fix(terminal): close retained string review gaps

src/main/daemon/osc7-uri-extraction.test.ts

@@ -33,9 +33,7 @@ describe('OSC-7 URI extraction', () => {
     expect(execSpy).not.toHaveBeenCalled()
   })
 
-  // The tail is parked per-PTY (orca-runtime osc7ScanTailByPtyId) and per
-  // daemon scanner until the next chunk, so an attached slice pins one whole
-  // source chunk per tracked PTY.
+  // Production persists one scan tail per PTY and daemon scanner.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind a carried OSC scan tail', () => {
@@ -46,7 +44,6 @@ describe('OSC-7 URI extraction', () => {
     const before = process.memoryUsage().heapUsed
     const carried = new Map<string, string>()
     for (let index = 0; index < ptys; index += 1) {
-      // No terminator yet, so the tail is carried to the next chunk.
       const chunk = `${'x'.repeat(chunkChars)}\x1b]7;file://host/Users/dev/orca/wt-${index}`
       carried.set(`pty-${index}`, extractOscScanTail(chunk, 4096))
     }
@@ -54,7 +51,6 @@ describe('OSC-7 URI extraction', () => {
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(carried.get('pty-0')).toBe('\x1b]7;file://host/Users/dev/orca/wt-0')
-    // 8 MiB of source chunks stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/main/daemon/osc7-uri-extraction.ts

@@ -67,10 +67,7 @@ export function extractLastOsc7Uri(data: string): string | null {
   return lastUri
 }
 
-// Callers park this tail until the next chunk arrives — osc7ScanTailByPtyId in
-// orca-runtime.ts, TerminalOscCwdTitleScanner.scanTail in the daemon — so the
-// returned slice is detached; otherwise each carried tail pins a whole source
-// chunk for as long as its PTY or scanner stays alive.
+// Persisted scan tails must not retain their source PTY chunks.
 export function extractOscScanTail(input: string, limit: number): string {
   const lastOsc = input.lastIndexOf('\x1b]')
   const lastEscape = input.endsWith('\x1b') ? input.length - 1 : -1

src/main/daemon/terminal-mouse-mode-mirror.test.ts

@@ -34,8 +34,7 @@ describe('TerminalMouseModeMirror', () => {
     expect(mirror.sgrMousePixelsMode).toBe(false)
   })
 
-  // One mirror per daemon session holds the scan tail until the next chunk, so
-  // an attached slice pins a whole PTY chunk per live session.
+  // Production persists one scan tail per daemon session.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind a carried private-mode tail', () => {
@@ -46,18 +45,15 @@ describe('TerminalMouseModeMirror', () => {
     const before = process.memoryUsage().heapUsed
     const mirrors = Array.from({ length: sessions }, () => {
       const mirror = new TerminalMouseModeMirror()
-      // Ends mid-DECSET, so the params tail is carried into the next chunk.
       mirror.scan(`${'x'.repeat(chunkChars)}\x1b[?1002;1006;100`)
       return mirror
     })
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The carry must still complete the split DECSET across the boundary.
     mirrors[0]!.scan('3h')
     expect(mirrors[0]!.mouseTrackingMode).toBe('any')
     expect(mirrors[0]!.sgrMouseMode).toBe(true)
-    // 8 MiB of source chunks stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/main/daemon/terminal-mouse-mode-mirror.ts

@@ -93,8 +93,7 @@ export class TerminalMouseModeMirror {
     }
   }
 
-  // Detached: one mirror per daemon session holds this tail until the next
-  // chunk, so an attached slice pins a whole PTY chunk per live session.
+  // Persisted per-session tails must not retain their source PTY chunks.
   private extractScanTail(input: string): string {
     const start = Math.max(input.lastIndexOf('\x1b'), input.lastIndexOf('\x9b'))
     if (start === -1) {

src/main/runtime/orca-runtime.ts

@@ -11,6 +11,7 @@ import {
 } from '../../shared/agent-detection'
 import { detachString } from '../../shared/detached-string'
 import { extractOscTitleScanTail } from '../../shared/osc-title-scan-tail'
+import { retainTerminalPendingAnsi } from './terminal-pending-ansi'
 import { isServerDriveListRequest, listWindowsDrives } from './windows-drive-listing'
 import { extractLastOsc7Uri, extractOscScanTail } from '../daemon/osc7-uri-extraction'
 import { parseFileUriPathParts } from '../daemon/osc7-file-uri'
@@ -8927,8 +8928,7 @@ export class OrcaRuntimeService {
     }
     const appendedLower = appendedText.toLowerCase()
     const keywordHit = WAIT_BLOCKED_KEYWORD_PATTERN.test(`${state.keywordCarry}${appendedLower}`)
-    // Detached: keywordCarry is parked per PTY until the next chunk, so an
-    // attached slice would pin the whole lowercased chunk per tracked PTY.
+    // Persisted keyword carries must not retain their source PTY chunks.
     state.keywordCarry = detachString(appendedLower.slice(-WAIT_BLOCKED_KEYWORD_CARRY_CHARS))
     // Why the cap keeps the tail: the accumulated text only anchors boundary-
     // spanning prompt detection; anything past the tail cap has scrolled out
@@ -32477,7 +32477,6 @@ const WAIT_BLOCKED_KEYWORD_CARRY_CHARS = 31
 const MAX_TAIL_LINES = 2000
 const MAX_TAIL_CHARS = 256 * 1024
 const MAX_TAIL_PARTIAL_CHARS = 4000
-const MAX_TAIL_PENDING_ANSI_CHARS = 4096
 const DEFAULT_TERMINAL_READ_LIMIT = 120
 const MAX_TERMINAL_READ_LIMIT = 2000
 const MAX_TERMINAL_PREVIEW_CHARS = 32 * 1024
@@ -34693,13 +34692,16 @@ function normalizeTerminalChunk(
     if (char === '\x1b') {
       appendTerminalNormalizedSpan(parts, combined, textStart, index)
       if (index + 1 >= combined.length) {
-        return { text: parts.join(''), pendingAnsi: combined.slice(index) }
+        return {
+          text: parts.join(''),
+          pendingAnsi: retainTerminalPendingAnsi(combined.slice(index))
+        }
       }
       const parsed = parseAnsiControlSequence(combined, index)
       if (!parsed) {
         return {
           text: parts.join(''),
-          pendingAnsi: trimPendingAnsiControl(combined.slice(index))
+          pendingAnsi: retainTerminalPendingAnsi(combined.slice(index))
         }
       }
       if (parsed.kind === 'csi' && isTerminalPreviewLineControl(parsed)) {
@@ -34763,15 +34765,6 @@ function terminalChunkNeedsNormalization(chunk: string): boolean {
   return false
 }
 
-function trimPendingAnsiControl(value: string): string {
-  if (value.length <= MAX_TAIL_PENDING_ANSI_CHARS) {
-    return value
-  }
-  const introducer = value.slice(0, Math.min(2, value.length))
-  const suffixBudget = Math.max(0, MAX_TAIL_PENDING_ANSI_CHARS - introducer.length)
-  return `${introducer}${value.slice(-suffixBudget)}`
-}
-
 function isTerminalPreviewLineControl(parsed: {
   final: string
   params: string

src/main/runtime/orchestration/setup-completion-signal.test.ts

@@ -59,8 +59,7 @@ describe('orchestration setup completion signal', () => {
   })
 })
 
