diff --git a/src/main/runtime/agent-launch-typed-line-shell.test.ts b/src/main/runtime/agent-launch-typed-line-shell.test.ts
index 1ef35b85292..9d17d07a49b 100644
--- a/src/main/runtime/agent-launch-typed-line-shell.test.ts
+++ b/src/main/runtime/agent-launch-typed-line-shell.test.ts
@@ -1,5 +1,8 @@
 import { describe, expect, it } from 'vitest'
-import { nameLocalTypedLineShell } from './agent-launch-typed-line-shell'
+import {
+  launchHostProvesAgentInFront,
+  nameLocalTypedLineShell
+} from './agent-launch-typed-line-shell'
 
 describe('naming the shell a local launch line is typed into', () => {
   it('takes the request’s shell first, then the setting, then SHELL, as the spawn does', () => {
@@ -28,3 +31,17 @@ describe('naming the shell a local launch line is typed into', () => {
     ).toBeUndefined()
   })
 })
+
+describe('whether a launch host can prove its agent is in front before a paste', () => {
+  it.each([
+    ['a local macOS host', false, 'darwin', 'darwin', true],
+    ['a local Linux host', false, 'linux', 'linux', true],
+    ['a local Windows host', false, 'win32', 'win32', false],
+    // The pane runs in the distro, but the reads run on the Windows host.
+    ['a local WSL pane', false, 'linux', 'win32', false],
+    ['an SSH Linux host from Windows', true, 'linux', 'win32', true],
+    ['an SSH Windows host', true, 'win32', 'darwin', false]
+  ] as const)('%s: %s', (_label, isRemote, launchPlatform, hostPlatform, proves) => {
+    expect(launchHostProvesAgentInFront({ isRemote, launchPlatform, hostPlatform })).toBe(proves)
+  })
+})
diff --git a/src/main/runtime/agent-launch-typed-line-shell.ts b/src/main/runtime/agent-launch-typed-line-shell.ts
index 6a7f83c048b..f733ade3385 100644
--- a/src/main/runtime/agent-launch-typed-line-shell.ts
+++ b/src/main/runtime/agent-launch-typed-line-shell.ts
@@ -25,3 +25,18 @@ export function nameLocalTypedLineShell(args: {
     '/bin/zsh'
   return basename(shellPath).toLowerCase()
 }
+
+/**
+ * Whether the execution host can prove a launched agent holds its terminal before a paste
+ * (`launched-agent-foreground`). A Windows host cannot, and a local WSL pane runs on one; an SSH
+ * host is judged by its own platform.
+ */
+export function launchHostProvesAgentInFront(args: {
+  isRemote: boolean
+  launchPlatform: NodeJS.Platform
+  hostPlatform?: NodeJS.Platform
+}): boolean {
+  return args.isRemote
+    ? args.launchPlatform !== 'win32'
+    : (args.hostPlatform ?? process.platform) !== 'win32'
+}
diff --git a/src/main/runtime/launched-agent-composer-readiness-claude.test.ts b/src/main/runtime/launched-agent-composer-readiness-claude.test.ts
index 889f08a2ef6..14647906266 100644
--- a/src/main/runtime/launched-agent-composer-readiness-claude.test.ts
+++ b/src/main/runtime/launched-agent-composer-readiness-claude.test.ts
@@ -131,7 +131,7 @@ describe('a fresh orchestration worker start for Claude', () => {
   })
 })
 
-describe('whether a shell is proven in front of a launched agent’s terminal', () => {
+describe('what holds a launched agent’s terminal', () => {
   const originalPlatform = Object.getOwnPropertyDescriptor(process, 'platform')!
   const setPlatform = (value: NodeJS.Platform): void => {
     Object.defineProperty(process, 'platform', { configurable: true, value })
@@ -140,315 +140,206 @@ describe('whether a shell is proven in front of a launched agent’s terminal',
     Object.defineProperty(process, 'platform', originalPlatform)
   })
 
-  describe('macOS and Linux: a fresh foreground read', () => {
+  describe('macOS and Linux: one fresh foreground read decides', () => {
     beforeEach(() => setPlatform('darwin'))
 
     it.each([
       // For its first 5 s the daemon's cached read names the launch agent while zsh is in front.
-      ['claude', 'zsh', true],
-      ['claude', 'claude', false],
-      ['zsh', 'zsh', true],
-      ['-zsh', '-zsh', true],
-      ['bash', 'bash', true],
-      [null, 'zsh', true],
-      [null, null, false]
-    ])('cached %s, scan %s: %s', async (foregroundProcess, scanned, shellInFront) => {
+      ['claude', 'zsh', 'shell'],
+      // A stub agent that crashed: the cached read can still name it after zsh took the terminal.
+      ['python3', 'zsh', 'shell'],
+      ['claude', 'claude', 'agent'],
+      // A native Claude names itself by its version.
+      ['zsh', '2.1.285', 'agent'],
+      ['zsh', 'node', 'agent'],
+      ['-zsh', '-zsh', 'shell'],
+      ['bash', 'bash', 'shell'],
+      [null, 'zsh', 'shell'],
+      ['claude', null, 'unknown']
+    ] as const)('cached %s, scan %s: %s', async (foregroundProcess, scanned, found) => {
       const scan = vi.fn()
+      const proof = vi.fn()
       const { runtime } = await createTranscriptPane({
         paneTitle: 'Claude Code',
         foregroundProcess,
         confirmedForegroundProcess: scanned,
         onForegroundScan: scan,
+        shellForegroundProven: false,
+        onShellForegroundProof: proof,
         launchAgent: 'claude',
         data: ''
       })
 
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'claude')).resolves.toBe(
-        shellInFront
-      )
+      await expect(
+        runtime.readLaunchedAgentForeground(TRANSCRIPT_PANE_PTY_ID, 'claude')
+      ).resolves.toBe(found)
       expect(scan).toHaveBeenCalledOnce()
+      expect(proof).not.toHaveBeenCalled()
     })
 
-    // The terminal daemon answers the shell-foreground check from its recovery state, which an
-    // agent that exits without leaving a full-screen mode up never sets.
-    it('an agent that exited: the scan names zsh while the daemon’s shell check says no', async () => {
-      const proof = vi.fn()
+    it('proves nothing on a controller without a fresh scan', async () => {
       const { runtime } = await createTranscriptPane({
-        paneTitle: 'copilot',
-        foregroundProcess: 'copilot',
-        confirmedForegroundProcess: 'zsh',
-        shellForegroundProven: false,
-        onShellForegroundProof: proof,
+        paneTitle: 'Claude Code',
+        foregroundProcess: 'claude',
+        launchAgent: 'claude',
+        data: ''
+      })
+
+      await expect(
+        runtime.readLaunchedAgentForeground(TRANSCRIPT_PANE_PTY_ID, 'claude')
+      ).resolves.toBe('unknown')
+    })
+
+    // Captured with `ps -o pid=,ppid=,pgid=,tpgid=,stat=,command=` on a pane spawned the way a macOS
+    // pane is, under `login`: zsh holds the terminal in its own group, never the root's.
+    it('reads a login-wrapped pane’s zsh at its prompt as the shell', async () => {
+      const rows: ProcessTableRow[] = [
+        {
+          pid: 60404,
+          ppid: 60394,
+          pgid: 60404,
+          tpgid: 60406,
+          stat: 'Ss',
+          tty: 'ttys004',
+          startTime: 'Fri Oct  2 01:20:11 2026',
+          command: '/usr/bin/login -flpq user /bin/bash --noprofile --norc -p -c'
+        },
+        {
+          pid: 60406,
+          ppid: 60404,
+          pgid: 60406,
+          tpgid: 60406,
+          stat: 'S+',
+          tty: 'ttys004',
+          startTime: 'Fri Oct  2 01:20:11 2026',
+          command: '-/bin/zsh -f'
+        }
+      ]
+      const inspection = {
+        foregroundProcess: 'zsh',
+        hasChildProcesses: true,
+        foregroundProcessEvidence: resolveRemoteForegroundEvidence(
+          { rootPid: 60404, fallbackProcess: 'zsh' },
+          {
+            ptyId: TRANSCRIPT_PANE_PTY_ID,
+            ptyIncarnationId: 'inc-1',
+            authorityGeneration: 'gen-1',
+            observationEpoch: 1,
+            capturedAgeMs: 0,
+            platform: 'darwin'
+          },
+          rows
+        )
+      }
+      // Presence precondition: the fence is live, and its foreground group is not the root's.
+      expect(inspection.foregroundProcessEvidence).toMatchObject({
+        verdict: 'live',
+        fence: { shellPid: 60404, foregroundPgid: 60406 }
+      })
+      const { runtime } = await createTranscriptPane({
+        paneTitle: 'Terminal',
+        foregroundProcess: 'python3',
+        processInspection: inspection,
+        confirmedForegroundProcess: '-zsh',
+        launchAgent: 'claude',
+        data: ''
+      })
+
+      await expect(
+        runtime.readLaunchedAgentForeground(TRANSCRIPT_PANE_PTY_ID, 'claude')
+      ).resolves.toBe('shell')
+    })
+
+    // A wrapper that does not `exec` an agent the scan cannot recognize keeps its own name in front,
+    // which reads as a shell: the prompt stays undelivered rather than risk the user's shell.
+    it('reads a bash wrapper still in front of an unrecognized agent as a shell', async () => {
+      const { runtime } = await createTranscriptPane({
+        paneTitle: 'Terminal',
+        foregroundProcess: 'bash',
+        confirmedForegroundProcess: 'bash',
         launchAgent: 'copilot',
         data: ''
       })
 
