diff --git a/src/renderer/src/components/terminal-pane/agent-completion-positive-exit-requirement.test.ts b/src/renderer/src/components/terminal-pane/agent-completion-positive-exit-requirement.test.ts
new file mode 100644
index 00000000000..e49824d1502
--- /dev/null
+++ b/src/renderer/src/components/terminal-pane/agent-completion-positive-exit-requirement.test.ts
@@ -0,0 +1,129 @@
+import { describe, expect, it, vi } from 'vitest'
+import { createAgentCompletionCoordinator } from './agent-completion-coordinator'
+import { useAgentCompletionCoordinatorLifecycle } from './agent-completion-coordinator-test-harness'
+import type { RuntimeTerminalProcessInspection } from '@/runtime/runtime-terminal-inspection'
+
+// Pins the positive-exited arm on its own, independently of the
+// out-of-vocabulary guard.
+//
+// Given the vocabulary guard, every sample that reaches the exit gate already
+// carries a children verdict of 'live' or 'exited'. Against a host that builds
+// its wire result with buildPtyProcessInspectionWireResult the legacy
+// `hasChildProcesses` boolean is exactly `children.verdict === 'live'`, so on
+// that host reading the boolean and requiring a positive 'exited' agree — the
+// arm is invisible.
+//
+// They part company on the one shape the reader is explicitly written to
+// survive: a peer this client cannot vouch for, whose two fields disagree.
+// `readPtyProcessInspectionEvidence` never cross-checks the legacy scalar
+// against the evidence union, so `hasChildProcesses: false` can arrive beside a
+// positively observed `children: 'live'`. The boolean's `false` is the lossy
+// collapse that conflates "no children" with "could not tell" — the polarity
+// bug from #16900/#16908. The evidence union is the authoritative vocabulary,
+// so `live` must refuse completion even though the legacy field would wave it
+// through. No mock is needed: this shape flows through the real reader.
+function liveChildrenWithLegacyFalse(
+  foregroundProcess: string | null
+): RuntimeTerminalProcessInspection {
+  return {
+    foregroundProcess,
+    // The lossy legacy collapse disagrees with the authoritative evidence.
+    hasChildProcesses: false,
+    processEvidence: {
+      foreground: { verdict: 'observed', processName: foregroundProcess },
+      children: { verdict: 'live' }
+    }
+  }
+}
+
+function conforming(
+  foregroundProcess: string | null,
+  childrenVerdict: 'live' | 'exited'
+): RuntimeTerminalProcessInspection {
+  return {
+    foregroundProcess,
+    hasChildProcesses: childrenVerdict === 'live',
+    processEvidence: {
+      foreground: { verdict: 'observed', processName: foregroundProcess },
+      children: { verdict: childrenVerdict }
+    }
+  }
+}
+
+describe('agent completion requires a positively matched exited verdict', () => {
+  useAgentCompletionCoordinatorLifecycle()
+
+  function startCoordinator(results: () => RuntimeTerminalProcessInspection) {
+    const dispatchCompletion = vi.fn()
+    const coordinator = createAgentCompletionCoordinator({
+      paneKey: 'tab-1:leaf-1',
+      getPtyId: () => 'pty-1',
+      getSettings: () => null,
+      inspectProcess: vi.fn(async () => results()),
+      dispatchCompletion,
+      isLive: () => true
+    })
+    coordinator.startProcessTracking()
+    return { coordinator, dispatchCompletion }
+  }
+
+  it('never completes while children are positively live, even when the legacy field says otherwise', async () => {
+    let result = conforming('codex', 'live')
+    const { dispatchCompletion } = startCoordinator(() => result)
+
+    await vi.advanceTimersByTimeAsync(2_000)
+
+    // Foreground is observed and inside the vocabulary, and so is the children
+    // verdict — the vocabulary guard has nothing to refuse here. The only arm
+    // that can still refuse is the one that demands a positive 'exited', and
+    // it must, because the host positively observed live children.
+    result = liveChildrenWithLegacyFalse('zsh')
+    await vi.advanceTimersByTimeAsync(60_000)
+
+    expect(dispatchCompletion).not.toHaveBeenCalled()
+  })
+
+  it('still completes once the same host positively reports exited', async () => {
+    let result = conforming('codex', 'live')
+    const { dispatchCompletion } = startCoordinator(() => result)
+
+    await vi.advanceTimersByTimeAsync(2_000)
+
+    result = liveChildrenWithLegacyFalse('zsh')
+    await vi.advanceTimersByTimeAsync(60_000)
+    expect(dispatchCompletion).not.toHaveBeenCalled()
+
+    // Refusing the disagreeing sample must not wedge the monitor.
+    result = conforming('zsh', 'exited')
+    await vi.advanceTimersByTimeAsync(30_000)
+
+    expect(dispatchCompletion).toHaveBeenCalledTimes(1)
+    expect(dispatchCompletion).toHaveBeenCalledWith('codex', {
+      source: 'process-exit',
+      quietedHookDone: false,
+      terminalIdleConfirmed: true
+    })
+  })
+
+  it('treats a positively live sample as a readable inspection, not a failed one', async () => {
+    // The vocabulary guard re-arms with an inspection-error backoff because the
+    // sample could not be read. A 'live' sample was read fine — it just is not
+    // an exit — so the exit arm must refuse without charging an error, or a
+    // long-running agent would slide down the error backoff for free.
+    let result = conforming('codex', 'live')
+    const { dispatchCompletion } = startCoordinator(() => result)
+
+    await vi.advanceTimersByTimeAsync(2_000)
+
+    result = liveChildrenWithLegacyFalse('zsh')
+    await vi.advanceTimersByTimeAsync(20_000)
+
+    // A single conforming exit sample confirms against the pending exit that
+    // the live samples left standing; had they been charged as errors, the
+    // pending exit would have been cleared and re-armed instead.
+    result = conforming('zsh', 'exited')
+    await vi.advanceTimersByTimeAsync(30_000)
+
+    expect(dispatchCompletion).toHaveBeenCalledTimes(1)
+  })
+})