const [sid, tag] of [[A, 'A'], [B, 'B']]) {
+      const label = `round${round}:${tag}:${round === 0 ? 'cold' : 'warm'}`
+      const r = await switchTo(sid, label)
+      results.push({ label, sid, ...r })
+      const ftp = r.firstPaintT != null ? (r.firstPaintT - r.t0).toFixed(0) : 'n/a'
+      const st = r.settledT != null ? (r.settledT - r.t0).toFixed(0) : 'TIMEOUT'
+      console.log(`${label.padEnd(18)} first-paint ${String(ftp).padStart(6)} ms   settled ${String(st).padStart(6)} ms   msgs ${r.msgs}  dom ${r.domNodes ?? '?'}`)
+      await new Promise((r2) => setTimeout(r2, 800))
+    }
+  }
+
+  const { profile } = await cdp.send('Profiler.stop')
+  writeFileSync(OUT, JSON.stringify(profile))
+  console.log('\nwrote', OUT)
+
+  // Longtasks per switch window.
+  const obs = await cdp.eval('window.__SWITCH_OBS__')
+  console.log('\n=== LONGTASKS (>=50ms main-thread blocks) ===')
+  for (let i = 0; i < obs.marks.length; i++) {
+    const m = obs.marks[i]
+    const end = obs.marks[i + 1]?.t ?? Infinity
+    const lts = obs.longtasks.filter((lt) => lt.t >= m.t && lt.t < end)
+    const total = lts.reduce((a, b) => a + b.dur, 0)
+    console.log(`${m.label.padEnd(18)} ${String(lts.length).padStart(2)} longtasks, ${total.toFixed(0).padStart(5)} ms total  ${lts.map((l) => Math.round(l.dur)).join(', ')}`)
+  }
+
+  // Self-time ranking.
+  const samples = profile.samples || []
+  const timeDeltas = profile.timeDeltas || []
+  const nodes = new Map(profile.nodes.map((n) => [n.id, n]))
+  const selfTime = new Map()
+  for (let i = 0; i < samples.length; i++) {
+    selfTime.set(samples[i], (selfTime.get(samples[i]) || 0) + (timeDeltas[i] ?? 0))
+  }
+  const ranked = [...selfTime.entries()]
+    .map(([id, us]) => {
+      const cf = nodes.get(id)?.callFrame || {}
+      return { ms: us / 1000, name: cf.functionName || '(anonymous)', url: (cf.url || '').slice(-70), line: cf.lineNumber }
+    })
+    .filter((x) => !/\(root\)|\(idle\)|\(garbage collector\)|\(program\)/.test(x.name))
+    .sort((a, b) => b.ms - a.ms)
+    .slice(0, 30)
+
+  console.log('\n=== TOP 30 SELF TIME (ms) ACROSS ALL SWITCHES ===')
+  for (const r of ranked) {
+    console.log(`${r.ms.toFixed(1).padStart(8)}  ${r.name.padEnd(44)}  ${r.url}:${r.line}`)
+  }
+
+  cdp.close()
+}
+
+main().catch((e) => { console.error(e); process.exit(1) })

apps/desktop/src/app/chat/composer/focus.ts

@@ -157,7 +157,14 @@ export const focusComposerInput = (el: HTMLElement | null) => {
     return
   }
 
-  const focus = () => el.focus({ preventScroll: true })
+  // Skip when already focused: focus() runs the full focusing steps (forcing
+  // layout) even on the active element, and during a session switch the DOM is
+  // large and dirty — the redundant retries were measurably expensive there.
+  const focus = () => {
+    if (document.activeElement !== el) {
+      el.focus({ preventScroll: true })
+    }
+  }
 
   focus()
   window.requestAnimationFrame(focus)

apps/desktop/src/components/assistant-ui/thread/list.test.ts

@@ -1,26 +1,82 @@
 import { describe, expect, it } from 'vitest'
 
-describe('ThreadMessageList render budget constants', () => {
-  it('defines FIRST_PAINT_BUDGET as smaller than RENDER_BUDGET', () => {
-    // These constants control the deferred render budget on session switch.
-    // FIRST_PAINT_BUDGET should be small enough to render quickly (just the
-    // bottom turn(s) visible after scroll-to-bottom), while RENDER_BUDGET
-    // is the full cap for "Show earlier" increments and the eventual full
-    // transcript after the first paint.
-    const RENDER_BUDGET = 300
-    const FIRST_PAINT_BUDGET = 60
+import { buildGroups, firstVisibleGroupIndex, type MessageGroup } from './list'
 
