diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py
index 7995b59f11..24ebd0edbb 100644
--- a/agent/conversation_loop.py
+++ b/agent/conversation_loop.py
@@ -8512,14 +8512,18 @@ def run_conversation(
                     _turn_exit_reason = f"local_processing_error({error_msg[:80]})"
                     final_response = f"I apologize, but I encountered an error while processing the model response: {error_msg}"
                 elif _outer_error_count >= _outer_error_cap:
+                    failed = True
                     _turn_exit_reason = f"repeated_outer_errors({error_msg[:80]})"
                     final_response = f"I apologize, but I encountered repeated errors: {error_msg}"
                 else:
                     _turn_exit_reason = f"error_near_max_iterations({error_msg[:80]})"
                     final_response = f"I apologize, but I encountered repeated errors: {error_msg}"
-                # Append as assistant so the history stays valid for
-                # session resume (avoids consecutive user messages).
-                append_message(messages, {"role": "assistant", "content": final_response})
+                # Don't append the assistant message here — a thinking-prefill
+                # or interim assistant may already be the tail, and appending
+                # would create assistant→assistant.  finalize_turn handles
+                # this case safely (lines 341-353: appends only when
+                # _tail_role != "assistant"), matching the pattern at other
+                # break sites that set final_response without appending.
                 break
     
     # Post-loop turn finalization extracted to agent/turn_finalizer.finalize_turn
diff --git a/tests/run_agent/test_92450_outer_error_retry_bound.py b/tests/run_agent/test_92450_outer_error_retry_bound.py
index 240a56d0d0..17e277da81 100644
--- a/tests/run_agent/test_92450_outer_error_retry_bound.py
+++ b/tests/run_agent/test_92450_outer_error_retry_bound.py
@@ -103,9 +103,9 @@ class TestOuterErrorRetryBound:
             result = loop_agent.run_conversation("hello")
 
         # The bound path returns an apology as final_response with
-        # failed=False (same contract as the legacy near-limit exit), so the
+        # failed=True (the turn did not complete successfully), so the
         # meaningful assertions are: bounded call count + dedicated exit
-        # reason + apology text.
+        # reason + apology text + failed flag.
         assert result["api_calls"] <= 8, (
             "The loop must give up within the per-turn error bound instead "
             "of retrying an unlimited-budget failure forever."
@@ -114,6 +114,8 @@ class TestOuterErrorRetryBound:
             f"unexpected exit reason: {result['turn_exit_reason']}"
         )
         assert "repeated errors" in (result["final_response"] or "")
+        assert result["failed"] is True
+        assert result["completed"] is False
 
     def test_recovery_after_failures_completes_normally(self, loop_agent):
         """Early escaping failures followed by success must NOT trip the new