diff --git a/src/main/runtime/terminal-tail-buffer.test.ts b/src/main/runtime/terminal-tail-buffer.test.ts
new file mode 100644
index 00000000000..9851bd7fdc6
--- /dev/null
+++ b/src/main/runtime/terminal-tail-buffer.test.ts
@@ -0,0 +1,157 @@
+import { describe, expect, it, vi } from 'vitest'
+import { appendNormalizedToTailBuffer } from './terminal-tail-buffer'
+import { MAX_TAIL_LINES } from './terminal-tail-limits'
+import type { RetainedTailRedrawCursor } from './terminal-tail-redraw-buffer'
+
+// Guards the per-chunk prefix work in appendNormalizedToTailBuffer: the retained char total is
+// carried across appends and the redraw prefix is not re-scanned, so a saturated tail must not be
+// walked once per chunk. Correctness is pinned by the cold/warm differential below — a "cold" run
+// hands every append a fresh array so the memo always misses and every total is summed in full.
+
+type TailSim = {
+  lines: string[]
+  partialLine: string
+  redrawCursor: RetainedTailRedrawCursor | null
+}
+
+function newSim(): TailSim {
+  return { lines: [], partialLine: '', redrawCursor: null }
+}
+
+type Step = ReturnType<typeof appendNormalizedToTailBuffer>
+
+function feed(sim: TailSim, chunk: string, cold: boolean): Step {
+  const next = appendNormalizedToTailBuffer(
+    cold ? [...sim.lines] : sim.lines,
+    sim.partialLine,
+    chunk,
+    sim.redrawCursor
+  )
+  sim.lines = next.lines
+  sim.partialLine = next.partialLine
+  sim.redrawCursor = next.redrawCursor
+  return next
+}
+
+function mulberry32(seed: number): () => number {
+  let state = seed >>> 0
+  return () => {
+    state = (state + 0x6d2b79f5) >>> 0
+    let t = Math.imul(state ^ (state >>> 15), 1 | state)
+    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t
+    return ((t ^ (t >>> 14)) >>> 0) / 4294967296
+  }
+}
+
+const ESC = String.fromCharCode(27)
+
+/**
+ * `short` fills the 2000-line cap and mixes in TUI redraws; `long` streams lines wide enough to
+ * hit the 256 KiB character cap first. Both eviction paths adjust the carried character total, so
+ * both need differential coverage, and a redraw's row truncation would keep `long` off its cap.
+ */
+function randomChunk(random: () => number, profile: 'short' | 'long'): string {
+  const roll = random()
+  if (profile === 'long') {
+    if (roll < 0.7) {
+      return `${'w'.repeat(1000 + Math.floor(random() * 3000))}\n`
+    }
+    if (roll < 0.8) {
+      return `\rspinner ${Math.floor(random() * 100)}%`
+    }
+    if (roll < 0.9) {
+      return 'trailing spaces here   \n'
+    }
+    return roll < 0.95 ? '' : `no newline ${Math.floor(random() * 1000)}`
+  }
+  if (roll < 0.22) {
+    const lines: string[] = []
+    for (let index = 0; index < 30; index += 1) {
+      lines.push(`burst ${Math.floor(random() * 1e6)}${random() < 0.3 ? '   ' : ''}`)
+    }
+    return `${lines.join('\n')}\n`
+  }
+  if (roll < 0.42) {
+    return `plain output ${Math.floor(random() * 1e6)}\n`
+  }
+  if (roll < 0.56) {
+    return `${'   '.repeat(Math.floor(random() * 3))}\n`
+  }
+  if (roll < 0.68) {
+    const rows = 1 + Math.floor(random() * 12)
+    return `${ESC}[${rows}A${ESC}[2Kredrawn ${Math.floor(random() * 1000)}\n`
+  }
+  if (roll < 0.8) {
+    return `\rspinner ${Math.floor(random() * 100)}%`
+  }
+  if (roll < 0.86) {
+    return 'trailing spaces here   \n'
+  }
+  if (roll < 0.92) {
+    return `multi\nline\nchunk ${Math.floor(random() * 1000)}\n`
+  }
+  if (roll < 0.96) {
+    return ''
+  }
+  return `no newline ${Math.floor(random() * 1000)}`
+}
+
+describe('retained tail buffer prefix reuse', () => {