-    expect(FIRST_PAINT_BUDGET).toBeLessThan(RENDER_BUDGET)
-    expect(FIRST_PAINT_BUDGET).toBeGreaterThan(0)
+// Signature rows are `${index}:${id}:${role}:${weight}` (see the useAuiState
+// selector in list.tsx).
+const signature = (rows: [string, string, number][]) =>
+  rows.map(([id, role, weight], index) => `${index}:${id}:${role}:${weight}`).join('\n')
+
+describe('buildGroups', () => {
+  it('returns no groups for an empty signature', () => {
+    expect(buildGroups('')).toEqual([])
   })
 
-  it('ensures FIRST_PAINT_BUDGET is sufficient for at least one visible turn', () => {
-    // A typical bottom turn (user + assistant) might have ~20-40 parts total
-    // (including tool calls), so 60 is enough for 1-2 visible turns at the
-    // bottom of the viewport.
-    const FIRST_PAINT_BUDGET = 60
-    const TYPICAL_TURN_PARTS = 30
+  it('groups a user message with the assistant turn(s) that follow it', () => {
+    const groups = buildGroups(
+      signature([
+        ['u1', 'user', 1],
+        ['a1', 'assistant', 4],
+        ['a2', 'assistant', 2],
+        ['u2', 'user', 1],
+        ['a3', 'assistant', 3]
+      ])
+    )
 
-    expect(FIRST_PAINT_BUDGET).toBeGreaterThanOrEqual(TYPICAL_TURN_PARTS)
+    expect(groups).toEqual([
+      { id: 'u1', indices: [0, 1, 2], kind: 'turn', weight: 7 },
+      { id: 'u2', indices: [3, 4], kind: 'turn', weight: 4 }
+    ])
+  })
+
+  it('keeps leading non-user messages as standalone groups', () => {
+    const groups = buildGroups(
+      signature([
+        ['s1', 'system', 1],
+        ['a0', 'assistant', 2],
+        ['u1', 'user', 1],
+        ['a1', 'assistant', 5]
+      ])
+    )
+
+    expect(groups).toEqual([
+      { id: 's1', index: 0, kind: 'standalone', weight: 1 },
+      { id: 'a0', index: 1, kind: 'standalone', weight: 2 },
+      { id: 'u1', indices: [2, 3], kind: 'turn', weight: 6 }
+    ])
+  })
+
+  it('defaults a missing/zero weight to 1', () => {
+    const groups = buildGroups('0:a:assistant:0')
+
+    expect(groups).toEqual([{ id: 'a', index: 0, kind: 'standalone', weight: 1 }])
+  })
+})
+
+describe('firstVisibleGroupIndex', () => {
+  const group = (id: string, weight: number): MessageGroup => ({ id, index: 0, kind: 'standalone', weight })
+
+  it('shows everything when total weight fits the budget', () => {
+    const groups = [group('a', 10), group('b', 10), group('c', 10)]
+
+    expect(firstVisibleGroupIndex(groups, 100)).toBe(0)
+  })
+
+  it('walks newest-first and hides everything before the turn that meets the budget', () => {
+    const groups = [group('old', 50), group('mid', 30), group('new', 30)]
+
+    // newest-first: 30 (new) < 60, +30 (mid) = 60 >= 60 → mid is the first
+    // visible group, old is hidden.
+    expect(firstVisibleGroupIndex(groups, 60)).toBe(1)
+  })
+
+  it('keeps whole turns intact — the turn that crosses the budget stays visible', () => {
+    const groups = [group('old', 5), group('huge', 500)]
+
+    expect(firstVisibleGroupIndex(groups, 60)).toBe(1)
+  })
+
+  it('returns groups.length for an empty list', () => {
+    expect(firstVisibleGroupIndex([], 60)).toBe(0)
   })
 })

apps/desktop/src/components/assistant-ui/thread/list.tsx

@@ -5,6 +5,7 @@ import {
   type FC,
   memo,
   type ReactNode,
+  startTransition,
   useCallback,
   useEffect,
   useLayoutEffect,
@@ -28,7 +29,7 @@ import { MessageRenderBoundary } from '../message-render-boundary'
 
 type ThreadMessageComponents = ComponentProps<typeof ThreadPrimitive.MessageByIndex>['components']
 