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'copilot')).resolves.toBe(
-        true
-      )
-      expect(proof).not.toHaveBeenCalled()
-    })
-
-    it('takes the cached read on a controller without a scan', async () => {
-      const { runtime } = await createTranscriptPane({
-        paneTitle: 'Claude Code',
-        foregroundProcess: 'zsh',
-        launchAgent: 'claude',
-        data: ''
-      })
-
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'claude')).resolves.toBe(
-        true
-      )
+      await expect(
+        runtime.readLaunchedAgentForeground(TRANSCRIPT_PANE_PTY_ID, 'copilot')
+      ).resolves.toBe('shell')
     })
   })
 
-  describe('Windows: only the shell-foreground check', () => {
+  // The scan names the pane's shell for an agent it cannot recognize (an npm agent as `node.exe`),
+  // and Git Bash and WSL keep other processes in the shell's job, so nothing proves the agent.
+  describe('Windows: only the shell-foreground check, and never the agent', () => {
     beforeEach(() => setPlatform('win32'))
 
-    // The scan names the pane's shell for an agent it cannot recognize (an npm agent as
-    // `node.exe`); only the shell alone in the pane's job proves one.
     it.each([
-      ['powershell.exe', false, false],
-      ['powershell.exe', true, true],
-      ['claude', true, true],
-      ['claude', false, false]
-    ])('cached %s, shell check %s: %s', async (foregroundProcess, proven, shellInFront) => {
+      ['powershell.exe', false, 'unknown'],
+      ['powershell.exe', true, 'shell'],
+      ['claude', true, 'shell'],
+      ['claude', false, 'unknown'],
+      ['node', false, 'unknown']
+    ] as const)('cached %s, shell check %s: %s', async (foregroundProcess, proven, found) => {
       const scan = vi.fn()
       const { runtime } = await createTranscriptPane({
         paneTitle: 'Claude Code',
         foregroundProcess,
-        confirmedForegroundProcess: 'powershell.exe',
+        confirmedForegroundProcess: 'claude',
         onForegroundScan: scan,
         shellForegroundProven: proven,
         launchAgent: 'claude',
         data: ''
       })
 
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'claude')).resolves.toBe(
-        shellInFront
-      )
+      await expect(
+        runtime.readLaunchedAgentForeground(TRANSCRIPT_PANE_PTY_ID, 'claude')
+      ).resolves.toBe(found)
       expect(scan).not.toHaveBeenCalled()
     })
-
-    it('proves nothing on a controller without a shell check', async () => {
-      const { runtime } = await createTranscriptPane({
-        paneTitle: 'Claude Code',
-        foregroundProcess: 'powershell.exe',
-        confirmedForegroundProcess: 'powershell.exe',
-        launchAgent: 'claude',
-        data: ''
-      })
-
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'claude')).resolves.toBe(
-        false
-      )
-    })
   })
 
   // The stubs always disagree with the relay's name, so asking either would flip the answer.
   it.each([
-    ['claude', false],
-    ['bash', true]
-  ])(
-    'SSH: takes the relay’s own read %s (shell %s), without a scan or check',
-    async (relayRead, shell) => {
+    ['claude', 'agent'],
+    ['bash', 'shell']
+  ] as const)(
+    'SSH: takes the relay’s own read %s (%s), without a scan or check',
+    async (relayRead, found) => {
+      setPlatform('darwin')
       const scan = vi.fn()
       const proof = vi.fn()
       const { runtime } = await createTranscriptPane({
         paneTitle: 'Claude Code',
         foregroundProcess: relayRead,
-        confirmedForegroundProcess: shell ? 'claude' : 'zsh',
+        confirmedForegroundProcess: found === 'shell' ? 'claude' : 'zsh',
         onForegroundScan: scan,
-        shellForegroundProven: !shell,
+        shellForegroundProven: found !== 'shell',
         onShellForegroundProof: proof,
         connectionId: 'ssh-1',
         launchAgent: 'claude',
         data: ''
       })
 
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'claude')).resolves.toBe(
-        shell
-      )
+      await expect(
+        runtime.readLaunchedAgentForeground(TRANSCRIPT_PANE_PTY_ID, 'claude')
+      ).resolves.toBe(found)
       expect(scan).not.toHaveBeenCalled()
       expect(proof).not.toHaveBeenCalled()
     }
   )
 
-  describe('macOS and Linux: by the shell’s identity, from the host’s process inspection', () => {
-    beforeEach(() => setPlatform('darwin'))
-
-    // Captured with `ps -axo pid=,ppid=,pgid=,tpgid=,stat=,tty=,lstart=,command=` on a pty running
-    // `zsh -f -i` (macOS 26): first while it ran `bash -c 'sleep 4; :'`, then back at its prompt.
-    const zsh = (tpgid: number, stat: string): ProcessTableRow => ({
-      pid: 20894,
-      ppid: 20891,
-      pgid: 20894,
-      tpgid,
-      stat,
-      tty: 'ttys000',
-      startTime: 'Wed Sep 30 09:18:50 2026',
-      command: '/bin/zsh -f -i'
-    })
-    const wrapperRunning: ProcessTableRow[] = [
-      zsh(20996, 'Ss'),
-      {
-        pid: 20996,
-        ppid: 20894,
-        pgid: 20996,
-        tpgid: 20996,
-        stat: 'S+',
-        tty: 'ttys000',
-        startTime: 'Wed Sep 30 09:18:52 2026',
-        command: 'bash -c sleep 4; :'
-      },
-      {
-        pid: 20997,
-        ppid: 20996,
-        pgid: 20996,
-        tpgid: 20996,
-        stat: 'S+',
-        tty: 'ttys000',
-        startTime: 'Wed Sep 30 09:18:52 2026',
-        command: 'sleep 4'
-      }
-    ]
-    const zshAtPrompt: ProcessTableRow[] = [zsh(20894, 'Ss+')]
-    const inspect = (rows: ProcessTableRow[]) => ({
-      foregroundProcess: 'zsh',
-      hasChildProcesses: rows.length > 1,
-      foregroundProcessEvidence: resolveRemoteForegroundEvidence(
-        { rootPid: 20894, fallbackProcess: 'zsh' },
-        {
-          ptyId: TRANSCRIPT_PANE_PTY_ID,
-          ptyIncarnationId: 'inc-1',
-          authorityGeneration: 'gen-1',
-          observationEpoch: 1,
-          capturedAgeMs: 0,
-          platform: 'darwin'
-        },
-        rows
-      )
-    })
-
-    it.each([
-      // A wrapper script's bash runs as its own job: named like a shell, but not the pane's shell.
-      ['a bash wrapper under zsh is not the shell', wrapperRunning, 'bash', false, false],
-      ['zsh back at its prompt is the shell', zshAtPrompt, 'zsh', true, true]
-    ] as const)('%s', async (_label, rows, foregroundName, shell, scanned) => {
-      const inspection = inspect([...rows])
-      // Presence precondition: the captured rows are a live observation, not an unreadable one.
-      expect(inspection.foregroundProcessEvidence).toMatchObject({ verdict: 'live' })
-      const scan = vi.fn()
-      const { runtime } = await createTranscriptPane({
-        paneTitle: 'Terminal',
-        foregroundProcess: foregroundName,
-        processInspection: inspection,
-        confirmedForegroundProcess: foregroundName,
-        onForegroundScan: scan,
-        launchAgent: 'copilot',
-        data: ''
-      })
-
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'copilot')).resolves.toBe(
-        shell
-      )
-      expect(scan).toHaveBeenCalledTimes(scanned ? 1 : 0)
-    })
-
-    it('SSH: takes the relay’s inspection the same way', async () => {
-      const { runtime } = await createTranscriptPane({
-        paneTitle: 'Terminal',
-        foregroundProcess: 'bash',
-        processInspection: inspect([...wrapperRunning]),
-        connectionId: 'ssh-1',
-        launchAgent: 'copilot',
-        data: ''
-      })
-
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'copilot')).resolves.toBe(
-        false
-      )
-    })
-
-    it('an agent in the shell’s own group (a shell without job control) is not the shell', async () => {
-      const { runtime } = await createTranscriptPane({
-        paneTitle: 'Copilot',
-        // node-pty names the group's leader, which is the shell itself here.
-        foregroundProcess: 'zsh',
-        processInspection: inspect([...zshAtPrompt]),
-        confirmedForegroundProcess: 'node',
-        launchAgent: 'copilot',
-        data: ''
-      })
-
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'copilot')).resolves.toBe(
-        false
-      )
-    })
-
-    it('an inspection that cannot observe proves nothing, and falls back to no name', async () => {
-      const scan = vi.fn()
-      const { runtime } = await createTranscriptPane({
-        paneTitle: 'Terminal',
-        foregroundProcess: 'zsh',
-        processInspection: inspect([]),
-        confirmedForegroundProcess: 'zsh',
-        onForegroundScan: scan,
-        launchAgent: 'claude',
-        data: ''
-      })
-
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'claude')).resolves.toBe(
-        false
-      )
-      expect(scan).not.toHaveBeenCalled()
-    })
-  })
-
   // Why: a Windows relay names the pane's shell for an agent its scan cannot recognize (node.exe).
-  it('SSH to a Windows host: the relay’s shell name proves nothing', async () => {
+  it('SSH to a Windows host: never the agent', async () => {
     const { runtime } = await createTranscriptPane({
       paneTitle: 'Copilot',
-      foregroundProcess: 'powershell.exe',
+      foregroundProcess: 'node',
       connectionId: 'ssh-1',
       remoteWindowsHost: true,
       launchAgent: 'copilot',
       data: ''
     })
 
