diff --git a/openhands-sdk/openhands/sdk/llm/llm.py b/openhands-sdk/openhands/sdk/llm/llm.py
index 85d39409a0..640f7b7fa6 100644
--- a/openhands-sdk/openhands/sdk/llm/llm.py
+++ b/openhands-sdk/openhands/sdk/llm/llm.py
@@ -1856,9 +1856,14 @@ def _one_attempt(**retry_kwargs: Any) -> ResponsesAPIResponse:
                     if stream_callback is not None and delta_chunk is not None:
                         stream_callback(delta_chunk)
 
-                completed_response = getattr(
-                    ret, "completed_response", completed_response
-                )
+                # Keep the completion the stream yielded. A wrapper may also
+                # carry a `completed_response` attribute that is still None
+                # after iteration, and getattr's default does not apply when the
+                # attribute exists, so reading it back unconditionally discards
+                # a valid event and the call then raises LLMNoResponseError.
+                wrapper_completed = getattr(ret, "completed_response", None)
+                if wrapper_completed is not None:
+                    completed_response = wrapper_completed
                 return self._finalize_stream_response(
                     completed_response, collected_output_items
                 )
@@ -2033,9 +2038,14 @@ async def _one_attempt(
                         if stream_cb is not None and delta_chunk is not None:
                             await _invoke_token_callback(stream_cb, delta_chunk)
 
-                completed_response = getattr(
-                    ret, "completed_response", completed_response
-                )
+                # Keep the completion the stream yielded. A wrapper may also
+                # carry a `completed_response` attribute that is still None
+                # after iteration, and getattr's default does not apply when the
+                # attribute exists, so reading it back unconditionally discards
+                # a valid event and the call then raises LLMNoResponseError.
+                wrapper_completed = getattr(ret, "completed_response", None)
+                if wrapper_completed is not None:
+                    completed_response = wrapper_completed
                 return self._finalize_stream_response(
                     completed_response, collected_output_items
                 )
diff --git a/tests/sdk/llm/test_responses_parsing_and_kwargs.py b/tests/sdk/llm/test_responses_parsing_and_kwargs.py
index 296d521412..93af0dcac4 100644
--- a/tests/sdk/llm/test_responses_parsing_and_kwargs.py
+++ b/tests/sdk/llm/test_responses_parsing_and_kwargs.py
@@ -596,3 +596,72 @@ async def _events():
     assert [chunk.choices[0].delta.content for chunk in received] == [
         "Hello wrapped stream"
     ]
+
+
+class _StaleCompletedResponseStream:
+    """A streaming wrapper that yields the completion and also exposes a stale None.
+
+    Reproduces https://github.com/OpenHands/software-agent-sdk/issues/4769: the
+    attribute exists, so `getattr(ret, "completed_response", default)` returns
+    its None rather than the default, discarding the event the stream yielded.
+    """
+
+    def __init__(self, events):
+        self._events = events
+        self.completed_response = None
+
+    def __iter__(self):
+        return iter(self._events)
+
+
+@patch("openhands.sdk.llm.llm.litellm_responses")
+def test_responses_streaming_keeps_the_yielded_completion(mock_responses):
+    events, completed_response = _make_wrapped_response_stream_events()
+    mock_responses.return_value = _StaleCompletedResponseStream(events)
+
+    llm = LLM(
+        model="gpt-4o",
+        api_key=SecretStr("test_key"),
+        usage_id="test-llm",
+        num_retries=1,
+        retry_min_wait=1,
+        retry_max_wait=2,
+    )
+
+    received = []
+    response = llm.responses(
+        [Message(role="user", content=[TextContent(text="Hello")])],
+        stream=True,
+        on_token=received.append,
+    )
+
+    assert response.raw_response is completed_response
+
+
+@pytest.mark.asyncio
+@patch("openhands.sdk.llm.llm.litellm_aresponses", new_callable=AsyncMock)
+async def test_aresponses_streaming_keeps_the_yielded_completion(mock_aresponses):
+    events, completed_response = _make_wrapped_response_stream_events()
+
+    def _return_stale_wrapper(*args, **kwargs):
+        return _StaleCompletedResponseStream(events)
+
+    mock_aresponses.side_effect = _return_stale_wrapper
+
+    llm = LLM(
+        model="gpt-4o",
+        api_key=SecretStr("test_key"),
+        usage_id="test-llm",
+        num_retries=1,
+        retry_min_wait=1,
+        retry_max_wait=2,
+    )
+
+    received = []
+    response = await llm.aresponses(
+        [Message(role="user", content=[TextContent(text="Hello")])],
+        stream=True,
+        on_token=received.append,
+    )
+
+    assert response.raw_response is completed_response