+  for (const profile of ['short', 'long'] as const) {
+    for (const seed of [3, 11, 91, 2024]) {
+      it(`carries the retained char total exactly (${profile}, seed ${seed})`, () => {
+        const random = mulberry32(seed)
+        const warm = newSim()
+        const cold = newSim()
+        let sawCap = false
+        for (let step = 0; step < 1400; step += 1) {
+          const chunk = randomChunk(random, profile)
+          const lineCountBefore = warm.lines.length
+          const warmStep = feed(warm, chunk, false)
+          const coldStep = feed(cold, chunk, true)
+          expect(warmStep.lines, `step ${step} lines`).toEqual(coldStep.lines)
+          expect(warmStep.partialLine, `step ${step} partial`).toBe(coldStep.partialLine)
+          expect(warmStep.truncated, `step ${step} truncated`).toBe(coldStep.truncated)
+          expect(warmStep.redrawCursor, `step ${step} cursor`).toEqual(coldStep.redrawCursor)
+          expect(warmStep.newCompleteLines, `step ${step} newCompleteLines`).toBe(
+            coldStep.newCompleteLines
+          )
+          expect(warmStep.newlyCompletedLines, `step ${step} newlyCompletedLines`).toEqual(
+            coldStep.newlyCompletedLines
+          )
+          sawCap =
+            sawCap ||
+            (profile === 'short'
+              ? warm.lines.length >= MAX_TAIL_LINES
+              : // Lines dropped below the line cap on an append-only chunk == character-cap eviction.
+                !chunk.includes(ESC) &&
+                warm.lines.length < MAX_TAIL_LINES &&
+                lineCountBefore + warmStep.newlyCompletedLines.length > warm.lines.length)
+        }
+        // Guard against a vacuous pass: the profile's eviction path must have run.
+        expect(sawCap).toBe(true)
+      })
+    }
+  }
+
+  it('does not walk the untouched redraw prefix on every chunk', () => {
+    const sim = newSim()
+    for (let index = 0; index < MAX_TAIL_LINES + 200; index += 1) {
+      feed(sim, `streaming build output line ${index}\n`, false)
+    }
+    expect(sim.lines.length).toBe(MAX_TAIL_LINES)
+
+    const redrawChunk = `${ESC}[3A${ESC}[2Krewritten row${ESC}[2B\n`
+    const spy = vi.spyOn(String.prototype, 'charCodeAt')
+    let prefixTouches = 0
+    try {
+      feed(sim, redrawChunk, false)
+      prefixTouches = spy.mock.calls.length
+    } finally {
+      spy.mockRestore()
+    }
+    // Before this change the prefix trailing-space scan alone cost one charCodeAt per retained
+    // row (~1990); the chunk itself accounts for well under a hundred.
+    expect(prefixTouches).toBeLessThan(300)
+  })
+})
diff --git a/src/main/runtime/terminal-tail-buffer.ts b/src/main/runtime/terminal-tail-buffer.ts
index 0cf551e92c6..38f929b5d7a 100644
--- a/src/main/runtime/terminal-tail-buffer.ts
+++ b/src/main/runtime/terminal-tail-buffer.ts
@@ -1,4 +1,5 @@
 import { containsTerminalVerticalLineControl } from './terminal-ansi-normalization'
+import { carryTerminalTailSentinelMatches } from './terminal-tail-sentinel-index'
 import {
   applyTerminalLineControls,
   processTerminalTailCompleteSegments,
@@ -11,6 +12,35 @@ import {
   type RetainedTailRedrawCursor
 } from './terminal-tail-redraw-buffer'
 
+type RetainedTailLineStats = {
+  totalChars: number
+  /** Whether every line is already right-trimmed, so the redraw prefix trim is a no-op. */
+  rightTrimmed: boolean
+}
+
+// Why weak + array-keyed: the tail is replaced (never mutated) on every append, so an entry dies
+// with the array it describes and only the live tail per PTY is retained. Carrying the char total
+// this way replaces a full-tail re-sum on every chunk.