-    await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'copilot')).resolves.toBe(
-      false
-    )
+    await expect(
+      runtime.readLaunchedAgentForeground(TRANSCRIPT_PANE_PTY_ID, 'copilot')
+    ).resolves.toBe('unknown')
   })
-
-  // Why no read: a native Claude names itself by its version, which the cached read gives at once.
-  it.each([
-    ['darwin', '2.1.285'],
-    ['darwin', 'node'],
-    ['win32', 'node']
-  ] as const)(
-    '%s: takes the non-shell name %s, other than the agent’s own, at once',
-    async (platform, foregroundProcess) => {
-      setPlatform(platform)
-      const scan = vi.fn()
-      const proof = vi.fn()
-      const { runtime } = await createTranscriptPane({
-        paneTitle: 'Claude Code',
-        foregroundProcess,
-        confirmedForegroundProcess: 'zsh',
-        onForegroundScan: scan,
-        shellForegroundProven: true,
-        onShellForegroundProof: proof,
-        launchAgent: 'claude',
-        data: ''
-      })
-
-      await expect(runtime.isLaunchShellInFront(TRANSCRIPT_PANE_PTY_ID, 'claude')).resolves.toBe(
-        false
-      )
-      expect(scan).not.toHaveBeenCalled()
-      expect(proof).not.toHaveBeenCalled()
-    }
-  )
 })
diff --git a/src/main/runtime/launched-agent-composer-readiness.ts b/src/main/runtime/launched-agent-composer-readiness.ts
index 7c8fa6871a0..49db277d0f3 100644
--- a/src/main/runtime/launched-agent-composer-readiness.ts
+++ b/src/main/runtime/launched-agent-composer-readiness.ts
@@ -9,7 +9,8 @@
  * a composer from a startup dialog drawn in the same mode, so it counts only while the pane shows no
  * startup dialog and no Codex provisional header (`readFreshComposerHold`). Unlike the desktop's
  * paste it reads only output after the shell's last `?2004l`, and drops a signal while a shell is
- * proven in front (`isLaunchShellInFront`).
+ * proven in front (`readLaunchedAgentForeground`). A signal is never proof by itself: a shell back
+ * at its prompt turns bracketed paste on too, so the write still needs the agent found in front.
  *
  * Where the desktop pasted blind once its budget ran out, the host falls back to the `tui-idle`
  * evidence ranking (idle titles, known ready screens), which also reports a dialog left up. A few
diff --git a/src/main/runtime/launched-agent-crash-guard.live-shell.test.ts b/src/main/runtime/launched-agent-crash-guard.live-shell.test.ts
new file mode 100644
index 00000000000..8dd13f8543c
--- /dev/null
+++ b/src/main/runtime/launched-agent-crash-guard.live-shell.test.ts
@@ -0,0 +1,174 @@
+/**
+ * A launched agent that exits at startup hands the terminal back to its shell, which turns bracketed
+ * paste on at its next prompt just as an agent's composer does. Real zsh with a slow user config,
+ * spawned the way a pane is (under `login` on macOS), read through the terminal daemon's own
+ * foreground tracker: the launch must find the shell and write nothing, and must still find a live
+ * agent.
+ */
+import { spawnSync } from 'node:child_process'
+import { chmodSync, mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import * as pty from 'node-pty'
+import { afterEach, beforeEach, describe, expect, it } from 'vitest'
+import { recognizeAgentProcess } from '../../shared/agent-process-recognition'
+import { getStrictProcessTableSnapshotWithAge } from '../../shared/process-table-snapshot-reader'
+import { createPtyForegroundProcessTracker } from '../daemon/pty-subprocess/foreground-process-tracker'
+import { resolveRemoteForegroundEvidence } from '../providers/agent-foreground-process'
+import {
+  prepareMacosTccLoginShell,
+  wrapShellSpawnForMacosTccAttribution
+} from '../providers/macos-tcc-login-shell'
+import { readLaunchedAgentForeground } from './launched-agent-foreground'
+import { createLaunchedAgentWriteGuard } from './launched-agent-write-guard'
+import { waitForWorktreeStartupDraft } from './runtime-worktree-startup-readiness'
+
+const PTY_ID = 'pty-1'
+const ZSH_PATH =
+  process.platform === 'win32'
+    ? ''
+    : (spawnSync('sh', ['-c', 'command -v zsh'], { encoding: 'utf8' }).stdout ?? '').trim()
+const describeWithZsh = ZSH_PATH ? describe : describe.skip
+
+const STUBS = {
+  // Exits at startup, as an agent with a bad config or a missing dependency does.
+  crash: "#!/bin/sh\nprintf 'stub: crashing at startup\\r\\n'\nexit 1\n",
+  // Opens its composer the way an agent does, then stays up.
+  ready: "#!/bin/sh\nprintf '\\033[?2004hstub agent\\r\\n> '\nexec sleep 30\n"
+} as const
+
+let home = ''
+let savedLoginOptOut: string | undefined
+
+beforeEach(() => {
+  home = mkdtempSync(join(tmpdir(), 'orca-crash-guard-'))
+  // The production spawn: a pane on macOS runs under `login` unless the user opted out.
+  savedLoginOptOut = process.env.ORCA_DISABLE_MACOS_LOGIN_SHELL
+  delete process.env.ORCA_DISABLE_MACOS_LOGIN_SHELL
+})
+
+afterEach(() => {
+  if (savedLoginOptOut !== undefined) {
+    process.env.ORCA_DISABLE_MACOS_LOGIN_SHELL = savedLoginOptOut
+  }
+  rmSync(home, { recursive: true, force: true, maxRetries: 5, retryDelay: 100 })
+})
+
+/** Launches the `claude` stub in a fresh zsh pane and runs the launch's own readiness wait. */
+async function launchStub(stub: keyof typeof STUBS, readyTimeoutMs: number) {
+  const bin = join(home, 'bin')
+  mkdirSync(bin)
+  // Typed by absolute path: `login` resets HOME and PATH, so a bare name would find a real agent.
+  const stubPath = join(bin, 'claude')
+  writeFileSync(stubPath, STUBS[stub])
+  chmodSync(stubPath, 0o755)
+  writeFileSync(join(home, '.zshrc'), 'sleep 2\n')
+  const env = { PATH: '/usr/bin:/bin', HOME: home, ZDOTDIR: home, TERM: 'xterm-256color' }
+  await prepareMacosTccLoginShell()
+  const spawn = wrapShellSpawnForMacosTccAttribution(ZSH_PATH, ['-i'], env)
+  const proc = pty.spawn(spawn.file, spawn.args, { cols: 80, rows: 24, cwd: home, env })
+  let dead = false
+  proc.onExit(() => {
+    dead = true
+  })
+  const listeners = new Set<(data: string) => void>()
+  let output = ''
+  const tracker = createPtyForegroundProcessTracker({
+    process: proc,
+    shellPath: ZSH_PATH,
+    cwd: home,
+    sessionId: 'wt-crash-guard@@pane',
+    startupAgentRecognition: recognizeAgentProcess('claude'),
+    isDead: () => dead
+  })
+  proc.onData((data) => {
+    tracker.recordOutput(data)
+    output += data
+    for (const listener of listeners) {
+      listener(data)
+    }
+  })
+  // The daemon's answers, each from the same pieces the daemon uses.
+  const controller = {
+    getForegroundProcess: async () => tracker.getForegroundProcess(),
+    confirmForegroundProcess: () => tracker.confirmForegroundProcess(),
+    confirmShellForeground: () => tracker.confirmShellForeground(),
+    inspectProcess: async () => {
+      const snapshot = await getStrictProcessTableSnapshotWithAge()
+      return {
+        foregroundProcess: tracker.getForegroundProcess(),
+        hasChildProcesses: true,
+        foregroundProcessEvidence: resolveRemoteForegroundEvidence(
+          { rootPid: proc.pid, fallbackProcess: tracker.getForegroundProcess() },
+          {
+            ptyId: PTY_ID,
+            ptyIncarnationId: 'inc-1',
+            authorityGeneration: 'gen-1',
+            observationEpoch: 1,
+            capturedAgeMs: snapshot.capturedAgeMs,
+            platform: process.platform
+          },
+          snapshot.rows
+        )
+      }
+    }
+  }
+  const readForeground = (ptyId: string) =>
+    readLaunchedAgentForeground(controller, { remote: false, windows: false }, ptyId, 'claude')
+  const subscribe = (_ptyId: string, listener: (data: string) => void) => {
+    listeners.add(listener)
+    return () => listeners.delete(listener)
+  }
+  const ready = waitForWorktreeStartupDraft(
+    {
+      getPtyId: () => PTY_ID,
+      getForegroundProcess: controller.getForegroundProcess,
+      subscribeToData: subscribe,
+      readRecentOutput: () => undefined,
+      write: () => {}
+    },
+    'term-1',
+    'claude',
+    {
+      timeoutMs: readyTimeoutMs,