-// The carry lives in the scanner closure — one per running setup — until the
-// next chunk, so an attached slice pins a whole setup-output chunk each.
+// Production persists one carry per running setup.
 describe('setup completion scanner retention', () => {
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
@@ -72,22 +71,18 @@ describe('setup completion scanner retention', () => {
     const before = process.memoryUsage().heapUsed
     const scanners = Array.from({ length: setups }, (_unused, index) => {
       const scanner = createSetupCompletionScanner(`token-${index}`, () => undefined)
-      // A distinct chunk per setup. Sharing one would leave a single parent
-      // alive and pass whether or not the carry is detached.
       scanner.scan(`${'x'.repeat(chunkChars)}setup-${index}`)
       return scanner
     })
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The carry must still complete a marker split across the chunk boundary.
     const exitCodes: number[] = []
     const split = createSetupCompletionScanner('tok', (code) => exitCodes.push(code))
     split.scan(`${'x'.repeat(chunkChars)}__ORCA_SETUP_COMPLETE__:tok`)
     split.scan(':7\n')
     expect(exitCodes).toEqual([7])
     expect(scanners).toHaveLength(setups)
-    // 8 MiB of source chunks stay alive if the carries are attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/main/runtime/orchestration/setup-completion-signal.ts

@@ -67,8 +67,7 @@ export function createSetupCompletionScanner(
           return
         }
       }
-      // Detached: the carry is held per running setup until its next chunk, so
-      // an attached slice would pin a whole setup-output chunk each.
+      // Persisted setup carries must not retain their source output chunks.
       carry = detachString(combined.slice(-SETUP_COMPLETION_CARRY_LENGTH))
     }
   }

src/main/runtime/terminal-pending-ansi.test.ts

@@ -0,0 +1,26 @@
+import { describe, expect, it } from 'vitest'
+import { createForceGc, resolveForcedGc } from '../../shared/forced-gc-for-retention-tests'
+import { retainTerminalPendingAnsi } from './terminal-pending-ansi'
+
+describe('retainTerminalPendingAnsi', () => {
+  it('bounds an incomplete control while preserving its introducer and suffix', () => {
+    expect(retainTerminalPendingAnsi(`\x1b]${'x'.repeat(5000)}`)).toBe(`\x1b]${'x'.repeat(4094)}`)
+  })
+
+  const forcedGc = resolveForcedGc()
+  const itWithGc = forcedGc ? it : it.skip
+  itWithGc('does not pin PTY chunks behind pending ANSI state', () => {
+    const forceGc = createForceGc(forcedGc!)
+    forceGc()
+    const before = process.memoryUsage().heapUsed
+    const held = Array.from({ length: 512 }, (_unused, index) => {
+      const chunk = `${'x'.repeat(16 * 1024)}\x1b]0;partial-${index}`
+      return retainTerminalPendingAnsi(chunk.slice(chunk.lastIndexOf('\x1b')))
+    })
+    forceGc()
+    const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
+
+    expect(held[0]).toBe('\x1b]0;partial-0')
+    expect(retainedMiB).toBeLessThan(2)
+  })
+})

src/main/runtime/terminal-pending-ansi.ts

@@ -0,0 +1,12 @@
+import { detachString } from '../../shared/detached-string'
+
+const MAX_TERMINAL_PENDING_ANSI_CHARS = 4096
+
+export function retainTerminalPendingAnsi(value: string): string {
+  if (value.length <= MAX_TERMINAL_PENDING_ANSI_CHARS) {
+    return detachString(value)
+  }
+  const introducer = value.slice(0, Math.min(2, value.length))
+  const suffixBudget = MAX_TERMINAL_PENDING_ANSI_CHARS - introducer.length
+  return detachString(`${introducer}${value.slice(-suffixBudget)}`)
+}

src/main/stats/agent-detector.test.ts

@@ -243,8 +243,7 @@ describe('AgentDetector', () => {
     expect(stats.onAgentStop).toHaveBeenCalledWith('pty-1', 120)
   })
 
-  // meaningfulContentScanTailByPtyId parks the tail until the next chunk, so
-  // an attached slice pins one whole source chunk per tracked PTY.
+  // Production persists one meaningful-content tail per PTY.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin source chunks behind carried per-PTY scan tails', () => {
@@ -260,14 +259,12 @@ describe('AgentDetector', () => {
     forceGc()
     const before = process.memoryUsage().heapUsed
     for (let index = 0; index < ptys; index += 1) {
-      // Ends mid-CSI, so the tail is carried into the next chunk.
       detector.onData(`pty-${index}`, `${'y'.repeat(chunkChars)}\x1b[38;2;12;34;5`, 200)
     }
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(detector.trackedPtyCount).toBe(ptys)
-    // 8 MiB of source chunks stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/main/stats/agent-detector.ts

@@ -286,8 +286,7 @@ function isStTerminatedStringControlIntroducer(introducer: string): boolean {
   return introducer === 'P' || introducer === 'X' || introducer === '^' || introducer === '_'
 }
 
-// Detached because meaningfulContentScanTailByPtyId parks this until the next
-// chunk: an attached slice pins a whole source chunk per tracked PTY.
+// Persisted per-PTY tails must not retain their source chunks.
 function trimMeaningfulContentScanTail(value: string): string {
   if (value.length <= MEANINGFUL_CONTENT_SCAN_TAIL_LIMIT) {
     return detachString(value)

src/renderer/src/components/terminal-pane/pty-connection.ts

@@ -21,10 +21,10 @@ import { createTerminalZeroDimensionsMessage } from '../../../../shared/terminal
 import { isWorktreeRemovalFenceError } from '../../../../shared/worktree-removal-fence-error'
 import { parseTerminalOscColorQuery } from '../../../../shared/terminal-osc-color-reply'
 import {
-  HIDDEN_STARTUP_RENDERER_QUERY_PENDING_CHARS,
   containsCsiRendererQuery,
   containsStatefulRendererQuery,
   extractHiddenStartupRendererQueryData,
+  extractHiddenStartupRendererQueryPending,
   findCsiFinalByteIndex,
   isStatefulRendererReplyCsiQuery,
   isStatelessRendererReplyCsiQuery
@@ -5224,8 +5224,7 @@ export function connectPanePty(
           return ''
         }
       }
-      // Detached: this tail is parked per pane until the next chunk, so an
-      // attached slice would pin a whole PTY chunk for the pane's lifetime.
+      // Persisted per-pane tails must not retain their source PTY chunks.
       return detachString(tail.slice(-TERMINAL_RENDERER_RISK_SCAN_TAIL_CHARS))
     }
 
@@ -6260,7 +6259,7 @@ export function connectPanePty(
       if (input.startsWith('\x1b[')) {
         const finalByteIndex = findCsiFinalByteIndex(input, 2)
         if (finalByteIndex === -1) {
-          nextPending = input.slice(0, HIDDEN_STARTUP_RENDERER_QUERY_PENDING_CHARS)
+          nextPending = extractHiddenStartupRendererQueryPending(input, 0)
           consumedInputChars = input.length
         } else {
           const sequence = input.slice(0, finalByteIndex + 1)
@@ -6274,7 +6273,7 @@ export function connectPanePty(
       } else if (input.startsWith('\x1b]')) {
         const query = parseTerminalOscColorQuery(input, 0)
         if (query.kind === 'partial') {
-          nextPending = input.slice(0, HIDDEN_STARTUP_RENDERER_QUERY_PENDING_CHARS)
+          nextPending = extractHiddenStartupRendererQueryPending(input, 0)
           consumedInputChars = input.length
         } else if (query.kind === 'match') {
           oscColorQueryData = input.slice(0, query.endIndex)

src/renderer/src/components/terminal-pane/pty-title-retention.test.ts

@@ -2,28 +2,12 @@ import { describe, expect, it } from 'vitest'
 import { createForceGc, resolveForcedGc } from '../../../../shared/forced-gc-for-retention-tests'
 import { createPtyOutputProcessor } from './pty-transport'
 