+const tailLineStatsByLines = new WeakMap<readonly string[], RetainedTailLineStats>()
+
+function getRetainedTailLineStats(lines: readonly string[]): RetainedTailLineStats {
+  const cached = tailLineStatsByLines.get(lines)
+  if (cached) {
+    return cached
+  }
+  let totalChars = 0
+  let rightTrimmed = true
+  for (const line of lines) {
+    totalChars += line.length
+    if (rightTrimmed && trimTerminalLineRight(line) !== line) {
+      rightTrimmed = false
+    }
+  }
+  const stats = { totalChars, rightTrimmed }
+  tailLineStatsByLines.set(lines, stats)
+  return stats
+}
+
 export function appendNormalizedToTailBuffer(
   previousLines: string[],
   previousPartialLine: string,
@@ -52,7 +82,13 @@ export function appendNormalizedToTailBuffer(
   // Why: status UIs redraw one line via CR/backspace/erase; retain the latest redraw segment instead of appending every spinner frame.
   const segments = splitRetainedTerminalTailSegments(combinedChunk)
   const pieces = processTerminalTailCompleteSegments(segments.completeSegments)
-  const newlyCompletedLines = pieces.map((line) => trimTerminalLineRight(line))
+  const newlyCompletedLines: string[] = []
+  let newlyCompletedChars = 0
+  for (const piece of pieces) {
+    const line = trimTerminalLineRight(piece)
+    newlyCompletedLines.push(line)
+    newlyCompletedChars += line.length
+  }
   const partialResult = applyTerminalLineControls(segments.partialSegment)
   const nextPartialLine = trimTerminalLineRight(partialResult.text)
   const retainedPartialLine = nextPartialLine.slice(-MAX_TAIL_PARTIAL_CHARS)
@@ -67,28 +103,51 @@ export function appendNormalizedToTailBuffer(
     omittedNewCompleteLines > 0 ||
     nextPartialLine.length > MAX_TAIL_PARTIAL_CHARS
 
+  // The plain path only ever appends, so the whole previous tail carries unless it was discarded.
+  const carriesPreviousLines = newCompleteLines === 0 || omittedNewCompleteLines === 0
+  const previousStats = carriesPreviousLines ? getRetainedTailLineStats(previousLines) : null
+  let carriedSourceStart = 0
+  let carriedCount = carriesPreviousLines ? previousLines.length : 0
+  let nextLinesChars =
+    (previousStats?.totalChars ?? 0) + (newCompleteLines > 0 ? newlyCompletedChars : 0)
+
   if (nextLines.length > MAX_TAIL_LINES) {
-    nextLines = nextLines.slice(nextLines.length - MAX_TAIL_LINES)
+    const evictedCount = nextLines.length - MAX_TAIL_LINES
+    for (let index = 0; index < evictedCount; index += 1) {
+      nextLinesChars -= nextLines[index]!.length
+    }
+    nextLines = nextLines.slice(evictedCount)
     truncated = true
+    const carriedShift = Math.min(evictedCount, carriedCount)
+    carriedSourceStart += carriedShift
+    carriedCount -= carriedShift
   }
 
   if (newCompleteLines > 0 || retainedPartialLine.length > previousPartialLine.length) {
-    if (nextLines === previousLines) {
-      nextLines = [...previousLines]
-    }
-    let totalChars =
-      nextLines.reduce((sum, line) => sum + line.length, 0) + retainedPartialLine.length
+    let totalChars = nextLinesChars + retainedPartialLine.length
     let trimStartIndex = 0
     while (trimStartIndex < nextLines.length && totalChars > MAX_TAIL_CHARS) {
-      totalChars -= nextLines[trimStartIndex].length
+      totalChars -= nextLines[trimStartIndex]!.length
       trimStartIndex += 1
     }
     if (trimStartIndex > 0) {
+      nextLinesChars = totalChars - retainedPartialLine.length
       nextLines = nextLines.slice(trimStartIndex)
       truncated = true
+      const carriedShift = Math.min(trimStartIndex, carriedCount)
+      carriedSourceStart += carriedShift
+      carriedCount -= carriedShift
     }
   }
 
+  if (nextLines !== previousLines) {
+    tailLineStatsByLines.set(nextLines, {
+      totalChars: nextLinesChars,
+      rightTrimmed: carriedCount === 0 || (previousStats?.rightTrimmed ?? true)
+    })
+    carryTerminalTailSentinelMatches(previousLines, nextLines, carriedSourceStart, carriedCount)
+  }
+
   const redrawCursor =
     !partialResult.hadControl || partialResult.cursorColumn === nextPartialLine.length
       ? null
@@ -145,13 +204,23 @@ function appendNormalizedToMultilineTailBuffer(
   const windowRows =
     maxUpwardCursorReach(normalizedChunk, previousRedrawCursor) + REDRAW_WINDOW_SAFETY_ROWS
   if (windowRows >= previousLines.length) {
-    return appendNormalizedToMultilineTailBufferUnwindowed(
+    const unwindowed = appendNormalizedToMultilineTailBufferUnwindowed(
       previousLines,
       boundedPreviousPartialLine,
       normalizedChunk,
       previousPartialWasCapped,
       previousRedrawCursor
     )
+    if (unwindowed.lines !== previousLines) {
+      let totalChars = 0
+      for (const line of unwindowed.lines) {
+        totalChars += line.length
+      }
+      // Why nothing carries: an unwindowed redraw may rewrite any retained row.