+      isShellInFront: async (ptyId) => (await readForeground(ptyId)) === 'shell'
+    }
+  )
+  proc.write(`'${stubPath}'\r`)
+  const guard = createLaunchedAgentWriteGuard(
+    { readLaunchedAgentForeground: readForeground, subscribeToTerminalData: subscribe },
+    'claude'
+  )
+  const close = async (): Promise<void> => {
+    guard.dispose()
+    proc.kill()
+    await new Promise((resolve) => setTimeout(resolve, 200))
+  }
+  return { ready, guard, close, output: () => output }
+}
+
+describeWithZsh('a launch prompt after the launched agent exits at startup', () => {
+  it('drops the shell’s prompt as a ready signal and refuses the write', async () => {
+    const launch = await launchStub('crash', 9_000)
+    try {
+      // The shell's prompt after the crash turned bracketed paste on and went quiet, yet no agent did.
+      await expect(launch.ready).resolves.toBeNull()
+      // Presence precondition: the stub, not anything else on the host, is what ran and exited.
+      expect(launch.output()).toContain('stub: crashing at startup')
+      await expect(launch.guard.beforeWrite(PTY_ID)).rejects.toThrow('agent_not_in_foreground')
+    } finally {
+      await launch.close()
+    }
+  }, 30_000)
+
+  it('still finds an agent that stays up, and lets the write through', async () => {
+    const launch = await launchStub('ready', 15_000)
+    try {
+      await expect(launch.ready).resolves.toBe(PTY_ID)
+      expect(launch.output()).toContain('stub agent')
+      await expect(launch.guard.beforeWrite(PTY_ID)).resolves.toBeUndefined()
+    } finally {
+      await launch.close()
+    }
+  }, 30_000)
+})
diff --git a/src/main/runtime/launched-agent-exit-delivery.test.ts b/src/main/runtime/launched-agent-exit-delivery.test.ts
new file mode 100644
index 00000000000..37c90e4354f
--- /dev/null
+++ b/src/main/runtime/launched-agent-exit-delivery.test.ts
@@ -0,0 +1,192 @@
+/**
+ * An agent that exits at startup leaves its shell at a prompt that turns bracketed paste on, which
+ * is the same signal an agent's composer gives. Driven through the launch's own readiness wait,
+ * foreground read and write guard, with each host answering the way it does after the exit.
+ */
+
+import { readFileSync } from 'node:fs'
+import { join } from 'node:path'
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+import { deliverTerminalAgentLaunchPrompt } from './rpc/methods/agent-launch-terminal-prompt'
+import { readLaunchedAgentForeground } from './launched-agent-foreground'
+import { waitForWorktreeStartupDraft } from './runtime-worktree-startup-readiness'
+
+const PTY_ID = 'pty-1'
+const SHELL_HANDOFF = '\x1b[?2004l'
+
+/** zsh's prompt, the launch line it runs, then its prompt again once that command exited. */
+function zshRunsLaunchLineThatExits(): string[] {
+  const zsh = readFileSync(join(__dirname, '__fixtures__', 'zsh-prompt-runs-command.txt'), 'utf8')
+  const split = zsh.indexOf('\n', zsh.indexOf(SHELL_HANDOFF)) + 1
+  return [zsh.slice(0, split), 'stub: crashing at startup\r\n', zsh.slice(split)]
+}
+
+/** Git Bash and WSL's bash: readline turns bracketed paste on at each prompt. */
+function bashRunsLaunchLineThatExits(): string[] {
+  return [
+    '\x1b[?2004hqa@host:~$ claude\r\n\x1b[?2004l\r',
+    'stub: crashing at startup\r\n',
+    '\x1b[?2004hqa@host:~$ '
+  ]
+}
+
+type HostAnswers = {
+  host: { remote: boolean; windows: boolean }
+  cached: string | null
+  scanned: string | null
+  shellAlone: boolean
+}
+
+const HOSTS: [string, HostAnswers, () => string[]][] = [
+  // The shell keeps other processes in its job, so the shell-alone check never answers.
+  [
+    'Windows Git Bash',
+    {
+      host: { remote: false, windows: true },
+      cached: 'claude',
+      scanned: 'bash',
+      shellAlone: false
+    },
+    bashRunsLaunchLineThatExits
+  ],
+  [
+    'Windows WSL',
+    { host: { remote: false, windows: true }, cached: 'claude', scanned: 'wsl', shellAlone: false },
+    bashRunsLaunchLineThatExits
+  ],
+  // A pane under `login`, whose cached name is still the exited stub's.
+  [
+    'macOS zsh',
+    {
+      host: { remote: false, windows: false },
+      cached: 'python3',
+      scanned: 'zsh',
+      shellAlone: false
+    },
+    zshRunsLaunchLineThatExits
+  ]
+]
+
+function launchedPane(answers: HostAnswers) {
+  const listeners = new Set<(data: string) => void>()
+  const subscribe = (_ptyId: string, listener: (data: string) => void) => {
+    listeners.add(listener)
+    return () => listeners.delete(listener)
+  }
+  const controller = {
+    getForegroundProcess: async () => answers.cached,
+    confirmForegroundProcess: async () => answers.scanned,
+    confirmShellForeground: async () => answers.shellAlone
+  }
+  const readForeground = (ptyId: string) =>
+    readLaunchedAgentForeground(controller, answers.host, ptyId, 'claude')
+  const writes: string[] = []
+  const runtime = {
+    waitForFreshWorkerComposer: async (_handle: string, agent: 'claude', timeoutMs: number) => {
+      const ptyId = await waitForWorktreeStartupDraft(
+        {
+          getPtyId: () => PTY_ID,
+          getForegroundProcess: controller.getForegroundProcess,
+          subscribeToData: subscribe,
+          readRecentOutput: () => undefined,
+          write: () => {}
+        },
+        'term-1',
+        agent,
+        {
+          timeoutMs,
+          requireComposerMarker: false,
+          isShellInFront: async (id) => (await readForeground(id)) === 'shell'
+        }
+      )
+      if (!ptyId) {
+        throw new Error('timeout')
+      }
+      return { handle: 'term-1', satisfied: true, status: 'running' }
+    },
+    // The idle evidence after the budget: the shell's prompt can look settled too.
+    waitForTerminal: vi.fn(async () => ({ handle: 'term-1', satisfied: true, status: 'idle' })),
+    readLaunchedAgentForeground: (ptyId: string) => readForeground(ptyId),
+    subscribeToTerminalData: subscribe,
+    sendTerminalAgentPrompt: vi.fn(
+      async (
+        handle: string,
+        text: string,
+        options: { beforeWrite?: (id: string) => Promise<void> }
+      ) => {
+        await options.beforeWrite?.(PTY_ID)
+        writes.push(text)
+        return { handle, accepted: true, bytesWritten: text.length }
+      }
+    )
+  }
+  return {
+    runtime,
+    writes,
+    emit: (data: string) => {
+      for (const listener of listeners) {
+        listener(data)
+      }
+    }
+  }
+}
+
+describe('a launch prompt after the launched agent exits at startup', () => {
+  beforeEach(() => {
+    vi.useFakeTimers()
+    vi.spyOn(console, 'warn').mockImplementation(() => {})
+  })
+  afterEach(() => {
+    vi.useRealTimers()
+    vi.restoreAllMocks()
+  })
+
+  it.each(HOSTS)(
+    '%s: types nothing and reports it undelivered',
+    async (_label, answers, transcript) => {
+      const pane = launchedPane(answers)
+      const delivery = deliverTerminalAgentLaunchPrompt({
+        // oxlint-disable-next-line typescript/consistent-type-assertions -- SAFETY: the deliverer reaches only the methods built above.
+        runtime: pane.runtime as unknown as Parameters<
+          typeof deliverTerminalAgentLaunchPrompt
+        >[0]['runtime'],
+        handle: 'term-1',
+        agent: 'claude',
+        freshLaunch: true,
+        text: 'QA prompt that must never run as a shell command'
+      })
+      for (const chunk of transcript()) {
+        pane.emit(chunk)
+      }
+      await vi.advanceTimersByTimeAsync(61_000)
+
+      await expect(delivery).resolves.toBe(false)
+      expect(pane.writes).toEqual([])
+    }
+  )
+
+  it('macOS zsh: still writes into an agent that stays up', async () => {
+    const pane = launchedPane({
+      host: { remote: false, windows: false },
+      cached: 'claude',
+      scanned: 'claude',
+      shellAlone: false
+    })
+    const delivery = deliverTerminalAgentLaunchPrompt({
+      // oxlint-disable-next-line typescript/consistent-type-assertions -- SAFETY: the deliverer reaches only the methods built above.
+      runtime: pane.runtime as unknown as Parameters<