-type MessageGroup = { id: string; weight: number } & (
+export type MessageGroup = { id: string; weight: number } & (
   | { index: number; kind: 'standalone' }
   | { indices: number[]; kind: 'turn' }
 )
@@ -58,7 +59,7 @@ interface ThreadMessageListProps {
 // Group each user message with the assistant turn(s) that follow it so the
 // human bubble can `position: sticky` against the scroller across its whole
 // turn (see StickyHumanMessageContainer in thread.tsx).
-function buildGroups(signature: string): MessageGroup[] {
+export function buildGroups(signature: string): MessageGroup[] {
   if (!signature) {
     return []
   }
@@ -94,6 +95,24 @@ function buildGroups(signature: string): MessageGroup[] {
   return groups
 }
 
+// Walk turns newest-first, summing their part weights until the budget is met;
+// everything before the first kept turn is hidden. Returns the index of that
+// first visible group.
+export function firstVisibleGroupIndex(groups: readonly MessageGroup[], budget: number): number {
+  let firstVisible = groups.length
+
+  for (let i = groups.length - 1, weight = 0; i >= 0; i--) {
+    weight += groups[i].weight
+    firstVisible = i
+
+    if (weight >= budget) {
+      break
+    }
+  }
+
+  return firstVisible
+}
+
 const ThreadMessageListInner: FC<ThreadMessageListProps> = ({
   clampToComposer,
   components,
@@ -121,22 +140,59 @@ const ThreadMessageListInner: FC<ThreadMessageListProps> = ({
     resize: 'instant'
   })
 
-  const [renderBudget, setRenderBudget] = useState(RENDER_BUDGET)
+  const [renderBudget, setRenderBudget] = useState(FIRST_PAINT_BUDGET)
 
-  // Walk turns newest-first, summing their part weights until the budget is met;
-  // everything before that first kept turn is hidden.
-  let firstVisible = groups.length
+  // Cut the budget during RENDER, not in the post-commit layout effect. An
+  // effect-time cut is too late: React would first build the whole tree with
+  // the full budget (up to 300 parts of markdown + syntax highlighting),
+  // commit it, and only then re-render at the small budget. The render-phase
+  // state adjustment restarts this component immediately — before any child
+  // renders — so the heavy commit never happens.
+  //
+  // Two triggers, because the transcript swap arrives differently per path:
+  // a WARM switch publishes sessionKey + messages in one commit (the key
+  // branch), while a COLD switch changes sessionKey with an empty transcript
+  // and the prefetched messages land hundreds of ms later under the SAME key
+  // (the empty→non-empty branch).
+  const hasGroups = groups.length > 0
+  const [budgetSessionKey, setBudgetSessionKey] = useState(sessionKey)
+  const [hadGroups, setHadGroups] = useState(hasGroups)
 
-  for (let i = groups.length - 1, weight = 0; i >= 0; i--) {
-    weight += groups[i].weight
-    firstVisible = i
+  if (budgetSessionKey !== sessionKey) {
+    setBudgetSessionKey(sessionKey)
+    setHadGroups(hasGroups)
+    setRenderBudget(FIRST_PAINT_BUDGET)
+  } else if (hadGroups !== hasGroups) {
+    setHadGroups(hasGroups)
 
-    if (weight >= renderBudget) {
-      break
+    if (hasGroups) {
+      setRenderBudget(FIRST_PAINT_BUDGET)
     }
   }
 