-// What this proves — and what it does not.
-//
-// Invariant under test: the title a pane hands the store (setRuntimePaneTitle /
-// updateTabTitle) is a standalone string, not a V8 SlicedString over the PTY
-// frame it was parsed from. Frames are sized at the production renderer-write
-// bound (PTY_BATCH_FLUSH_CHUNK_CHARS = 16 KiB, src/main/ipc/pty.ts), not the
-// multi-MiB frames an earlier draft of this test assumed.
-//
-// Panes are held live simultaneously with no store overwrite, which is the
-// unbounded-in-principle shape. Production writes `[paneId]: title`, an
-// overwrite that releases the previous parent, so the real live set is
-// O(panes). This is a regression guard on the retention property, NOT a
-// reproduction of a shipped out-of-memory crash.
-//
+// This guards 16 KiB frame detachment, not production's O(panes) overwrite model.
 const MIB = 1024 * 1024
 const CHUNK_CHARS = 16 * 1024
 const PANES = 2048
-// '█' is 2 bytes/char in V8, so a pinned frame costs twice its char count.
 const PINNED_FRAME_MIB = (CHUNK_CHARS * PANES * 2) / MIB
 
-// resolveForcedGc works without --expose-gc, so this runs in CI's plain
-// `vitest run`. Skips only if a hardened host refuses to expose gc at all.
 const forcedGc = resolveForcedGc()
 const describeWithGc = forcedGc ? describe : describe.skip
 const forceGc = forcedGc ? createForceGc(forcedGc) : (): void => undefined
@@ -36,7 +20,6 @@ describeWithGc('renderer PTY title retention', () => {
   it('does not pin each pane PTY frame behind the title handed to the store', async () => {
     forceGc()
     const before = process.memoryUsage().heapUsed
-    // Stands in for store.runtimePaneTitlesByTabId / tab titles: one live title per pane.
     const storedTitles: string[] = []
 
     for (let pane = 0; pane < PANES; pane += 1) {
@@ -46,7 +29,6 @@ describeWithGc('renderer PTY title retention', () => {
         }
       })
       processor.processData(makeAgentFrame(pane), { onData: () => undefined })
-      // Drain the deferred side-effect queue the way the real drain timer does.
       await new Promise((resolve) => setTimeout(resolve, 0))
     }
 
@@ -54,7 +36,6 @@ describeWithGc('renderer PTY title retention', () => {
     const retainedMiB = (process.memoryUsage().heapUsed - before) / MIB
     const titleChars = storedTitles.reduce((total, title) => total + title.length, 0)
 
-    // eslint-disable-next-line no-console
     console.log(
       `${storedTitles.length} stored pane titles (${titleChars} chars) -> heap delta ${retainedMiB.toFixed(2)} MiB; ` +
         `source frames total ${PINNED_FRAME_MIB} MiB if every title pinned its parent`
@@ -62,8 +43,6 @@ describeWithGc('renderer PTY title retention', () => {
 
     expect(storedTitles).toHaveLength(PANES)
     expect(storedTitles[0]).toBe('✳ Working… (esc to interrupt) pane-0')
-    // Titles hold well under 1 MiB of text. A delta near the frame total means
-    // the slices still point at their parents.
     expect(retainedMiB).toBeLessThan(PINNED_FRAME_MIB / 8)
   })
 })

src/renderer/src/components/terminal-pane/terminal-bracketed-paste.test.ts

@@ -5,10 +5,7 @@ import {
   pasteTerminalText,
   sanitizeTerminalPasteText
 } from './terminal-bracketed-paste'
-import {
-  createForceGc,
-  resolveForcedGc
-} from '../../../../shared/forced-gc-for-retention-tests'
+import { createForceGc, resolveForcedGc } from '../../../../shared/forced-gc-for-retention-tests'
 
 function createTerminal(bracketedPasteMode = true) {
   const terminal = {
@@ -257,8 +254,7 @@ describe('terminal bracketed paste policy', () => {
     expect(splitCallCount).toBe(0)
   })
 
-  // The tail sits in a per-terminal WeakMap until the pane sees a 2004 toggle,
-  // so an attached slice pins a whole PTY chunk per interrupted pane.
+  // Production persists one interrupted-mode tail per terminal.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind the interrupted-pane mode tail', () => {
@@ -266,8 +262,7 @@ describe('terminal bracketed paste policy', () => {
     const panes = 512
     const chunkChars = 16 * 1024
 
-    // Plain stubs, not createTerminal: a vi.fn() per pane costs megabytes of
-    // mock bookkeeping and would swamp the retention signal being measured.
+    // Avoid mock bookkeeping that would swamp the retention signal.
     const terminals = Array.from({ length: panes }, () => ({
       modes: { bracketedPasteMode: true }
     }))
@@ -276,14 +271,12 @@ describe('terminal bracketed paste policy', () => {
     const before = process.memoryUsage().heapUsed
     for (const [index, terminal] of terminals.entries()) {
       markTerminalBracketedPasteInterrupted(terminal)
-      // No 2004 sequence, so the tail is carried rather than cleared.
       observeTerminalBracketedPasteModeOutput(terminal, `${'x'.repeat(chunkChars)}pane-${index}`)
     }
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(terminals).toHaveLength(panes)
-    // ~8 MiB of source chunks stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/renderer/src/components/terminal-pane/terminal-bracketed-paste.ts

@@ -99,8 +99,7 @@ export function observeTerminalBracketedPasteModeOutput(
     return
   }
   const combined = (bracketedPasteModeOutputTail.get(terminal) ?? '') + data
-  // Detached: this tail is parked in a per-terminal WeakMap until the pane sees
-  // a 2004 toggle, so an attached slice pins a whole PTY chunk per interrupted pane.
+  // Persisted per-terminal tails must not retain their source PTY chunks.
   bracketedPasteModeOutputTail.set(
     terminal,
     detachString(combined.slice(-BRACKETED_PASTE_MODE_TAIL_MAX))

src/renderer/src/lib/pane-manager/terminal-complex-script.test.ts

@@ -406,13 +406,7 @@ describe('nativeWindowsRewriteNeedsFollowupRenderRefresh', () => {
   })
 })
 