+        typeof deliverTerminalAgentLaunchPrompt
+      >[0]['runtime'],
+      handle: 'term-1',
+      agent: 'claude',
+      freshLaunch: true,
+      text: 'QA prompt'
+    })
+    pane.emit(zshRunsLaunchLineThatExits()[0])
+    pane.emit('\x1b[?2004hClaude Code\r\n> ')
+    await vi.advanceTimersByTimeAsync(2_000)
+
+    await expect(delivery).resolves.toBe(true)
+    expect(pane.writes).toEqual(['QA prompt'])
+  })
+})
diff --git a/src/main/runtime/launched-agent-foreground.ts b/src/main/runtime/launched-agent-foreground.ts
new file mode 100644
index 00000000000..2d1abb1f9bf
--- /dev/null
+++ b/src/main/runtime/launched-agent-foreground.ts
@@ -0,0 +1,60 @@
+import { isExpectedAgentProcess } from '../../shared/agent-process-recognition'
+import { isShellProcess } from '../../shared/shell-process-detection'
+import type { TuiAgent } from '../../shared/tui-agent'
+import { TUI_AGENT_CONFIG } from '../../shared/tui-agent-config'
+import type { RuntimePtyController } from './runtime-pty-controller-contract'
+
+/**
+ * What holds a launched agent's terminal: the agent (anything but the pane's shell), the shell
+ * (the launch line has not run yet, or the agent exited), or `unknown` when the host cannot tell.
+ * Only `agent` lets a launch write its prompt; only `shell` drops a ready signal.
+ */
+export type LaunchedAgentForeground = 'agent' | 'shell' | 'unknown'
+
+/** A login shell is reported as `-zsh`. */
+function isLaunchShell(processName: string): boolean {
+  return isShellProcess(processName.replace(/^-/, ''))
+}
+
+/**
+ * One fresh read of the terminal's foreground process: on a local POSIX host the scan behind
+ * `confirmForegroundProcess`, and on an SSH host the relay's name, which it reads from the terminal
+ * when asked. Never the cached name a tab icon uses: it can still name a process that already
+ * exited. Not the process-group fence either: a macOS pane runs its shell under `login`, so the
+ * fence's root is never the shell's group.
+ *
+ * Windows has no foreground process group, and its scan names the pane's shell for an agent it
+ * cannot recognize, while Git Bash and WSL keep other processes in the shell's job, so nothing
+ * there proves the agent. A Windows host can still prove its shell alone (`confirmShellForeground`).
+ */
+export async function readLaunchedAgentForeground(
+  controller: Pick<
+    RuntimePtyController,
+    'getForegroundProcess' | 'confirmForegroundProcess' | 'confirmShellForeground'
+  > | null,
+  host: { remote: boolean; windows: boolean },
+  ptyId: string,
+  agent: TuiAgent
+): Promise<LaunchedAgentForeground> {
+  if (!controller) {
+    return 'unknown'
+  }
+  try {
+    if (host.windows) {
+      // An SSH pane has no such check, so on a Windows relay nothing proves a shell.
+      return (await controller.confirmShellForeground?.(ptyId)) ? 'shell' : 'unknown'
+    }
+    const foreground = host.remote
+      ? await controller.getForegroundProcess(ptyId)
+      : ((await controller.confirmForegroundProcess?.(ptyId)) ?? null)
+    if (!foreground) {
+      return 'unknown'
+    }
+    return isLaunchShell(foreground) &&
+      !isExpectedAgentProcess(foreground, TUI_AGENT_CONFIG[agent].expectedProcess)
+      ? 'shell'
+      : 'agent'
+  } catch {
+    return 'unknown'
+  }
+}
diff --git a/src/main/runtime/launched-agent-shell-foreground.ts b/src/main/runtime/launched-agent-shell-foreground.ts
deleted file mode 100644
index 9b5c20bb7f1..00000000000
--- a/src/main/runtime/launched-agent-shell-foreground.ts
+++ /dev/null
@@ -1,92 +0,0 @@
-import {
-  isExpectedAgentProcess,
-  recognizeAgentProcess
-} from '../../shared/agent-process-recognition'
-import { isShellProcess } from '../../shared/shell-process-detection'
-import type { TuiAgent } from '../../shared/tui-agent'
-import { TUI_AGENT_CONFIG } from '../../shared/tui-agent-config'
-import type { RuntimePtyController } from './runtime-pty-controller-contract'
-
-/** A login shell is reported as `-zsh`. */
-function isLaunchShell(processName: string): boolean {
-  return isShellProcess(processName.replace(/^-/, ''))
-}
-
-/**
- * Whether the pane's own shell, rather than the agent a launch started, is proven to be in the
- * terminal's foreground: the launch line has not run yet, or the agent exited. A read that fails or
- * cannot tell proves nothing, so it never delays a launch.
- *
- * On macOS and Linux identity decides first: the host's process inspection reports the terminal's
- * foreground process group, and the pane's shell can be in front only when that group is the
- * shell's own. A wrapper script's bash, or a conda hook, runs as its own job, so it is not the
- * pane's shell even though it is named like one. A fresh read must then also name a shell.
- *
- * Windows has no foreground process group, and its scan names the pane's shell for any agent it
- * cannot recognize (an npm agent running as `node.exe`), so it takes only the host's
- * shell-foreground check (the shell alone in the pane's job). A daemon pane answers that from its
- * recovery state, which a plain exit leaves unset, and an SSH pane has none, so there it proves
- * nothing.
- *
- * Why no check for a non-shell name other than the agent's own: for its first 5 s the daemon's
- * cached read names the launch agent whenever a shell is in front (measured on a zsh pane: cached
- * `copilot`, scan `zsh`), so any other non-shell name it gives is the process itself. macOS
- * reports the native Claude by its version (`2.1.258`), and a runtime such as node can front an
- * agent.
- */
-export async function isShellInFrontOfLaunchedAgent(
-  controller: Pick<
-    RuntimePtyController,
-    | 'getForegroundProcess'
-    | 'inspectProcess'
-    | 'confirmForegroundProcess'
-    | 'confirmShellForeground'
-  > | null,
-  host: { remote: boolean; windows: boolean },
-  ptyId: string,
-  agent: TuiAgent
-): Promise<boolean> {
-  if (!controller) {
-    return false
-  }
-  try {
-    const cached = await controller.getForegroundProcess(ptyId)
-    if (cached && !isLaunchShell(cached) && recognizeAgentProcess(cached)?.agent !== agent) {
-      return false
-    }
-    if (host.windows) {
-      // An SSH pane has no such check, so on a Windows relay nothing proves a shell.
-      return (await controller.confirmShellForeground?.(ptyId)) ?? false
-    }
-    const evidence = (await controller.inspectProcess?.(ptyId))?.foregroundProcessEvidence
-    if (
-      evidence &&
-      !(
-        evidence.verdict === 'live' &&
-        evidence.fence.platform === 'posix' &&
-        evidence.fence.foregroundPgid === evidence.fence.shellPid
-      )
-    ) {
-      return false
-    }
-    // Why the name too: a shell without job control runs its commands in its own group.
-    // A host that predates the evidence (an older daemon or relay, or a pane outside the daemon)
-    // has only the name, which reads a wrapper script's bash as the shell. A relay's cached name is
-    // read live, and it has no scan.
-    const foreground =
-      !host.remote && controller.confirmForegroundProcess
-        ? await controller.confirmForegroundProcess(ptyId)
-        : cached
-    return isShellName(foreground, agent)
-  } catch {
-    return false
-  }
-}
-
-function isShellName(processName: string | null, agent: TuiAgent): boolean {
-  return (
-    !!processName &&
-    isLaunchShell(processName) &&
-    !isExpectedAgentProcess(processName, TUI_AGENT_CONFIG[agent].expectedProcess)
-  )
-}
diff --git a/src/main/runtime/launched-agent-write-guard.test.ts b/src/main/runtime/launched-agent-write-guard.test.ts
index 4af5f64c3b7..29d6aed51ca 100644
--- a/src/main/runtime/launched-agent-write-guard.test.ts
+++ b/src/main/runtime/launched-agent-write-guard.test.ts
@@ -1,17 +1,20 @@
 import { describe, expect, it, vi } from 'vitest'
+import type { LaunchedAgentForeground } from './launched-agent-foreground'
 import {
   createLaunchedAgentWriteGuard,
   type LaunchedAgentWriteGuardRuntime
 } from './launched-agent-write-guard'
 