+      tailLineStatsByLines.set(unwindowed.lines, { totalChars, rightTrimmed: true })
+      carryTerminalTailSentinelMatches(previousLines, unwindowed.lines, 0, 0)
+    }
+    return unwindowed
   }
   const prefixLength = previousLines.length - windowRows
   const suffix = previousLines.slice(prefixLength)
@@ -162,36 +231,59 @@ function appendNormalizedToMultilineTailBuffer(
     previousPartialWasCapped,
     previousRedrawCursor
   )
+  const previousStats = getRetainedTailLineStats(previousLines)
   let lines = previousLines.slice(0, prefixLength)
-  // Why: the shared prefix must match the unwindowed finalize's trailing-space trim without paying a regex per untouched row.
-  for (let index = 0; index < lines.length; index += 1) {
-    const line = lines[index]!
-    const lastChar = line.charCodeAt(line.length - 1)
-    if (lastChar === 32 || lastChar === 9) {
-      lines[index] = line.replace(/[ \t]+$/g, '')
+  let linesChars = previousStats.totalChars
+  for (const line of suffix) {
+    linesChars -= line.length
+  }
+  let carriedSourceStart = 0
+  let carriedCount = prefixLength
+  if (!previousStats.rightTrimmed) {
+    // Why: the shared prefix must match the unwindowed finalize's trailing-space trim without paying a regex per untouched row.
+    for (let index = 0; index < lines.length; index += 1) {
+      const line = lines[index]!
+      const lastChar = line.charCodeAt(line.length - 1)
+      if (lastChar === 32 || lastChar === 9) {
+        const trimmed = line.replace(/[ \t]+$/g, '')
+        lines[index] = trimmed
+        linesChars -= line.length - trimmed.length
+        carriedCount = 0
+      }
     }
   }
   for (const line of windowed.lines) {
     lines.push(line)
+    linesChars += line.length
   }
   let truncated = windowed.truncated
   if (lines.length > MAX_TAIL_LINES) {
-    lines = lines.slice(lines.length - MAX_TAIL_LINES)
+    const evictedCount = lines.length - MAX_TAIL_LINES
+    for (let index = 0; index < evictedCount; index += 1) {
+      linesChars -= lines[index]!.length
+    }
+    lines = lines.slice(evictedCount)
     truncated = true
+    const carriedShift = Math.min(evictedCount, carriedCount)
+    carriedSourceStart += carriedShift
+    carriedCount -= carriedShift
   }
-  let totalChars = windowed.partialLine.length
-  for (const line of lines) {
-    totalChars += line.length
-  }
+  let totalChars = linesChars + windowed.partialLine.length
   let dropCount = 0
   while (dropCount < lines.length && totalChars > MAX_TAIL_CHARS) {
     totalChars -= lines[dropCount]!.length
     dropCount += 1
   }
   if (dropCount > 0) {
+    linesChars = totalChars - windowed.partialLine.length
     lines = lines.slice(dropCount)
     truncated = true
+    const carriedShift = Math.min(dropCount, carriedCount)
+    carriedSourceStart += carriedShift
+    carriedCount -= carriedShift
   }
+  tailLineStatsByLines.set(lines, { totalChars: linesChars, rightTrimmed: true })
+  carryTerminalTailSentinelMatches(previousLines, lines, carriedSourceStart, carriedCount)
   return {
     lines,
     partialLine: windowed.partialLine,
diff --git a/src/main/runtime/terminal-tail-sentinel-index.test.ts b/src/main/runtime/terminal-tail-sentinel-index.test.ts
new file mode 100644
index 00000000000..fcab502da96
--- /dev/null
+++ b/src/main/runtime/terminal-tail-sentinel-index.test.ts
@@ -0,0 +1,277 @@
+import { describe, expect, it, vi } from 'vitest'
+import { appendNormalizedToTailBuffer } from './terminal-tail-buffer'
+import { MAX_TAIL_CHARS, MAX_TAIL_LINES } from './terminal-tail-limits'
+import { buildPreview } from './terminal-tail-state'
+import { tailMayContainBlockedSignal } from './terminal-tail-sentinel-index'
+import { computeTerminalTailWaitState } from './terminal-wait-tail-state'
+import { TERMINAL_WAIT_BLOCKED_SENTINEL_RE } from './terminal-wait-detection'
+import type { RetainedTailRedrawCursor } from './terminal-tail-redraw-buffer'
+
+// The definition the incremental index must reproduce: does ANY retained line (or the
+// partial line) match the sentinel? Written out independently of the implementation.