-// The rewrite-CSI scan tail is parked in per-pane closure state
-// (foreground/hiddenRewriteCsiScanTail in pty-connection.ts) until the next
-// chunk, so an attached slice pins a whole PTY chunk per open pane.
-//
-// Why the split CSI is long: V8 only builds a SlicedString past
-// SlicedString::kMinLength (13 chars); a shorter tail is already copied flat
-// and would make this assertion pass with or without the detach.
+// The ≥13-char control arm proves detachment for one persisted rewrite tail per pane.
 describe('rewrite CSI scan tail retention', () => {
   const splitCsi = '\x1b[1;2;3;4;5;6;7;8;9'
   const forcedGc = resolveForcedGc()
@@ -427,8 +421,6 @@ describe('rewrite CSI scan tail retention', () => {
     const tails = Array.from(
       { length: panes },
       (_unused, index) =>
-        // A distinct chunk per pane, each ending mid-CSI. Sharing one chunk
-        // would leave a single parent alive and pass either way.
         terminalRewriteOutputRenderRefreshDecision(
           `${'x'.repeat(chunkChars)}pane-${index}${splitCsi}`,
           { previousChunkEndsWithCarriageReturn: false, previousRewriteCsiScanTail: '' }
@@ -437,7 +429,6 @@ describe('rewrite CSI scan tail retention', () => {
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The tail must still complete the erase sequence across the boundary.
     expect(tails[0]).toBe(splitCsi)
     expect(
       terminalRewriteOutputRenderRefreshDecision('K', {
@@ -445,7 +436,6 @@ describe('rewrite CSI scan tail retention', () => {
         previousRewriteCsiScanTail: tails[0]!
       }).prefersRenderRefresh
     ).toBe(true)
-    // 8 MiB of source chunks stay alive if the carried tails are attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/renderer/src/lib/pane-manager/terminal-complex-script.ts

@@ -141,9 +141,7 @@ function containsRewriteEraseSequence(data: string): boolean {
   return false
 }
 
-// Detached: callers park this tail in per-pane closure state (foreground/hidden
-// RewriteCsiScanTail in pty-connection.ts) until the next chunk, so an attached
-// slice pins a whole PTY chunk per open pane.
+// Persisted per-pane rewrite tails must not retain their source PTY chunks.
 function trailingIncompleteRewriteCsiTail(data: string): string {
   const escapeIndex = data.lastIndexOf(ESCAPE_CHARACTER)
   if (escapeIndex === -1) {

src/shared/agent-status-field-normalization.ts

@@ -190,11 +190,7 @@ export function normalizeInteractivePromptField(
   if (typeof value !== 'string' || value.length === 0) {
     return undefined
   }
-  // Why detach: this field is cached in the store, so an over-cap value's
-  // truncated slice would otherwise pin the whole original string.
-  // Why only when truncation happened: an under-cap value is returned by
-  // identity, is already flat (JSON.parse / JSON.stringify output), and pins
-  // nothing — detaching it would copy the payload on every hook event.
+  // Only truncated values can retain oversized payloads; preserve hot-path identity otherwise.
   const truncated = truncatePreservingSurrogates(value, maxLength)
   if (truncated.length === 0) {
     return undefined

src/shared/agent-status-osc.test.ts

@@ -35,9 +35,7 @@ describe('createAgentStatusOscProcessor', () => {
     ])
   })
 
-  // The unterminated-payload carry lives in the processor closure until the
-  // next chunk, and there is one processor per pane, so an attached slice
-  // would pin a whole PTY chunk per pane.
+  // Production persists one unterminated payload per pane.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind a carried OSC 9999 payload', () => {
@@ -48,14 +46,12 @@ describe('createAgentStatusOscProcessor', () => {
     const before = process.memoryUsage().heapUsed
     const processors = Array.from({ length: panes }, (_unused, index) => {
       const processChunk = createAgentStatusOscProcessor()
-      // No terminator, so the payload is carried into the next chunk.
       processChunk(`${'x'.repeat(chunkChars)}\x1b]9999;{"state":"working","prompt":"p-${index}"`)
       return processChunk
     })
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The carry must still work: 8 MiB of chunks stay alive if it is attached.
     expect(processors[0]!('}\x07').payloads[0]).toEqual({ state: 'working', prompt: 'p-0' })
     expect(retainedMiB).toBeLessThan(2)
   })

src/shared/agent-status-osc.ts

@@ -71,8 +71,7 @@ export function createAgentStatusOscProcessor(): (data: string) => ProcessedAgen
 
       if (terminator === null) {
         const candidate = combined.slice(start)
-        // Detached: `pending` survives until the next chunk, so an attached
-        // slice would pin this whole chunk per processor (one per pane).
+        // Persisted per-pane payloads must not retain their source PTY chunks.
         pending = candidate.length > MAX_PENDING ? '' : detachString(candidate)
         break
       }

src/shared/agent-status-types.test.ts

@@ -242,8 +242,7 @@ Fix dispatch fallback preview for normalized status prompts`
     expect(AGENT_STATUS_INTERACTIVE_PROMPT_MAX_LENGTH).toBe(16000)
   })
 
-  // The capped value is cached in the store; if truncation returned a raw
-  // slice it would keep the whole oversized payload alive.
+  // The store caches capped interactive prompts.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the oversized payload behind a capped interactivePrompt', () => {
@@ -262,13 +261,10 @@ Fix dispatch fallback preview for normalized status prompts`
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(held[0]!.interactivePrompt).toHaveLength(AGENT_STATUS_INTERACTIVE_PROMPT_MAX_LENGTH)
-    // 128 MiB of source payloads stay alive if the capped slice is still attached.
     expect(retainedMiB).toBeLessThan(16)
   })
 
-  // An under-cap prompt pins nothing (it is the caller's own flat string), so
-  // it must be passed through by identity rather than copied on every hook
-  // event — this path fires many times per second during a tool-use run.
+  // Under-cap prompts stay on the allocation-free hook path.
   itWithGc('returns an under-cap interactivePrompt without copying it', () => {
     const source = JSON.stringify({
       questions: [{ question: 'Which approach?', options: ['a'.repeat(4000), 'b'.repeat(4000)] }]
@@ -289,7 +285,6 @@ Fix dispatch fallback preview for normalized status prompts`
     const bytesPerCall = (process.memoryUsage().heapUsed - before) / repeats
 
     expect(held[0]).toBe(source)
-    // A copy would cost ~source.length bytes per call; identity costs a pointer.
     expect(bytesPerCall).toBeLessThan(source.length / 4)
   })
 

src/shared/codex-subagent-transcript.test.ts

@@ -157,10 +157,7 @@ describe('Codex subagent transcript reconciliation', () => {
     expect(roster.size).toBe(0)
   })
 
-  // The trailing partial line is parked per cursor in
-  // codexSubagentTranscriptByPaneKey until the next poll, and `split` yields
-  // slices of the up-to-1 MiB read buffer, so an attached carry pins a whole
-  // buffer per tracked pane.
+  // Production persists one partial line per transcript cursor.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the read buffer behind a carried partial line', () => {
@@ -172,7 +169,6 @@ describe('Codex subagent transcript reconciliation', () => {
 
     const states = Array.from({ length: panes }, (_unused, index) => {
       const parentPath = join(dir, `rollout-parent-${index}.jsonl`)
-      // Trailing partial line (no newline) is what gets carried.
       writeFileSync(parentPath, `${jsonl([activity('started')])}${bulkLine}\n{"partial":${index}`)
       return { state: createCodexSubagentTranscriptState(), parentPath }
     })
@@ -186,7 +182,6 @@ describe('Codex subagent transcript reconciliation', () => {
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(states[0]!.state.parent.carry).toBe('{"partial":0')
-    // ~64 MiB of read buffers stay alive if the carries are still attached.
     expect(retainedMiB).toBeLessThan(8)
   })
 })

src/shared/codex-subagent-transcript.ts

@@ -78,9 +78,7 @@ function readJsonlCursor(cursor: JsonlCursor): JsonRecord[] | undefined {
   const content = `${skippedPrefix ? '' : cursor.carry}${buffer.toString('utf8', 0, bytesRead)}`
   const lines = content.split('\n')
   cursor.offset = start + bytesRead
-  // Detached: `split` yields slices of the up-to-1 MiB read buffer, and this
-  // carry is parked per cursor (one per pane, plus one per tracked subagent, in
-  // codexSubagentTranscriptByPaneKey) until the next poll.
+  // Persisted cursor carries must not retain the transcript read buffer.
   cursor.carry = detachString(lines.pop() ?? '')
   if (skippedPrefix) {
     lines.shift()

src/shared/command-code-output-status.test.ts

@@ -464,8 +464,7 @@ describe('terminal control stripping', () => {
     }
   })
 