-function guardHarness(shellInFront: boolean[]) {
-  const answers = [...shellInFront]
+function guardHarness(foregrounds: LaunchedAgentForeground[]) {
+  const answers = [...foregrounds]
   let listener: ((data: string) => void) | null = null
   const unsubscribe = vi.fn(() => {
     listener = null
   })
   const runtime: LaunchedAgentWriteGuardRuntime = {
-    isLaunchShellInFront: vi.fn(async () => answers.shift() ?? false),
+    readLaunchedAgentForeground: vi.fn(
+      async (): Promise<LaunchedAgentForeground> => answers.shift() ?? 'agent'
+    ),
     subscribeToTerminalData: vi.fn((_ptyId: string, next: (data: string) => void) => {
       listener = next
       return unsubscribe
@@ -22,23 +25,35 @@ function guardHarness(shellInFront: boolean[]) {
 }
 
 describe('the check before each write of a launch prompt', () => {
-  it('reuses a read that found no shell for the writes after it', async () => {
-    const { runtime, guard } = guardHarness([false])
+  it('reuses a read that found the agent for the writes after it', async () => {
+    const { runtime, guard } = guardHarness(['agent'])
 
     await guard.beforeWrite('pty-1')
     // The agent's own output after the paste says nothing about a shell.
     await guard.beforeWrite('pty-1')
     await guard.beforeWrite('pty-1')
 
-    expect(runtime.isLaunchShellInFront).toHaveBeenCalledTimes(1)
+    expect(runtime.readLaunchedAgentForeground).toHaveBeenCalledTimes(1)
   })
 
+  // Why: a ready signal can come from a shell whose agent exited, so only a read that finds the
+  // agent may let the text through; an unanswered read is not that.
+  it.each(['shell', 'unknown'] as const)(
+    'refuses a write when the read finds %s',
+    async (found) => {
+      const { guard, unsubscribe } = guardHarness([found])
+
+      await expect(guard.beforeWrite('pty-1')).rejects.toThrow('agent_not_in_foreground')
+      expect(unsubscribe).toHaveBeenCalledTimes(1)
+    }
+  )
+
   it.each([
     ['a shell turning bracketed paste on', '$ \x1b[?2004h'],
     ['a shell integration prompt mark', '\x1b]133;D;1\x07\x1b]133;A\x07$ '],
     ['a mark split across two chunks', ['\x1b[?20', '04h']]
   ] as const)('reads again after %s, and refuses a shell', async (_label, output) => {
-    const { runtime, guard, emit } = guardHarness([false, true])
+    const { runtime, guard, emit } = guardHarness(['agent', 'shell'])
 
     await guard.beforeWrite('pty-1')
     for (const chunk of typeof output === 'string' ? [output] : output) {
@@ -46,17 +61,17 @@ describe('the check before each write of a launch prompt', () => {
     }
 
     await expect(guard.beforeWrite('pty-1')).rejects.toThrow('agent_not_in_foreground')
-    expect(runtime.isLaunchShellInFront).toHaveBeenCalledTimes(2)
+    expect(runtime.readLaunchedAgentForeground).toHaveBeenCalledTimes(2)
   })
 
   it('catches a shell that returns while the first read is still running', async () => {
     const { runtime, guard, emit } = guardHarness([])
-    vi.mocked(runtime.isLaunchShellInFront)
+    vi.mocked(runtime.readLaunchedAgentForeground)
       .mockImplementationOnce(async () => {
         emit('\x1b[?2004h')
-        return false
+        return 'agent'
       })
-      .mockResolvedValueOnce(true)
+      .mockResolvedValueOnce('shell')
 
     await guard.beforeWrite('pty-1')
 
@@ -64,11 +79,11 @@ describe('the check before each write of a launch prompt', () => {
   })
 
   it('stops watching when it refuses and when it is disposed', async () => {
-    const refused = guardHarness([true])
+    const refused = guardHarness(['shell'])
     await expect(refused.guard.beforeWrite('pty-1')).rejects.toThrow('agent_not_in_foreground')
     expect(refused.unsubscribe).toHaveBeenCalledTimes(1)
 
-    const cleared = guardHarness([false])
+    const cleared = guardHarness(['agent'])
     await cleared.guard.beforeWrite('pty-1')
     cleared.guard.dispose()
     expect(cleared.unsubscribe).toHaveBeenCalledTimes(1)
diff --git a/src/main/runtime/launched-agent-write-guard.ts b/src/main/runtime/launched-agent-write-guard.ts
index 80f0bf71d6b..7e29f3008c4 100644
--- a/src/main/runtime/launched-agent-write-guard.ts
+++ b/src/main/runtime/launched-agent-write-guard.ts
@@ -1,5 +1,6 @@
 import type { TuiAgent } from '../../shared/tui-agent'
 import type { OrcaRuntimeService } from './orca-runtime'
+import type { LaunchedAgentForeground } from './launched-agent-foreground'
 
 /** A shell back at its prompt turns bracketed paste on, and Orca's shell integration marks it. */
 const SHELL_RETURN_MARKERS = ['\x1b[?2004', '\x1b]133;'] as const
@@ -7,14 +8,15 @@ const MARKER_CARRY_CHARS = Math.max(...SHELL_RETURN_MARKERS.map((marker) => mark
 
 export type LaunchedAgentWriteGuardRuntime = Pick<
   OrcaRuntimeService,
-  'isLaunchShellInFront' | 'subscribeToTerminalData'
+  'readLaunchedAgentForeground' | 'subscribeToTerminalData'
 >
 
 /**
  * The check before each write of a launch prompt: the paste, its Enter, and Codex's second Enter.
- * A shell proven in front refuses the write. Once a read finds none, later writes reuse it until the
- * terminal shows a shell coming back to its prompt, so Enter follows the paste on the desktop's
- * timing instead of waiting out another process read.
+ * A write needs a fresh read that finds the agent in front; a shell, or a host that cannot tell,
+ * refuses it, since a ready signal alone can come from a shell back at its prompt. Once a read finds
+ * the agent, later writes reuse it until the terminal shows a shell coming back to its prompt, so
+ * Enter follows the paste on the desktop's timing instead of waiting out another process read.
  */
 export function createLaunchedAgentWriteGuard(
   runtime: LaunchedAgentWriteGuardRuntime,
@@ -41,14 +43,14 @@ export function createLaunchedAgentWriteGuard(
         watch.shellMayHaveReturned = true
       }
     })
-    let shellInFront: boolean
+    let foreground: LaunchedAgentForeground
     try {
-      shellInFront = await runtime.isLaunchShellInFront(ptyId, agent)
+      foreground = await runtime.readLaunchedAgentForeground(ptyId, agent)
     } catch (error) {
       watch.unsubscribe()
       throw error
     }
-    if (shellInFront) {
+    if (foreground !== 'agent') {
       watch.unsubscribe()
       throw new Error('agent_not_in_foreground')
     }
diff --git a/src/main/runtime/orca-runtime-activate-managed-worktree.ts b/src/main/runtime/orca-runtime-activate-managed-worktree.ts
index cecf3223b10..7a17aadfc3c 100644
--- a/src/main/runtime/orca-runtime-activate-managed-worktree.ts
+++ b/src/main/runtime/orca-runtime-activate-managed-worktree.ts
@@ -33,7 +33,10 @@ import type { CreateWorktreeResult } from '../../shared/worktree/create-types'
 import { provisionWorktreeTerminals } from './runtime-worktree-terminal-provisioning'
 import { readFreshComposerHold } from './launched-agent-composer-readiness'
 import { buildTerminalWaitText } from './terminal-wait-tail-state'
-import { isShellInFrontOfLaunchedAgent } from './launched-agent-shell-foreground'
+import {
+  readLaunchedAgentForeground,
+  type LaunchedAgentForeground
+} from './launched-agent-foreground'
 
 export class OrcaRuntimeWithActivateManagedWorktree extends OrcaRuntimeWithListManagedWorktrees {
   async activateManagedWorktree(
@@ -208,7 +211,8 @@ export class OrcaRuntimeWithActivateManagedWorktree extends OrcaRuntimeWithListM
         timeoutMs,
         requireComposerMarker,
         signal: stop.signal,
-        isShellInFront: (ownerPtyId) => this.isLaunchShellInFront(ownerPtyId, agent),
+        isShellInFront: async (ownerPtyId) =>
+          (await this.readLaunchedAgentForeground(ownerPtyId, agent)) === 'shell',
         accept: (readyPtyId) => {
           const pty = this.ptysById.get(readyPtyId)
           const hold = pty
@@ -234,12 +238,17 @@ export class OrcaRuntimeWithActivateManagedWorktree extends OrcaRuntimeWithListM
     return this.buildTuiIdleProbeResult(handle, null)
   }
 