+function referenceMayContainBlockedSignal(lines: string[], partialLine: string): boolean {
+  for (const line of lines) {
+    if (TERMINAL_WAIT_BLOCKED_SENTINEL_RE.test(line)) {
+      return true
+    }
+  }
+  return TERMINAL_WAIT_BLOCKED_SENTINEL_RE.test(partialLine)
+}
+
+function indexedMayContainBlockedSignal(lines: string[], partialLine: string): boolean {
+  return tailMayContainBlockedSignal(lines) || TERMINAL_WAIT_BLOCKED_SENTINEL_RE.test(partialLine)
+}
+
+type TailSim = {
+  lines: string[]
+  partialLine: string
+  redrawCursor: RetainedTailRedrawCursor | null
+  preview: string
+}
+
+function newSim(): TailSim {
+  return { lines: [], partialLine: '', redrawCursor: null, preview: '' }
+}
+
+function feed(sim: TailSim, chunk: string): void {
+  const next = appendNormalizedToTailBuffer(sim.lines, sim.partialLine, chunk, sim.redrawCursor)
+  sim.lines = next.lines
+  sim.partialLine = next.partialLine
+  sim.redrawCursor = next.redrawCursor
+  sim.preview = buildPreview(next.lines, next.partialLine)
+}
+
+/** A structurally identical tail the index has never seen, so it takes the full-scan path. */
+function unindexed(sim: TailSim): string[] {
+  return [...sim.lines]
+}
+
+function assertMatchesFullScan(sim: TailSim): void {
+  expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(
+    referenceMayContainBlockedSignal(sim.lines, sim.partialLine)
+  )
+  expect(computeTerminalTailWaitState(sim.lines, sim.partialLine, sim.preview)).toEqual(
+    computeTerminalTailWaitState(unindexed(sim), sim.partialLine, sim.preview)
+  )
+}
+
+const BLOCKED_LINE = 'Update available! Press Enter to continue.'
+
+function countSentinelTests(run: () => void): number {
+  const spy = vi.spyOn(TERMINAL_WAIT_BLOCKED_SENTINEL_RE, 'test')
+  try {
+    run()
+    return spy.mock.calls.length
+  } finally {
+    spy.mockRestore()
+  }
+}
+
+function saturatedSim(): TailSim {
+  const sim = newSim()
+  for (let index = 0; index < MAX_TAIL_LINES + 400; index += 1) {
+    feed(sim, `streaming build output line ${index}\n`)
+  }
+  expect(sim.lines.length).toBe(MAX_TAIL_LINES)
+  return sim
+}
+
+describe('terminal tail sentinel index', () => {
+  it('tests only the lines an append produced, not the whole retained tail', () => {
+    const sim = saturatedSim()
+    // Warm the index for the current tail identity.
+    computeTerminalTailWaitState(sim.lines, sim.partialLine, sim.preview)
+
+    const incrementalTests = countSentinelTests(() => {
+      for (let index = 0; index < 20; index += 1) {
+        feed(sim, `fresh line ${index}\n`)
+      }
+      computeTerminalTailWaitState(sim.lines, sim.partialLine, sim.preview)
+    })
+
+    const fullScanTests = countSentinelTests(() => {
+      computeTerminalTailWaitState(unindexed(sim), sim.partialLine, sim.preview)
+    })
+
+    expect(fullScanTests).toBeGreaterThanOrEqual(MAX_TAIL_LINES)
+    // 20 appended lines + one partial-line test per compute call.