-  // Every pane runs this detector (the ring is written before the Command Code
-  // early-out), so an attached ring pins one whole PTY chunk per open pane.
+  // Every pane persists this ring before Command Code detection.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind the recent raw-text ring', () => {
@@ -484,7 +483,6 @@ describe('terminal control stripping', () => {
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(detectors).toHaveLength(panes)
-    // ~8 MiB of source chunks stay alive if the rings are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/command-code-output-status.ts

@@ -137,10 +137,7 @@ function rawTextMayContainCommandCodeBanner(rawText: string): boolean {
   return rawText.includes('C') && rawText.includes('o') && rawText.includes('d')
 }
 
-// Detached: every pane parks this ring until its next chunk (the detector runs
-// before the Command Code early-out), so an attached slice pins a whole PTY
-// chunk per pane. The empty-prefix case matters too: V8 returns `data` itself
-// from `'' + data`, so that branch slices the raw chunk as well.
+// Every pane persists this ring before Command Code detection.
 function appendRecentRawText(previousRawText: string, data: string): string {
   if (data.length >= RECENT_TEXT_LIMIT) {
     return detachString(data.slice(-RECENT_TEXT_LIMIT))

src/shared/detached-string.test.ts

@@ -20,21 +20,13 @@ describe('detachString', () => {
     expect(detachString(value)).toBe(value)
   })
 
-  // Why the control arm: V8 only builds a SlicedString when the result is at
-  // least SlicedString::kMinLength (13) chars; shorter slices are already
-  // copied flat. A retention assertion over a short result therefore passes
-  // with or without detachString and proves nothing. Measuring the raw slice
-  // in the same run makes the test fail if that premise ever stops holding.
-  //
-  // resolveForcedGc works without --expose-gc, so this guard runs under the
-  // plain `vitest run` CI uses instead of silently skipping.
+  // A ≥13-char raw control arm proves both SlicedString retention and default-run GC.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('detaches a slice from its parent buffer', () => {
     const chunkChars = 16 * 1024
     const slices = 4096
     const forceGc = createForceGc(forcedGc!)
-    // Comfortably over kMinLength so the untouched slice really is a SlicedString.
     const measure = (detach: boolean): number => {
       const held: string[] = []
       forceGc()
@@ -52,7 +44,6 @@ describe('detachString', () => {
 
     const attachedMiB = measure(false)
     const detachedMiB = measure(true)
-    // ~128 MiB of parents stay alive when the slices are left attached.
     expect(attachedMiB).toBeGreaterThan(16)
     expect(detachedMiB).toBeLessThan(attachedMiB / 8)
   })

src/shared/detached-string.ts

@@ -1,15 +1,4 @@
-/**
- * Return a copy of `value` that shares no backing store with the string it was
- * sliced from.
- *
- * Why: `String.prototype.slice` on a long string produces a V8 SlicedString
- * that keeps a pointer to its parent, so a short slice retained in long-lived
- * state (a store entry, a cached field) pins the entire source buffer — a raw
- * PTY chunk or a hook JSON payload — for as long as the slice is reachable.
- * Concatenating forces V8 to flatten into a fresh flat string, dropping the
- * parent pointer. The leading space is stripped back off so the value is
- * unchanged.
- */
+/** Break V8 SlicedString parent retention by forcing a fresh flat string. */
 export function detachString(value: string): string {
   if (value.length === 0) {
     return ''

src/shared/draft-paste-ready-scanner.test.ts

@@ -138,9 +138,7 @@ describe('createDraftPasteReadyScanner', () => {
   })
 })
 
-// The 512-char rings live in the scanner closure — one scanner per pane
-// waiting to deliver a startup draft — so an attached slice pins a whole PTY
-// chunk per waiter instead of the 512 chars the ring documents.
+// Production persists two 512-char rings per startup-draft waiter.
 describe('draft paste scanner retention', () => {
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
@@ -152,17 +150,13 @@ describe('draft paste scanner retention', () => {
     const before = process.memoryUsage().heapUsed
     const scanners = Array.from({ length: waiters }, (_unused, index) => {
       const scanner = createDraftPasteReadyScanner('codex-composer-prompt')
-      // A distinct chunk per waiter. Sharing one would leave a single parent
-      // alive and pass whether or not the rings are detached.
       scanner.observe(`${DECSET_BRACKETED_PASTE}${'x'.repeat(chunkChars)}pane-${index}`)
       return scanner
     })
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The rings must still detect the prompt glyph on the next chunk.
     expect(scanners[0]!.observe(CODEX_PROMPT)).toEqual({ ready: true, armQuietTimer: false })
-    // 8 MiB of source chunks stay alive if the rings are attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/draft-paste-ready-scanner.ts

@@ -62,9 +62,7 @@ export function createDraftPasteReadyScanner(readySignal: DraftPasteReadySignal)
   return {
     observe(data: string): DraftPasteReadyScanResult {
       const combined = recent + data
-      // Detached: both rings survive until the next chunk (and past settle, until
-      // the scanner itself is dropped), so an attached slice would pin a whole
-      // PTY chunk per waiting pane instead of the 512 chars the ring promises.
+      // Persisted scanner rings must not retain their source PTY chunks.
       recent = detachString(combined.slice(-RING_CHARS))
       if (!saw2004) {
         const markerIndex = combined.indexOf(DECSET_BRACKETED_PASTE)

src/shared/forced-gc-for-retention-tests.ts

@@ -1,18 +1,7 @@
 import { setFlagsFromString } from 'node:v8'
 import { runInNewContext } from 'node:vm'
 
-/**
- * Resolve a real forced-GC function for retention tests, without requiring the
- * runner to pass `--expose-gc`.
- *
- * Why: CI runs plain `pnpm exec vitest run` (.github/workflows/pr.yml), so
- * `globalThis.gc` is undefined and every `describe.skip`-on-missing-gc
- * retention guard silently stops running — a reintroduced raw slice would land
- * green. `v8.setFlagsFromString('--expose-gc')` plus a throwaway VM context
- * hands back a real `gc` binding in the worker; the flag is reset immediately
- * so nothing else in the process observes it. Returns null only if the host
- * genuinely refuses (hardened runtimes), which keeps callers skippable.
- */
+/** Resolve forced GC under plain Vitest; hardened runtimes may return null. */
 export function resolveForcedGc(): (() => void) | null {
   const existing = (globalThis as { gc?: () => void }).gc
   if (typeof existing === 'function') {
@@ -28,7 +17,7 @@ export function resolveForcedGc(): (() => void) | null {
   }
 }
 
-/** Double-collect so a freshly unreachable parent is actually released. */
+/** Collect twice so freshly unreachable parents are released. */
 export function createForceGc(gc: () => void): () => void {
   return () => {
     gc()

src/shared/osc-title-extraction.retention.test.ts

@@ -2,31 +2,17 @@ import { describe, expect, it } from 'vitest'
 import { createForceGc, resolveForcedGc } from './forced-gc-for-retention-tests'
 import { extractAllOscTitles } from './osc-title-extraction'
 