-  /** Whether a shell, not the agent a launch started, is proven to be in the terminal's foreground. */
-  isLaunchShellInFront(ptyId: string, agent: TuiAgent): Promise<boolean> {
+  /** What holds the terminal a launch started its agent in, read fresh from the execution host. */
+  readLaunchedAgentForeground(ptyId: string, agent: TuiAgent): Promise<LaunchedAgentForeground> {
     const pty = this.ptysById.get(ptyId)
-    return isShellInFrontOfLaunchedAgent(
+    const remote = !!pty?.connectionId
+    return readLaunchedAgentForeground(
       this.ptyController,
-      { remote: !!pty?.connectionId, windows: this.pathFlavorForPty(pty) === 'win32' },
+      // A local WSL pane still runs on a Windows host, whose process reads cannot see into it.
+      {
+        remote,
+        windows: remote ? this.pathFlavorForPty(pty) === 'win32' : process.platform === 'win32'
+      },
       ptyId,
       agent
     )
diff --git a/src/main/runtime/orca-runtime-resolve-worktree-removal-target.ts b/src/main/runtime/orca-runtime-resolve-worktree-removal-target.ts
index cbbf85aa37c..643441467ec 100644
--- a/src/main/runtime/orca-runtime-resolve-worktree-removal-target.ts
+++ b/src/main/runtime/orca-runtime-resolve-worktree-removal-target.ts
@@ -21,7 +21,10 @@ import { resolveLocalProjectRuntimeForWorktreeId } from '../local-project-runtim
 import { resolveBareAgentLaunchCommand } from './runtime-agent-launch-resolution'
 import { agentPromptRidesLaunchCommand } from '../../shared/tui-agent-startup'
 import { planStartupWithPromptCandidate } from '../../shared/startup-line-prompt-carry'
-import { nameLocalTypedLineShell } from './agent-launch-typed-line-shell'
+import {
+  launchHostProvesAgentInFront,
+  nameLocalTypedLineShell
+} from './agent-launch-typed-line-shell'
 import { resolveAgentStartupPlanInputs } from '../../shared/agent-startup-plan-inputs'
 import { agentStartedTelemetry } from '../agent-launch/agent-started-telemetry'
 import { LOCAL_EXECUTION_HOST_ID } from '../../shared/execution-host'
@@ -273,13 +276,16 @@ export class OrcaRuntimeWithResolveWorktreeRemovalTarget extends OrcaRuntimeWith
         sessionOptions: this.toAgentSessionOptions(opts.launchPreferences)
       }),
       opts.startupPrompt ?? '',
-      nameLocalTypedLineShell({
-        isRemote,
-        ...(opts.shellOverride ? { shellOverride: opts.shellOverride } : {}),
-        ...(settings.terminalDefaultShell
-          ? { defaultShellSetting: settings.terminalDefaultShell }
-          : {})
-      })
+      {
+        shellName: nameLocalTypedLineShell({
+          isRemote,
+          ...(opts.shellOverride ? { shellOverride: opts.shellOverride } : {}),
+          ...(settings.terminalDefaultShell
+            ? { defaultShellSetting: settings.terminalDefaultShell }
+            : {})
+        }),
+        provesAgentInFront: launchHostProvesAgentInFront({ isRemote, launchPlatform: platform })
+      }
     )
     if (!startupPlan) {
       // Why: an explicit agent that yields no plan would otherwise spawn a bare
diff --git a/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.test.ts b/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.test.ts
index 0c43c77b40d..fc58a875377 100644
--- a/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.test.ts
+++ b/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.test.ts
@@ -26,8 +26,8 @@ function runtimeStub(overrides: {
   send?: SendFn
   /** Whether the agent's composer signal fires within its budget; else the idle evidence decides. */
   composerSignal?: boolean
-  /** Whether a shell is proven in the terminal's foreground. */
-  shellInFront?: boolean
+  /** What a fresh read finds in the terminal's foreground; default: the agent. */
+  foreground?: 'agent' | 'shell' | 'unknown'
 }) {
   const queued = [...(overrides.waits ?? [])]
   const waitForTerminal = vi.fn(
@@ -40,7 +40,7 @@ function runtimeStub(overrides: {
     }
     return COMPOSER_READY
   })
-  const isLaunchShellInFront = vi.fn(async () => overrides.shellInFront ?? false)
+  const readLaunchedAgentForeground = vi.fn(async () => overrides.foreground ?? 'agent')
   const subscribeToTerminalData = vi.fn(() => () => {})
   const sendTerminalAgentPrompt = vi.fn<SendFn>(
     overrides.send ?? (async () => ({ handle: 'term_1', accepted: true, bytesWritten: 12 }))
@@ -49,13 +49,13 @@ function runtimeStub(overrides: {
     waitForTerminal,
     waitForFreshWorkerComposer,
     sendTerminalAgentPrompt,
-    isLaunchShellInFront,
+    readLaunchedAgentForeground,
     // oxlint-disable-next-line typescript/consistent-type-assertions -- SAFETY: the deliverer reaches exactly these five runtime methods; anything else would throw rather than read a wrong value.
     runtime: {
       waitForTerminal,
       waitForFreshWorkerComposer,
       sendTerminalAgentPrompt,
-      isLaunchShellInFront,
+      readLaunchedAgentForeground,
       subscribeToTerminalData
     } as unknown as Parameters<typeof deliverTerminalAgentLaunchPrompt>[0]['runtime']
   }
@@ -228,12 +228,13 @@ describe('writing a launch prompt into a terminal agent', () => {
   })
 
   it.each([
-    ['refuses the write when the shell is back in front', true, false],
-    // An unread foreground is no proof the agent exited; the composer was just seen ready.
-    ['writes unless a shell is proven in front', false, true]
-  ])('%s', async (_label, shellInFront, delivered) => {
+    ['writes when a fresh read finds the agent in front', 'agent', true],
+    ['refuses the write when the shell is back in front', 'shell', false],
+    // A ready signal can be a shell's own prompt, so a read that cannot tell proves nothing.
+    ['refuses the write when the host cannot tell what is in front', 'unknown', false]
+  ] as const)('%s', async (_label, foreground, delivered) => {
     // An agent that exits at startup hands its shell back, which must never run the prompt.
-    const stub = runtimeStub({ composerSignal: true, shellInFront })
+    const stub = runtimeStub({ composerSignal: true, foreground })
     stub.sendTerminalAgentPrompt.mockImplementation(async (handle, _text, options) => {
       const { beforeWrite } = options
       if (typeof beforeWrite === 'function') {
@@ -251,7 +252,7 @@ describe('writing a launch prompt into a terminal agent', () => {
         text: 'do the thing'
       })
     ).resolves.toBe(delivered)
-    expect(stub.isLaunchShellInFront).toHaveBeenCalledWith('pty-1', 'claude')
+    expect(stub.readLaunchedAgentForeground).toHaveBeenCalledWith('pty-1', 'claude')
   })
 
   it('checks the foreground once for the paste, its Enter and the second Enter', async () => {
@@ -275,7 +276,7 @@ describe('writing a launch prompt into a terminal agent', () => {
         text: 'do the thing'
       })
     ).resolves.toBe(true)
-    expect(stub.isLaunchShellInFront).toHaveBeenCalledTimes(1)
+    expect(stub.readLaunchedAgentForeground).toHaveBeenCalledTimes(1)
   })
 
   it('writes as soon as the composer signal fires, without consulting the idle evidence', async () => {
@@ -381,6 +382,27 @@ describe('writing a launch prompt into a terminal agent', () => {
     // A reused pane's render state is only inferred, so its Enter still waits for the render.
     expect(stub.sendTerminalAgentPrompt.mock.calls[0]?.[2]?.composerReady).toBe(false)
   })
+
+  it('writes nothing into a reused terminal whose agent is no longer in front', async () => {
+    const stub = runtimeStub({ foreground: 'shell' })
+    stub.sendTerminalAgentPrompt.mockImplementation(async (handle, _text, options) => {
+      const { beforeWrite } = options
+      if (typeof beforeWrite === 'function') {
+        await beforeWrite('pty-1')
+      }
+      return { handle, accepted: true, bytesWritten: 12 }
+    })
+
+    await expect(
+      deliverTerminalAgentLaunchPrompt({
+        runtime: stub.runtime,
+        handle: 'term_existing',
+        agent: 'claude',
+        freshLaunch: false,
+        text: 'do the thing'
+      })
+    ).resolves.toBe(false)
+  })
 })
 