-  const hiddenCount = firstVisible
+  // Backfill from FIRST_PAINT_BUDGET to the full budget after the small
+  // commit painted — as a TRANSITION, so the heavy markdown + syntax
+  // highlight render of the older turns is interruptible instead of one long
+  // synchronous commit that freezes input right after the switch. Route
+  // changes stay urgent (main.tsx disables router transitions); it's exactly
+  // this backfill that belongs at background priority. "Show earlier" pages
+  // (budget > RENDER_BUDGET) never re-enter here.
+  useEffect(() => {
+    if (renderBudget >= RENDER_BUDGET) {
+      return
+    }
+
+    const rafId = requestAnimationFrame(() => {
+      // Functional max, not a plain set: an urgent "Show earlier" click can
+      // land between scheduling and committing this transition, and a plain
+      // set would rebase over it and shrink the budget back down.
+      startTransition(() => setRenderBudget(budget => Math.max(budget, RENDER_BUDGET)))
+    })
+
+    return () => cancelAnimationFrame(rafId)
+  }, [renderBudget])
+
+  const hiddenCount = firstVisibleGroupIndex(groups, renderBudget)
   const visibleGroups = hiddenCount > 0 ? groups.slice(hiddenCount) : groups
   const restoreFromBottomRef = useRef<number | null>(null)
   // Secondary windows (new-session scratch, subagent watch, cmd-click pop-out)
@@ -192,12 +248,6 @@ const ThreadMessageListInner: FC<ThreadMessageListProps> = ({
   // Instead: quiet it, glue to the true bottom until the height holds steady,
   // then hand back locked. Live streaming afterward uses the normal resize follow.
   useLayoutEffect(() => {
-    // Start with a small first-paint budget (enough for the bottom turn(s) the
-    // user sees after scroll-to-bottom), then defer the full budget bump to a
-    // requestAnimationFrame so the heavy markdown render happens after the
-    // initial commit.
-    setRenderBudget(FIRST_PAINT_BUDGET)
-
     const el = scrollRef.current
 
     if (!el) {
@@ -238,17 +288,7 @@ const ThreadMessageListInner: FC<ThreadMessageListProps> = ({
 
     let rafId = requestAnimationFrame(settle)
 
-    // After the settle loop starts, bump the render budget to the full value
-    // in a subsequent rAF so the full transcript becomes available after the
-    // first paint.
-    let budgetRafId = requestAnimationFrame(() => {
-      setRenderBudget(RENDER_BUDGET)
-    })
-
-    return () => {
-      cancelAnimationFrame(rafId)
-      cancelAnimationFrame(budgetRafId)
-    }
+    return () => cancelAnimationFrame(rafId)
   }, [scrollRef, scrollToBottom, sessionKey, stopScroll])
 
   // Prepend an older page while preserving the on-screen position. The user is

apps/desktop/src/components/assistant-ui/thread/message-parts.tsx

@@ -84,7 +84,8 @@ const ThinkingDisclosure: FC<{
       el.scrollTop = el.scrollHeight
     }
 
-    pin()
+    // No sync pin(): the observer's guaranteed initial delivery runs it with
+    // layout already clean (still before paint), avoiding a forced reflow.
     const observer = new ResizeObserver(pin)
     observer.observe(content)
 

apps/desktop/src/components/assistant-ui/thread/timeline.tsx

@@ -206,7 +206,13 @@ export const ThreadTimeline: FC = () => {
       }
     }
 
-    compute()
+    // Initial compute rides the same rAF batching as scroll. A sync call here
+    // reads getBoundingClientRect for every user message while other commit
+    // effects are still writing styles — on a session switch that interleaving
+    // forces a full reflow per read on a large transcript. One rAF later the
+    // reads batch into a single layout pass, and back-to-back entries updates
+    // (prefetch paint, then resume reconcile) coalesce into one compute.
+    onScroll()
     viewport.addEventListener('scroll', onScroll, { passive: true })
 
     return () => {

apps/desktop/src/components/assistant-ui/tool/fallback.tsx

@@ -659,7 +659,10 @@ function useToolWindow(enabled: boolean) {
       syncFade()
     }
 
-    pin()
+    // No sync pin() here: the observer's guaranteed initial delivery runs it
+    // inside RO timing (layout already clean, still before paint). A sync call
+    // at effect time reads scrollHeight while the commit's layout is dirty —
+    // one forced reflow per tool group, which cascades on a session switch.
     const observer = new ResizeObserver(pin)
     observer.observe(content)
 

apps/desktop/src/components/chat/expandable-block.tsx

@@ -1,7 +1,8 @@
 'use client'
 