-// What this proves — and what it does not.
-//
-// Invariant under test: an extracted OSC title is a standalone string, not a V8
-// SlicedString pinning the PTY chunk it was cut from. Chunks are sized at the
-// production renderer-write bound (PTY_BATCH_FLUSH_CHUNK_CHARS = 16 KiB, see
-// src/main/ipc/pty.ts), not the multi-MiB chunks an earlier draft assumed.
-//
-// This is the unbounded-in-principle case: N titles from N distinct chunks all
-// held live at once. Production's per-pane store write is an overwrite, so the
-// live set there is O(panes) and small. So this is a regression guard on the
-// retention property, NOT a reproduction of a shipped out-of-memory crash.
-//
+// This guards 16 KiB chunk detachment, not a production OOM reproduction.
 const MIB = 1024 * 1024
 const CHUNK_CHARS = 16 * 1024
 const RETAINED_TITLES = 4096
 const PINNED_CHUNK_MIB = (CHUNK_CHARS * RETAINED_TITLES) / MIB
 
-// resolveForcedGc works without --expose-gc, so this runs in CI's plain
-// `vitest run`. Skips only if a hardened host refuses to expose gc at all.
 const forcedGc = resolveForcedGc()
 const describeWithGc = forcedGc ? describe : describe.skip
 const forceGc = forcedGc ? createForceGc(forcedGc) : (): void => undefined
 
 function makePtyChunk(index: number): string {
-  // A realistic dense agent frame: bulk redraw plus one OSC 0 working title.
   return `${'x'.repeat(CHUNK_CHARS)}\x1b]0;✳ Working… (esc to interrupt) ${index}\x07`
 }
 
@@ -36,7 +22,6 @@ describeWithGc('OSC title retention', () => {
     const before = process.memoryUsage().heapUsed
     const retained: string[] = []
     for (let index = 0; index < RETAINED_TITLES; index += 1) {
-      // The chunk itself goes out of scope every iteration; only the title survives.
       const titles = extractAllOscTitles(makePtyChunk(index))
       retained.push(titles.at(-1) as string)
     }
@@ -44,7 +29,6 @@ describeWithGc('OSC title retention', () => {
     const retainedMiB = (process.memoryUsage().heapUsed - before) / MIB
     const titleChars = retained.reduce((total, title) => total + title.length, 0)
 
-    // eslint-disable-next-line no-console
     console.log(
       `retained ${retained.length} titles (${titleChars} chars total) -> heap delta ${retainedMiB.toFixed(2)} MiB; ` +
         `source chunks total ${PINNED_CHUNK_MIB} MiB if every title pinned its parent`
@@ -52,8 +36,6 @@ describeWithGc('OSC title retention', () => {
 
     expect(retained).toHaveLength(RETAINED_TITLES)
     expect(retained[0]).toBe('✳ Working… (esc to interrupt) 0')
-    // The titles hold well under 1 MiB of actual text. A delta anywhere near the
-    // source-chunk total means the slices still point at their parents.
     expect(retainedMiB).toBeLessThan(PINNED_CHUNK_MIB / 8)
   })
 })

src/shared/osc-title-extraction.ts

@@ -61,9 +61,7 @@ function parseOscTitleAt(data: string, index: number): OscTitleParseResult {
 function readBoundedOscTitle(data: string, titleStart: number, titleEnd: number): string {
   // Why: PTY output can contain pasted or remote-controlled OSC titles; keep
   // downstream title detection bounded while preserving trailing status words.
-  // Why detach here: titles outlive their chunk in pane/tab state, and a raw
-  // slice would pin the whole PTY chunk. Detaching at the extraction boundary
-  // means no consumer can bypass it.
+  // Extracted titles must not retain their source PTY chunks.
   const length = titleEnd - titleStart
   if (length <= MAX_OSC_TITLE_CHARS) {
     return detachString(data.slice(titleStart, titleEnd))

src/shared/osc-title-scan-tail.test.ts

@@ -16,16 +16,13 @@ describe('extractOscTitleScanTail', () => {
     expect(extractOscTitleScanTail('\x1b]133;D;0\x07\x1b')).toBe('\x1b')
   })
 
-  // A carried tail lives in a per-PTY map (agent-detector.ts,
-  // orca-runtime.ts) until the next chunk, so a raw slice would pin one whole
-  // source chunk per tracked PTY.
+  // Production persists one title scan tail per PTY.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind a carried tail', () => {
     const chunkChars = 16 * 1024
     const ptys = 512
     const forceGc = createForceGc(forcedGc!)
-    // A split title: no terminator yet, so the tail is carried to the next chunk.
     const makeChunk = (index: number): string =>
       `${'x'.repeat(chunkChars)}\x1b]0;✳ Working… (esc to interrupt) pty-${index}`
 
@@ -40,7 +37,6 @@ describe('extractOscTitleScanTail', () => {
 
     expect(carried.size).toBe(ptys)
     expect(carried.get('pty-0')).toBe('\x1b]0;✳ Working… (esc to interrupt) pty-0')
-    // 16 MiB of source chunks stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/osc-title-scan-tail.ts

@@ -4,10 +4,7 @@ const OSC_TITLE_SCAN_TAIL_LIMIT = 4096
 const OSC_TITLE_PREFIX_LENGTH = 4
 const OSC_TITLE_CODES = new Set(['0', '1', '2'])
 
-// Callers park the returned tail in a per-PTY map until the next chunk
-// arrives. The non-empty tails are slices of the raw chunk, so they are
-// detached (see trimOscTitleScanTail) — otherwise each carried tail would pin
-// a whole source chunk for as long as its PTY stays tracked.
+// Persisted per-PTY tails must not retain their source chunks.
 export function extractOscTitleScanTail(input: string): string {
   const lastOsc = input.lastIndexOf('\x1b]')
   if (lastOsc !== -1) {

src/shared/terminal-color-scheme-protocol.test.ts

@@ -53,8 +53,7 @@ describe('terminal color scheme protocol', () => {
     })
   })
 
-  // The incomplete private-mode tail is carried in per-pane scan state until the
-  // next chunk, so an attached slice pins a whole PTY chunk per pane.
+  // Production persists one private-mode tail per pane.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind a carried private-mode tail', () => {
@@ -65,14 +64,12 @@ describe('terminal color scheme protocol', () => {
     const before = process.memoryUsage().heapUsed
     const tails = Array.from(
       { length: panes },
-      // No final byte yet, so the params tail is carried to the next chunk.
       () => scanMode2031Sequences('', `${'x'.repeat(chunkChars)}\x1b[?1049;2004;2026;12345`).tail
     )
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(tails[0]).toBe('\x1b[?1049;2004;2026;12345')
-    // 8 MiB of source chunks stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/terminal-color-scheme-protocol.ts

@@ -132,8 +132,7 @@ function hasMode2031(params: string): boolean {
   return params.split(';').some((param) => Number(param) === 2031)
 }
 
-// Detached: the retained tail is carried in per-pane scan state until the next
-// chunk, so an attached slice pins a whole PTY chunk per pane.
+// Persisted per-pane tails must not retain their source PTY chunks.
 function extractPrivateModeScanTail(input: string): string {
   const start = Math.max(input.lastIndexOf('\x1b'), input.lastIndexOf('\x9b'))
   if (start === -1) {

src/shared/terminal-github-pr-link-detector.test.ts

@@ -293,9 +293,7 @@ describe('createTerminalGitHubPRLinkDetector', () => {
     ])
   })
 