+    expect(incrementalTests).toBeLessThanOrEqual(25)
+  })
+
+  it('keeps a retained sentinel visible and drops it exactly when it is evicted', () => {
+    const sim = saturatedSim()
+    feed(sim, `${BLOCKED_LINE}\n`)
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(true)
+    assertMatchesFullScan(sim)
+
+    // Push the prompt to the very last retained slot.
+    for (let index = 0; index < MAX_TAIL_LINES - 1; index += 1) {
+      feed(sim, `after prompt ${index}\n`)
+      expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(true)
+    }
+    expect(sim.lines[0]).toBe(BLOCKED_LINE)
+
+    // One more line evicts it.
+    feed(sim, 'evicting line\n')
+    expect(sim.lines.includes(BLOCKED_LINE)).toBe(false)
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(false)
+    assertMatchesFullScan(sim)
+
+    // And it stays gone many chunks later.
+    for (let index = 0; index < 200; index += 1) {
+      feed(sim, `long after eviction ${index}\n`)
+    }
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(false)
+    assertMatchesFullScan(sim)
+  })
+
+  it('drops a sentinel evicted by the retained-character cap', () => {
+    const sim = newSim()
+    feed(sim, `${BLOCKED_LINE}\n`)
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(true)
+    const bulkLine = `${'x'.repeat(4000)}\n`
+    for (let index = 0; index * 4001 < MAX_TAIL_CHARS + 20000; index += 1) {
+      feed(sim, bulkLine)
+    }
+    expect(sim.lines.includes(BLOCKED_LINE)).toBe(false)
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(false)
+    assertMatchesFullScan(sim)
+  })
+
+  it('finds a sentinel split across two chunks once the line completes', () => {
+    const sim = saturatedSim()
+    feed(sim, 'Codex asks: press ent')
+    // Still only a partial line, and no alternative matches the fragment yet.
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(false)
+    assertMatchesFullScan(sim)
+
+    feed(sim, 'er to confirm')
+    // Now complete, but still the partial line — the partial is always tested directly.
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(true)
+    assertMatchesFullScan(sim)
+
+    feed(sim, '\n')
+    // And once it becomes a retained line the index carries it.
+    expect(sim.partialLine).toBe('')
+    expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(true)
+    assertMatchesFullScan(sim)
+  })
+
+  it('full-scans a tail array the index has never seen (seed/restore path)', () => {
+    // primeWaitBlockedBaselineFromSeededTail reads whatever tail the restore seed installed.
+    const seeded = ['boot log', BLOCKED_LINE, 'trailing']
+    expect(tailMayContainBlockedSignal(seeded)).toBe(true)
+    const state = computeTerminalTailWaitState(seeded, '', '')
+    expect(state.fromTail).toBe(true)
+    expect(state.signal?.reason).toBe('codex-update-prompt')
+
+    const clean = ['boot log', 'no prompt here', 'trailing']
+    expect(tailMayContainBlockedSignal(clean)).toBe(false)
+    expect(computeTerminalTailWaitState(clean, '', '').signal).toBeNull()
+  })
+
+  it('reports fromTail from a blank tail without consulting the index', () => {
+    const sim = newSim()
+    feed(sim, '   \n\t\n')
+    expect(computeTerminalTailWaitState(sim.lines, sim.partialLine, '').fromTail).toBe(false)
+    feed(sim, 'now visible\n')
+    expect(computeTerminalTailWaitState(sim.lines, sim.partialLine, '').fromTail).toBe(true)
+  })
+})
+
+// Deterministic PRNG so a divergence is reproducible from the seed alone.
+function mulberry32(seed: number): () => number {
+  let state = seed >>> 0
+  return () => {
+    state = (state + 0x6d2b79f5) >>> 0
+    let t = Math.imul(state ^ (state >>> 15), 1 | state)
+    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t
+    return ((t ^ (t >>> 14)) >>> 0) / 4294967296
+  }
+}
+
+const ESC = String.fromCharCode(27)
+
+/**
+ * `streaming` saturates and evicts the retained tail; `tui` trades saturation for redraw
+ * coverage (cursor-up rewrites of retained rows, and reaches past the redraw window).