-import { type ReactNode, useLayoutEffect, useRef, useState } from 'react'
+import { type ReactNode, useCallback, useRef, useState } from 'react'
 
+import { useResizeObserver } from '@/hooks/use-resize-observer'
 import { ChevronDown } from '@/lib/icons'
 import { cn } from '@/lib/utils'
 
@@ -15,21 +16,19 @@ export function ExpandableBlock({ children, className }: ExpandableBlockProps) {
   const [expanded, setExpanded] = useState(false)
   const [overflowing, setOverflowing] = useState(false)
 
-  useLayoutEffect(() => {
+  // Measure inside ResizeObserver timing only (layout is clean there). A
+  // synchronous mount-time scrollHeight read forces a reflow per instance,
+  // and a tool-heavy transcript mounts dozens of these on a session switch.
+  const measure = useCallback(() => {
     const el = innerRef.current
 
-    if (!el) {
-      return
+    if (el) {
+      setOverflowing(el.scrollHeight > 121)
     }
-
-    const measure = () => setOverflowing(el.scrollHeight > 121)
-    measure()
-    const observer = new ResizeObserver(measure)
-    observer.observe(el)
-
-    return () => observer.disconnect()
   }, [])
 
+  useResizeObserver(measure, innerRef)
+
   return (
     <div className="relative">
       <div className={cn('overflow-y-auto', expanded ? 'max-h-[40dvh]' : 'max-h-[7.5rem]', className)} ref={innerRef}>

apps/desktop/src/components/chat/terminal-output.tsx

@@ -1,5 +1,6 @@
-import { useEffect, useLayoutEffect, useRef } from 'react'
+import { useCallback, useEffect, useRef } from 'react'
 
+import { useResizeObserver } from '@/hooks/use-resize-observer'
 import { cn } from '@/lib/utils'
 
 interface TerminalOutputProps {
@@ -22,15 +23,25 @@ const NEAR_BOTTOM_PX = 24
 export function TerminalOutput({ className, text }: TerminalOutputProps) {
   const ref = useRef<HTMLDivElement>(null)
 
-  // On open: jump straight to the latest output (no animation, before paint).
-  useLayoutEffect(() => {
+  // On open: jump straight to the latest output. Rides the ResizeObserver's
+  // initial delivery (layout-clean, still before paint) instead of a layout
+  // effect — a sync scrollHeight read there forces a reflow per instance, and
+  // a tool-heavy transcript mounts dozens of these on a session switch.
+  const jumpedRef = useRef(false)
+
+  const jumpToBottom = useCallback(() => {
     const el = ref.current
 
-    if (el) {
+    // First delivery only; later box resizes must not yank the user back
+    // down while they're scrolled up reading earlier output.
+    if (el && !jumpedRef.current) {
+      jumpedRef.current = true
       el.scrollTop = el.scrollHeight
     }
   }, [])
 
+  useResizeObserver(jumpToBottom, ref)
+
   // On growth: tail only when already pinned near the bottom.
   useEffect(() => {
     const el = ref.current

apps/desktop/src/hooks/use-resize-observer.ts

@@ -2,9 +2,16 @@ import { type RefObject, useLayoutEffect, useRef } from 'react'
 
 /**
  * Observe element resizes. The callback receives the ResizeObserver entries
- * (empty on the initial synchronous call and in non-RO environments) so
- * callers can read the observed size off the entry instead of forcing a
- * fresh layout read.
+ * (empty only in non-RO environments) so callers can read the observed size
+ * off the entry instead of forcing a fresh layout read.
+ *
+ * The initial measurement rides the observer's spec-guaranteed first delivery
+ * (same frame, after layout, before paint) instead of a synchronous call from
+ * the layout effect. A sync call here runs while the commit's layout is still
+ * dirty, so any size read in the callback forces a full reflow — and with many
+ * instances mounting at once (every user bubble on a session switch), the
+ * interleaved read→write→read pattern cascades into seconds of layout thrash.
+ * Inside RO timing, layout is already clean and the same reads are ~free.
  */
 export function useResizeObserver(
   onResize: (entries: readonly ResizeObserverEntry[]) => void,
@@ -40,8 +47,6 @@ export function useResizeObserver(
       return
     }
 
-    onResize([])
-
     return () => observer.disconnect()
   }, [onResize])
 }