-  // The URL carry lives in the detector closure until the next chunk, and there
-  // is one detector per pane/PTY, so an attached slice pins a whole PTY chunk
-  // per pane for as long as the pane stays open.
+  // Production persists one URL carry per pane or PTY.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind a carried PR-URL prefix', () => {
@@ -306,14 +304,12 @@ describe('createTerminalGitHubPRLinkDetector', () => {
     const before = process.memoryUsage().heapUsed
     const detectors = Array.from({ length: panes }, (_unused, index) => {
       const observe = createTerminalGitHubPRLinkDetector()
-      // Ends mid-URL, so the prefix is carried into the next chunk.
       observe(`${'x'.repeat(chunkChars)}https://github.com/acme/repo-${index}/pul`)
       return observe
     })
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The carry must still complete the split URL across the chunk boundary.
     expect(detectors[0]!('l/12\n')).toEqual([
       {
         url: 'https://github.com/acme/repo-0/pull/12',
@@ -321,7 +317,6 @@ describe('createTerminalGitHubPRLinkDetector', () => {
         number: 12
       }
     ])
-    // 8 MiB of source chunks stay alive if the carries are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/terminal-github-pr-link-detector.ts

@@ -66,8 +66,7 @@ function lastIndexOfHttpScheme(value: string, fromIndex?: number): number {
   return lastIndex
 }
 
-// Detached: the carry survives in the detector closure until the next chunk, and
-// there is one detector per pane/PTY, so an attached slice pins a whole chunk each.
+// Long URL carries must not retain their source PTY chunk.
 function getPotentialGitHubPRCarry(value: string): string {
   // Why bounded: carry is always a suffix of at most MAX_CARRY_LENGTH, so a scheme
   // further back can only ever be dropped — scanning to it is O(chunk) per PTY write.
@@ -83,11 +82,11 @@ function getPotentialGitHubPRCarry(value: string): string {
 
   const fragment = endsWithHttpSchemePrefixFragment(tailWindow)
   if (fragment === '' || windowStart === 0) {
-    return detachString(fragment)
+    return fragment
   }
   // Why look behind: an older scheme means the URL already overran the cap, so the
   // carry is abandoned rather than restarted from this fragment.
-  return lastIndexOfHttpScheme(value, windowStart - 1) === -1 ? detachString(fragment) : ''
+  return lastIndexOfHttpScheme(value, windowStart - 1) === -1 ? fragment : ''
 }
 
 function hasTerminalUrlWhitespace(value: string, start: number, end: number): boolean {

src/shared/terminal-kitty-keyboard-mode-tracker.test.ts

@@ -149,12 +149,7 @@ describe('TerminalKittyKeyboardModeTracker', () => {
     expect(ranAndExited.flags).toBe(0)
   })
 
-  // One tracker per pane (renderer) and per PTY (main) holds the scan tail
-  // until the next chunk, so an attached slice pins a whole PTY chunk each.
-  //
-  // Why the tail is long: V8 only builds a SlicedString past
-  // SlicedString::kMinLength (13 chars); a shorter tail is already copied flat
-  // and would make this assertion pass with or without the detach.
+  // The ≥13-char control arm proves detachment for one persisted tail per tracker.
   const splitPrivateModeTail = '\x1b[?1049;2004;2026;1234'
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
@@ -167,24 +162,19 @@ describe('TerminalKittyKeyboardModeTracker', () => {
     const before = process.memoryUsage().heapUsed
     const trackers = Array.from({ length: panes }, (_unused, index) => {
       const tracker = new TerminalKittyKeyboardModeTracker()
-      // Ends mid-DECSET, so the params tail is carried into the next chunk.
       tracker.scan(`${'x'.repeat(chunkChars)}pty-${index}${splitPrivateModeTail}`)
       return tracker
     })
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The carry must still complete the split sequence across the boundary.
     trackers[0]!.scan('9h')
     expect(trackers[0]!.isAlternateScreen).toBe(true)
-    // 8 MiB of source chunks stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 
-  // scanReplay is fed whole reattach snapshots, not 16 KiB PTY chunks, so an
-  // attached tail pins far more per pane than the live path does.
+  // Replay consumes whole reattach snapshots rather than 16 KiB live chunks.
   itWithGc('does not pin a reattach snapshot behind a carried kitty scan tail', () => {
-    // A default 5000-row scrollback snapshot at ~120 chars/row.
     const snapshotChars = 5000 * 120
     const panes = 20
     const forceGc = createForceGc(forcedGc!)
@@ -200,7 +190,6 @@ describe('TerminalKittyKeyboardModeTracker', () => {
 
     trackers[0]!.scan('9h')
     expect(trackers[0]!.isAlternateScreen).toBe(true)
-    // ~11 MiB of snapshots stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/terminal-kitty-keyboard-mode-tracker.ts

@@ -171,10 +171,7 @@ export class TerminalKittyKeyboardModeTracker {
     }
   }
 
-  // Detached: this tracker is per-pane (renderer) and per-PTY (main), and the
-  // tail outlives its input until the next chunk. An attached slice would pin
-  // the whole input — a live PTY chunk, or a full reattach/replay snapshot on
-  // the scanReplay path — for as long as the pane stays open.
+  // Persisted tracker tails must not retain live chunks or replay snapshots.
   private extractScanTail(input: string): string {
     const start = Math.max(input.lastIndexOf('\x1b'), input.lastIndexOf('\x9b'))
     if (start === -1) {

src/shared/terminal-osc133-command-finished.test.ts

@@ -34,8 +34,7 @@ describe('createOsc133CommandFinishedScanner', () => {
     expect(onCommandFinished).not.toHaveBeenCalled()
   })
 
-  // The carry lives in the scanner closure until the next chunk, and there is
-  // one scanner per pane, so an attached slice pins a whole PTY chunk per pane.
+  // Production persists one unterminated payload per pane scanner.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the source chunk behind a carried OSC 133 payload', () => {
@@ -47,17 +46,14 @@ describe('createOsc133CommandFinishedScanner', () => {
     const before = process.memoryUsage().heapUsed
     const scanners = Array.from({ length: panes }, (_unused, index) => {
       const scanner = createOsc133CommandFinishedScanner(finished)
-      // No terminator yet, so the payload is carried into the next chunk.
       scanner.scan(`${'x'.repeat(chunkChars)}\x1b]133;D;${index} pending payload tail text`)
       return scanner
     })
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The carry must still resolve across the chunk boundary.
     scanners[0]!.scan('\x07')
     expect(finished).toHaveBeenCalledWith(0)
-    // 8 MiB of source chunks stay alive if the carries are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/terminal-osc133-command-finished.ts

@@ -90,8 +90,7 @@ export function createOsc133CommandFinishedScanner(
       const terminator = findOscTerminator(combined, payloadStart)
       if (!terminator) {
         const pending = combined.slice(start)
-        // Detached: the carry outlives this chunk in the scanner closure (one per
-        // pane), so an attached slice would pin the whole chunk until the next one.
+        // Persisted per-pane carries must not retain their source PTY chunks.
         carry = detachString(
           pending.length > MAX_OSC_CARRY_LENGTH
             ? pending.slice(pending.length - MAX_OSC_CARRY_LENGTH)

src/shared/terminal-partial-escape-tail.test.ts

@@ -85,9 +85,7 @@ describe('advancePartialEscapeTail', () => {
     expect(advancePartialEscapeTail('', huge)).toBe('')
   })
 
