diff --git a/src/servers/sidecar/claude/claude-manager.ts b/src/servers/sidecar/claude/claude-manager.ts
index 147e0f55..b3842bc0 100644
--- a/src/servers/sidecar/claude/claude-manager.ts
+++ b/src/servers/sidecar/claude/claude-manager.ts
@@ -6,7 +6,7 @@ import { query, type Query, type SpawnOptions, type SpawnedProcess } from '@anth
 import type { ChatEvent, PromptImage } from '../../api/chat/types';
 import type { ClaudeSpawnParams, ClaudeSpawnStreamingParams, ClaudeCodeResult, LiveClaudeSession } from '../protocol';
 import { setClaudeSession, clearClaudeSession, getClaudeSession } from './state';
-import { diagnoseExit, stderrReason } from './exit-diagnosis';
+import { countsAsOutput, diagnoseExit, stderrReason } from './exit-diagnosis';
 import { createParseState, processMessage } from './stream-parser';
 import { spawnClaudeAsMember } from './spawn-as-member';
 import { claudeBinIn } from '@@/os-user-claude';
@@ -620,13 +620,12 @@ function createSession(params: ClaudeSpawnStreamingParams, onEvent: (event: Chat
       // Any event at all is a sign of life, so push the stall deadline back. Placed after the branch
       // above so a terminal event disarms rather than re-arms it.
       armStall(session);
+      // See `countsAsOutput`: an error is the one event that does not prove the harness was working.
+      if (countsAsOutput(event.type)) session.sawOutput = true;
       session.emit(event);
     };
     try {
       for await (const msg of q as AsyncGenerator<Record<string, unknown>>) {
-        // One message — any message — means the child got far enough to talk. That distinguishes a resume
-        // that could not even load from a turn that broke halfway, which want opposite handling.
-        session.sawOutput = true;
         processMessage(msg, state, {
           onEvent: emit,
           onSessionId: (id: string) => setClaudeSession(sessionKey, id, params.userId),
diff --git a/src/servers/sidecar/claude/exit-diagnosis.test.ts b/src/servers/sidecar/claude/exit-diagnosis.test.ts
index 14760e77..6714bd90 100644
--- a/src/servers/sidecar/claude/exit-diagnosis.test.ts
+++ b/src/servers/sidecar/claude/exit-diagnosis.test.ts
@@ -1,5 +1,5 @@
 import { describe, expect, test } from 'bun:test';
-import { diagnoseExit, stderrReason } from './exit-diagnosis';
+import { countsAsOutput, diagnoseExit, stderrReason } from './exit-diagnosis';
 
 // The bug these encode: a member's chat answered "Claude Code process exited with code 1" four times in a
 // row and could never recover, because every message retried the same `--resume` id and nothing ever
@@ -55,3 +55,39 @@ describe('stderrReason', () => {
     expect(stderrReason('x'.repeat(5000)).length).toBeLessThanOrEqual(300);
   });
 });
+
+// ── What counts as the harness having worked ──
+//
+// These exist because the first version of this rule was wrong in the only case it was written for, and
+// nothing failed: a broken resume emits a terminal error result, that was counted as output, and the
+// recovery it was supposed to trigger could never fire.
+
+describe('countsAsOutput', () => {
+  test('an error does NOT count — it is what a failed resume emits', () => {
+    expect(countsAsOutput('error')).toBe(false);
+  });
+
+  test('real work counts', () => {
+    for (const type of ['assistant:text', 'assistant:delta', 'tool:start', 'tool:result', 'result']) {
+      expect(countsAsOutput(type)).toBe(true);
+    }
+  });
+
+  test('a stopped turn counts — it was running, so the conversation loaded', () => {
+    // Clearing someone's transcript because they pressed stop quickly would be its own bug.
+    expect(countsAsOutput('stopped')).toBe(true);
+  });
+
+  // The end-to-end shape of the reported failure: no real output, an error result, a resuming turn.
+  test('the reported failure now resolves to "give up on the resume"', () => {
+    const events = ['error'];
+    const sawOutput = events.some(countsAsOutput);
+    expect(diagnoseExit({ sawOutput, resumeId: '49527b33', stderrTail: '' }).clearResume).toBe(true);
+  });
+
+  test('and a turn that spoke before failing does not', () => {
+    const events = ['assistant:text', 'error'];
+    const sawOutput = events.some(countsAsOutput);
+    expect(diagnoseExit({ sawOutput, resumeId: '49527b33', stderrTail: '' }).clearResume).toBe(false);
+  });
+});
diff --git a/src/servers/sidecar/claude/exit-diagnosis.ts b/src/servers/sidecar/claude/exit-diagnosis.ts
index d30df2d1..29726f63 100644
--- a/src/servers/sidecar/claude/exit-diagnosis.ts
+++ b/src/servers/sidecar/claude/exit-diagnosis.ts
@@ -17,6 +17,24 @@
  * turn that broke later. Clearing the resume on the second would trade one bad turn for a lost
  * conversation, which is the wrong direction.
  */
+/**
+ * Whether an event proves the harness was actually working.
+ *
+ * This is the flag's definition and the first version got it backwards in exactly the case it exists for.
+ * It counted ANY SDK message as output — but a turn that cannot load its conversation still emits a
+ * terminal `result` with `is_error` and no text, which is what reaches the user as "Claude Code returned an
+ * error". So the very event that proves the resume failed was being read as proof it succeeded, and the
+ * recovery could never fire. Caught by tracing the two error strings in a screenshot to their sources
+ * rather than by testing, which is luck; hence this being a named function with tests rather than an
+ * inline `!==`.
+ *
+ * `stopped` DOES count. A turn someone interrupted was running, which means the conversation loaded — and
+ * discarding the transcript because somebody pressed stop quickly would be its own bug.
+ */
+export function countsAsOutput(eventType: string): boolean {
+  return eventType !== 'error';
+}
+
 export type ExitDiagnosis = {
   /** Forget the stored transcript id, so the next message starts a new session instead of failing again. */
   clearResume: boolean;