 function fakeClock() {
diff --git a/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.ts b/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.ts
index 5af709eb0b1..8cffda5cf01 100644
--- a/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.ts
+++ b/src/main/runtime/rpc/methods/agent-launch-terminal-prompt.ts
@@ -107,9 +107,9 @@ export async function deliverTerminalAgentLaunchPrompt(args: {
   if (args.text.trim().length === 0) {
     return false
   }
-  // Before the paste and again before Enter: an agent that exited in between leaves its shell in
-  // the foreground, which must never receive the prompt or its Enter.
-  const guard = args.freshLaunch ? createLaunchedAgentWriteGuard(args.runtime, args.agent) : null
+  // Before the paste and again before Enter, for a reused pane too: a ready signal can come from a
+  // shell whose agent exited, so only a read that finds the agent in front lets the text through.
+  const guard = createLaunchedAgentWriteGuard(args.runtime, args.agent)
   try {
     const wait = await waitThroughBlockingPrompts(
       args.runtime,
@@ -130,7 +130,7 @@ export async function deliverTerminalAgentLaunchPrompt(args: {
       inputKind: 'launch',
       // A fresh launch's composer was just seen ready; a reused pane's state is only inferred.
       composerReady: args.freshLaunch,
-      ...(guard ? { beforeWrite: guard.beforeWrite } : {}),
+      beforeWrite: guard.beforeWrite,
       // Paired: together these take the queued path, which settles an unobserved turn start into
       // an `input_accepted` receipt rather than raising it. Without the id the write is verified
       // strictly and a slow first turn throws.
@@ -151,6 +151,6 @@ export async function deliverTerminalAgentLaunchPrompt(args: {
     )
     return false
   } finally {
-    guard?.dispose()
+    guard.dispose()
   }
 }
diff --git a/src/main/runtime/runtime-worktree-agent-startup.ts b/src/main/runtime/runtime-worktree-agent-startup.ts
index 139679ba705..26fd683074f 100644
--- a/src/main/runtime/runtime-worktree-agent-startup.ts
+++ b/src/main/runtime/runtime-worktree-agent-startup.ts
@@ -10,7 +10,10 @@ import { isTuiAgentEnabled, pickTuiAgent } from '../../shared/tui-agent-selectio
 import { resolveAgentStartupPlanInputs } from '../../shared/agent-startup-plan-inputs'
 import { buildAgentDraftLaunchPlan, buildAgentStartupPlan } from '../../shared/tui-agent-startup'
 import { planStartupWithPromptCandidate } from '../../shared/startup-line-prompt-carry'
-import { nameLocalTypedLineShell } from './agent-launch-typed-line-shell'
+import {
+  launchHostProvesAgentInFront,
+  nameLocalTypedLineShell
+} from './agent-launch-typed-line-shell'
 import {
   detectInstalledAgentsWithShellPathHydration,
   detectRemoteAgents
@@ -153,16 +156,18 @@ export function buildWorktreeStartupForAgent(
   const prompt = environment.prompt ?? ''
   let startupPlan: ReturnType<typeof buildAgentStartupPlan>
   if (environment.onPromptCarry) {
-    const offered = planStartupWithPromptCandidate(
-      planInputs,
-      prompt,
-      nameLocalTypedLineShell({
+    const offered = planStartupWithPromptCandidate(planInputs, prompt, {
+      shellName: nameLocalTypedLineShell({
         isRemote: repoIsRemote(repo),
         ...(settings.terminalDefaultShell
           ? { defaultShellSetting: settings.terminalDefaultShell }
           : {})
+      }),
+      provesAgentInFront: launchHostProvesAgentInFront({
+        isRemote: repoIsRemote(repo),
+        launchPlatform: environment.getLaunchPlatform()
       })
-    )
+    })
     startupPlan = offered.plan
     if (startupPlan && prompt.trim()) {
       environment.onPromptCarry(offered.promptCarried)
diff --git a/src/shared/startup-line-prompt-carry.test.ts b/src/shared/startup-line-prompt-carry.test.ts
index 360fd2b484c..4520b3635d1 100644
--- a/src/shared/startup-line-prompt-carry.test.ts
+++ b/src/shared/startup-line-prompt-carry.test.ts
@@ -12,16 +12,24 @@ import { AGENT_LAUNCH_PROMPT_CARRY_RUNTIME_CAPABILITY } from './agent-launch-run
 function offer(
   agent: TuiAgent,
   prompt: string,
-  extra: { cmdOverride?: string; shellName?: string | undefined } = {}
+  extra: {
+    cmdOverride?: string
+    shellName?: string | undefined
+    platform?: NodeJS.Platform
+    provesAgentInFront?: boolean
+  } = {}
 ) {
   return planStartupWithPromptCandidate(
     {
       agent,
       cmdOverrides: extra.cmdOverride ? { [agent]: extra.cmdOverride } : {},
-      platform: 'darwin'
+      platform: extra.platform ?? 'darwin'
     },
     prompt,
-    extra.shellName
+    {
+      ...(extra.shellName ? { shellName: extra.shellName } : {}),
+      provesAgentInFront: extra.provesAgentInFront ?? true
+    }
   )
 }
 
@@ -173,6 +181,24 @@ describe('a multi-line prompt typed into a shell the host names', () => {
   })
 })
 
+// Why: a host that cannot prove the agent took the terminal could paste into the shell of one that
+// exited, so there the line carries the prompt whatever its size, as it did before.
+describe('on a host that cannot prove the launched agent is in front', () => {
+  it.each([
+    ['a long single line', 'x'.repeat(TYPED_STARTUP_LINE_PROMPT_BUDGET_BYTES * 4)],
+    ['a multi-line prompt', linesOf(5, 40)]
+  ])('carries %s on the launch line', (_label, prompt) => {
+    const { plan, promptCarried } = offer('claude', prompt, {
+      platform: 'win32',
+      provesAgentInFront: false
+    })
+    expect(promptCarried).toBe(true)
+    expect(plan?.launchCommand).toContain(prompt.split('\n')[0])
+    // Control: the same prompt is pasted where the host can prove the agent.
+    expect(offer('claude', prompt, { platform: 'win32' }).promptCarried).toBe(false)
+  })
+})
+
 describe('the capability clients gate a prompted launch on', () => {
   it('is advertised by every host that applies the typed-line rule', () => {
     // An older host folds any prompt into the typed line, so its absence is the gate.
diff --git a/src/shared/startup-line-prompt-carry.ts b/src/shared/startup-line-prompt-carry.ts
index 65e70e2d11d..e808cb9c74d 100644
--- a/src/shared/startup-line-prompt-carry.ts
+++ b/src/shared/startup-line-prompt-carry.ts
@@ -1,6 +1,7 @@
 /**
  * Whether a launch prompt rides the command line that gets TYPED into the user's shell, or the agent
- * starts clean and the prompt is pasted once it is ready.
+ * starts clean and the prompt is pasted once it is ready. A paste needs a host that can prove the
+ * agent holds its terminal (`launched-agent-foreground`); elsewhere the line carries it, as on main.
  *
  * Measured on the built line, not the raw prompt: quoting, the launcher, its arguments and session
  * options all land on that line, and every failure a long or multi-line typed line has is a property
@@ -37,11 +38,18 @@ export function startupLineCarriesPrompt(args: {
   withPrompt: AgentStartupPlan | null
   /** The shell the line is typed into, when the host can name it before the spawn. */
   shellName?: string
+  /** Whether the host can prove the launched agent holds its terminal before it pastes. */
+  hostProvesAgentInFront: boolean
 }): boolean {
   const { withPrompt } = args
   if (!withPrompt || withPrompt.followupPrompt !== null) {
     return false
   }
+  // Where nothing can prove the agent took the terminal, a paste could land in the shell of an
+  // agent that exited, so the line carries the prompt at any size, as it always did.
+  if (!args.hostProvesAgentInFront) {
+    return true
+  }
   // Hermes types a fixed line that reads the prompt from the spawn env, so the line never grows
   // with the text; its own env budget already returned null above when the text did not fit.
   if (TUI_AGENT_CONFIG[args.agent].promptInjectionMode === 'hermes-query') {
@@ -93,7 +101,7 @@ type StartupPlanInputs = Omit<
 export function planStartupWithPromptCandidate(
   inputs: StartupPlanInputs,
   prompt: string,
-  shellName?: string
+  host: { shellName?: string; provesAgentInFront: boolean }
 ): { plan: AgentStartupPlan | null; promptCarried: boolean } {
   if (prompt.trim()) {
     const withPrompt = buildAgentStartupPlan({ ...inputs, prompt, allowEmptyPromptLaunch: true })
@@ -101,7 +109,8 @@ export function planStartupWithPromptCandidate(
       startupLineCarriesPrompt({
         agent: inputs.agent,
         withPrompt,
-        ...(shellName ? { shellName } : {})
+        ...(host.shellName ? { shellName: host.shellName } : {}),
+        hostProvesAgentInFront: host.provesAgentInFront
       })
     ) {
       return { plan: withPrompt, promptCarried: true }