+ */
+function randomChunk(random: () => number, profile: 'streaming' | 'tui'): string {
+  const roll = random()
+  if (roll < 0.2) {
+    const lines: string[] = []
+    for (let index = 0; index < 30; index += 1) {
+      lines.push(`burst line ${Math.floor(random() * 1e6)}`)
+    }
+    return `${lines.join('\n')}\n`
+  }
+  if (roll < 0.42) {
+    return `plain output ${Math.floor(random() * 1e6)}\n`
+  }
+  if (roll < 0.48) {
+    return `${'   '.repeat(Math.floor(random() * 3))}\n`
+  }
+  if (roll < 0.54) {
+    return `${BLOCKED_LINE}\n`
+  }
+  if (roll < 0.58) {
+    return 'do you trust the files in this folder?\n'
+  }
+  if (roll < 0.63) {
+    // Sentinel split across a chunk boundary.
+    return random() < 0.5 ? 'Codex asks: press ent' : 'er to confirm\n'
+  }
+  if (roll < 0.7) {
+    // TUI redraw: move the cursor up a few rows and rewrite them.
+    const rows = 1 + Math.floor(random() * 12)
+    return `${ESC}[${rows}A${ESC}[2Kredrawn row ${Math.floor(random() * 1000)}\n`
+  }
+  if (roll < (profile === 'tui' ? 0.76 : 0.7)) {
+    // Deep redraw that outruns the window and forces the unwindowed path.
+    return `${ESC}[${1500 + Math.floor(random() * 800)}A${ESC}[2Kdeep redraw\n`
+  }
+  if (roll < 0.82) {
+    return `\rspinner ${Math.floor(random() * 100)}%`
+  }
+  if (roll < 0.87) {
+    return 'trailing spaces here   \n'
+  }
+  if (roll < 0.91) {
+    return `${'y'.repeat(3000)}\n`
+  }
+  if (roll < 0.95) {
+    return `multi\nline\nchunk ${Math.floor(random() * 1000)}\n`
+  }
+  if (roll < 0.97) {
+    return ''
+  }
+  return `no newline ${Math.floor(random() * 1000)}`
+}
+
+describe('terminal tail sentinel index property', () => {
+  for (const profile of ['streaming', 'tui'] as const) {
+    for (const seed of [1, 7, 42, 1337]) {
+      it(`matches a full scan on every step of a random ${profile} sequence (seed ${seed})`, () => {
+        const random = mulberry32(seed)
+        const sim = newSim()
+        let sawSentinel = false
+        let sawSaturation = false
+        for (let step = 0; step < 1200; step += 1) {
+          feed(sim, randomChunk(random, profile))
+          const expected = referenceMayContainBlockedSignal(sim.lines, sim.partialLine)
+          expect(indexedMayContainBlockedSignal(sim.lines, sim.partialLine)).toBe(expected)
+          expect(computeTerminalTailWaitState(sim.lines, sim.partialLine, sim.preview)).toEqual(
+            computeTerminalTailWaitState(unindexed(sim), sim.partialLine, sim.preview)
+          )
+          sawSentinel = sawSentinel || expected
+          sawSaturation = sawSaturation || sim.lines.length >= MAX_TAIL_LINES
+        }
+        // Guard against a vacuous pass.
+        expect(sawSentinel).toBe(true)
+        if (profile === 'streaming') {
+          expect(sawSaturation).toBe(true)
+        }
+      })
+    }
+  }
+})
diff --git a/src/main/runtime/terminal-tail-sentinel-index.ts b/src/main/runtime/terminal-tail-sentinel-index.ts
new file mode 100644
index 00000000000..1c4d67bd0a3
--- /dev/null
+++ b/src/main/runtime/terminal-tail-sentinel-index.ts
@@ -0,0 +1,77 @@
+import { TERMINAL_WAIT_BLOCKED_SENTINEL_RE } from './terminal-wait-detection'
+
+/**
+ * Which retained tail lines match the wait-blocked sentinel, memoized per
+ * lines-array identity.
+ *
+ * Why: `computeTerminalTailWaitState` must prove the ABSENCE of a signal, so it
+ * cannot early-exit and re-tested all 2000 retained lines on every scan (20/s
+ * per streaming PTY) even though only ~20 lines were new. Keyed weakly by the
+ * array so an entry dies with the tail it describes; the tail array is replaced
+ * on every append and never mutated in place, so at most one entry per PTY
+ * stays live.