-  // The headless emulator parks this tail per session until the next write and
-  // ships it in every snapshot, so an attached slice pins a whole written
-  // stream per live session.
+  // Production persists and snapshots one partial escape tail per emulator.
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
   itWithGc('does not pin the written stream behind a carried escape tail', () => {
@@ -97,16 +95,13 @@ describe('advancePartialEscapeTail', () => {
     forceGc()
     const before = process.memoryUsage().heapUsed
     const tails = Array.from({ length: sessions }, (_unused, index) =>
-      // Unterminated OSC, so the tail is carried into the next write.
       advancePartialEscapeTail('', `${'x'.repeat(chunkChars)}\x1b]0;pane-${index} title`)
     )
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
     expect(tails[0]).toBe('\x1b]0;pane-0 title')
-    // The fold must still complete the split sequence on the next chunk.
     expect(advancePartialEscapeTail(tails[0]!, '\x07rest')).toBe('')
-    // 8 MiB of source streams stay alive if the tails are still attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/terminal-partial-escape-tail.ts

@@ -147,9 +147,6 @@ export function extractPartialEscapeTail(stream: string): string {
  *  (tracking abandoned) when the tail exceeds the cap — see the cap comment. */
 export function advancePartialEscapeTail(pendingTail: string, chunk: string): string {
   const tail = extractPartialEscapeTail(pendingTail + chunk)
-  // Detached: this is the ingest-time fold, so callers park the result per
-  // emulator/session until the next write; an attached slice would pin the
-  // whole written stream. Detaching after the cap check keeps the discarded
-  // over-cap tail (an unterminated OSC can span the entire chunk) free.
+  // Persisted emulator tails must not retain the whole written stream.
   return tail.length > MAX_PARTIAL_ESCAPE_TAIL_LENGTH ? '' : detachString(tail)
 }

src/shared/terminal-reply-query-extraction.test.ts

@@ -30,13 +30,7 @@ describe('extractHiddenStartupRendererQueryData', () => {
     expect(pending.length).toBe(HIDDEN_STARTUP_RENDERER_QUERY_PENDING_CHARS)
   })
 
-  // `pending` is parked per pane (hiddenStartupRendererQueryPending in
-  // pty-connection.ts) until the next chunk, so an attached slice pins a whole
-  // PTY chunk per open pane.
-  //
-  // Why the split CSI is long: V8 only builds a SlicedString past
-  // SlicedString::kMinLength (13 chars); a shorter prefix is already copied
-  // flat and would make this assertion pass with or without the detach.
+  // The ≥13-char control arm proves detachment for one persisted query per pane.
   const splitCsi = '\x1b[?1049;2004;2026'
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
@@ -50,8 +44,6 @@ describe('extractHiddenStartupRendererQueryData', () => {
     const pendings = Array.from(
       { length: panes },
       (_unused, index) =>
-        // A distinct chunk per pane, each ending mid-CSI. Sharing one chunk
-        // would leave a single parent alive and pass either way.
         extractHiddenStartupRendererQueryData(
           `${'x'.repeat(chunkChars)}pane-${index}${splitCsi}`,
           ''
@@ -60,12 +52,10 @@ describe('extractHiddenStartupRendererQueryData', () => {
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The pending prefix must still assemble the query across the boundary.
     expect(pendings[0]).toBe(splitCsi)
     expect(extractHiddenStartupRendererQueryData('$p', pendings[0]!).statefulQueryData).toBe(
       `${splitCsi}$p`
     )
-    // 8 MiB of source chunks stay alive if the pending prefixes are attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/terminal-reply-query-extraction.ts

@@ -16,10 +16,8 @@ export type ExtractedRendererQueryData = {
   pending: string
 }
 
-// Detached: callers park `pending` until the next chunk — one per pane in the
-// renderer (hiddenStartupRendererQueryPending, pty-connection.ts) — so an
-// attached slice would pin a whole PTY chunk for as long as the pane lives.
-function detachedQueryPending(input: string, start: number): string {
+// Pending query state must not retain the source PTY chunk.
+export function extractHiddenStartupRendererQueryPending(input: string, start: number): string {
   return detachString(input.slice(start, start + HIDDEN_STARTUP_RENDERER_QUERY_PENDING_CHARS))
 }
 
@@ -53,7 +51,7 @@ export function extractHiddenStartupRendererQueryData(
           statelessQueryData,
           statefulQueryData,
           oscColorQueryData,
-          pending: detachedQueryPending(input, candidateIndex)
+          pending: extractHiddenStartupRendererQueryPending(input, candidateIndex)
         }
       }
       const sequence = input.slice(candidateIndex, finalByteIndex + 1)
@@ -73,7 +71,7 @@ export function extractHiddenStartupRendererQueryData(
           statelessQueryData,
           statefulQueryData,
           oscColorQueryData,
-          pending: detachedQueryPending(input, candidateIndex)
+          pending: extractHiddenStartupRendererQueryPending(input, candidateIndex)
         }
       }
       if (query.kind === 'none') {

src/shared/terminal-reply-query-scan.test.ts

@@ -40,13 +40,7 @@ describe('terminal reply query scan', () => {
     expect(second.queries).toEqual([])
   })
 
-  // The pending prefix is parked in per-stream scan state until the next chunk
-  // for the whole mobile-stream buffering window, so an attached slice pins a
-  // whole PTY chunk per buffering stream.
-  //
-  // Why the pending prefix is long: V8 only builds a SlicedString past
-  // SlicedString::kMinLength (13 chars); a shorter one is already copied flat
-  // and would make this assertion pass with or without the detach.
+  // The ≥13-char control arm proves detachment for one persisted mobile-stream query.
   const splitQueryPrefix = '\x1b[?1049;2004;2026'
   const forcedGc = resolveForcedGc()
   const itWithGc = forcedGc ? it : it.skip
@@ -60,9 +54,6 @@ describe('terminal reply query scan', () => {
     const states = Array.from(
       { length: streams },
       (_unused, index) =>
-        // A distinct chunk per stream, each ending mid-CSI so the partial query
-        // is carried. Sharing one chunk would leave a single parent alive and
-        // make the assertion pass whether or not the pending prefix is detached.
         scanTerminalReplyQuerySequences(
           `${'x'.repeat(chunkChars)}pty-${index}${splitQueryPrefix}`,
           0,
@@ -72,7 +63,6 @@ describe('terminal reply query scan', () => {
     forceGc()
     const retainedMiB = (process.memoryUsage().heapUsed - before) / (1024 * 1024)
 
-    // The pending prefix must still assemble the query across the boundary.
     const pendingStartSeq = states[0]!.pendingStartSeq!
     const resumed = scanTerminalReplyQuerySequences(
       '$p',
@@ -86,7 +76,6 @@ describe('terminal reply query scan', () => {
         endSeq: pendingStartSeq + splitQueryPrefix.length + 2
       }
     ])
-    // 8 MiB of source chunks stay alive if the pending prefixes are attached.
     expect(retainedMiB).toBeLessThan(2)
   })
 })

src/shared/terminal-reply-query-scan.ts

@@ -46,9 +46,7 @@ function isReplyElicitingCsi(sequence: string): boolean {
   )
 }
 
-// Detached: the returned pending prefix is parked in per-stream scan state
-// until the next chunk, so an attached slice would pin a whole PTY chunk for
-// every buffering terminal stream.
+// Persisted query prefixes must not retain their source PTY chunks.
 function boundedPending(input: string, startIndex: number): string {
   return detachString(input.slice(startIndex, startIndex + MAX_PENDING_QUERY_CHARS))
 }