+ */
+const sentinelMatchesByTailLines = new WeakMap<readonly string[], number[]>()
+
+function collectSentinelMatches(
+  lines: readonly string[],
+  startIndex: number,
+  into: number[]
+): void {
+  for (let index = startIndex; index < lines.length; index += 1) {
+    if (TERMINAL_WAIT_BLOCKED_SENTINEL_RE.test(lines[index]!)) {
+      into.push(index)
+    }
+  }
+}
+
+/** Ascending indices of sentinel-matching lines; full-scans an unseen array. */
+export function getTerminalTailSentinelMatches(lines: readonly string[]): readonly number[] {
+  const cached = sentinelMatchesByTailLines.get(lines)
+  if (cached) {
+    return cached
+  }
+  const matches: number[] = []
+  collectSentinelMatches(lines, 0, matches)
+  sentinelMatchesByTailLines.set(lines, matches)
+  return matches
+}
+
+export function tailMayContainBlockedSignal(lines: readonly string[]): boolean {
+  return getTerminalTailSentinelMatches(lines).length > 0
+}
+
+/**
+ * Derive `nextLines`' match index from `previousLines`', testing only the lines
+ * the append actually produced.
+ *
+ * The caller guarantees `nextLines[0 … carriedCount)` are the very same strings
+ * as `previousLines[carriedSourceStart … carriedSourceStart + carriedCount)`,
+ * and that every later line is newly produced. Matches outside that carried
+ * window are dropped because their lines were evicted or rewritten, which is
+ * exactly what the full scan would conclude.
+ */
+export function carryTerminalTailSentinelMatches(
+  previousLines: readonly string[],
+  nextLines: readonly string[],
+  carriedSourceStart: number,
+  carriedCount: number
+): void {
+  if (nextLines === previousLines) {
+    return
+  }
+  const matches: number[] = []
+  if (carriedCount > 0) {
+    const carriedEnd = carriedSourceStart + carriedCount
+    for (const index of getTerminalTailSentinelMatches(previousLines)) {
+      if (index >= carriedEnd) {
+        break
+      }
+      if (index >= carriedSourceStart) {
+        matches.push(index - carriedSourceStart)
+      }
+    }
+  }
+  collectSentinelMatches(nextLines, carriedCount, matches)
+  sentinelMatchesByTailLines.set(nextLines, matches)
+}
diff --git a/src/main/runtime/terminal-wait-tail-state.ts b/src/main/runtime/terminal-wait-tail-state.ts
index 2b9cafee27b..712b301a961 100644
--- a/src/main/runtime/terminal-wait-tail-state.ts
+++ b/src/main/runtime/terminal-wait-tail-state.ts
@@ -1,5 +1,6 @@
 import type { RuntimeTerminalWaitBlockedReason } from '../../shared/runtime-types'
 import { buildTailLines } from './terminal-tail-state'
+import { tailMayContainBlockedSignal } from './terminal-tail-sentinel-index'
 import {
   findActionableTerminalWaitBlockedSignal,
   TERMINAL_WAIT_BLOCKED_SENTINEL_RE
@@ -60,23 +61,22 @@ function inspectTerminalWaitTail(
   lines: string[],
   partialLine: string
 ): { fromTail: boolean; mayContainBlockedSignal: boolean } {
-  let fromTail = false
-  let mayContainBlockedSignal = false
+  return {
+    fromTail: hasVisibleTailLine(lines) || partialLine.trim().length > 0,
+    // Why the index: proving a signal is ABSENT can't early-exit, so a full re-test of the
+    // 2000-line tail ran per scan; the index tests only the lines each append produced.
+    mayContainBlockedSignal:
+      tailMayContainBlockedSignal(lines) || TERMINAL_WAIT_BLOCKED_SENTINEL_RE.test(partialLine)
+  }
+}
+
+function hasVisibleTailLine(lines: string[]): boolean {
   for (const line of lines) {
-    if (!fromTail && line.trim().length > 0) {
-      fromTail = true
-    }
-    if (!mayContainBlockedSignal && TERMINAL_WAIT_BLOCKED_SENTINEL_RE.test(line)) {
-      mayContainBlockedSignal = true
+    if (line.trim().length > 0) {
+      return true
     }
   }
-  if (!fromTail && partialLine.trim().length > 0) {
-    fromTail = true
-  }
-  if (!mayContainBlockedSignal && TERMINAL_WAIT_BLOCKED_SENTINEL_RE.test(partialLine)) {
-    mayContainBlockedSignal = true
-  }
-  return { fromTail, mayContainBlockedSignal }
+  return false
 }
 
 // Why: consumes precomputed wait states so full-tail scans aren't repeated per chunk (replaces the former inline double full-tail scan).