Fix greptile review

litellm/llms/gemini/realtime/transformation.py

@@ -667,6 +667,8 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
     def transform_tool_call_events(
         self,
         tool_call_message: dict,
+        response_id: Optional[str] = None,
+        output_item_id: Optional[str] = None,
     ) -> List[Dict[str, Any]]:
         """
         Transform Gemini toolCall message to OpenAI function call events.
@@ -674,6 +676,8 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
         Converts Gemini's functionCalls format to OpenAI's response.function_call_arguments.done events.
         """
         function_calls = tool_call_message.get("functionCalls", [])
+        resolved_response_id = response_id or f"resp_{uuid.uuid4()}"
+        resolved_output_item_id = output_item_id or f"item_{uuid.uuid4()}"
         
         verbose_logger.debug(
             f"Gemini Realtime: Transforming {len(function_calls)} tool call(s) to OpenAI format"
@@ -683,11 +687,14 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
             {
                 "type": "response.function_call_arguments.done",
                 "event_id": f"event_{uuid.uuid4()}",
+                "response_id": resolved_response_id,
+                "item_id": resolved_output_item_id,
+                "output_index": idx,
                 "call_id": fc.get("id", ""),
                 "name": fc.get("name", ""),
                 "arguments": json.dumps(fc.get("args", {})),
             }
-            for fc in function_calls
+            for idx, fc in enumerate(function_calls)
         ]
 
     @staticmethod
@@ -1067,7 +1074,11 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                 returned_message.append(transformed_message)
             elif openai_event == ResponsesAPIStreamEvents.FUNCTION_CALL_ARGUMENTS_DONE:
                 # Handle toolCall from Gemini
-                tool_call_events = self.transform_tool_call_events(value)
+                tool_call_events = self.transform_tool_call_events(
+                    value,
+                    response_id=current_response_id,
+                    output_item_id=current_output_item_id,
+                )
                 returned_message.extend(tool_call_events)
             elif openai_event == OpenAIRealtimeEventTypes.RESPONSE_DONE:
                 transformed_response_done_event = self.transform_response_done_event(

tests/test_litellm/litellm_core_utils/test_realtime_streaming.py

@@ -254,6 +254,50 @@ async def test_transcription_captured_in_backend_to_client():
     assert logging_obj.model_call_details["messages"] == streaming.input_messages
 
 
+@pytest.mark.asyncio
+async def test_client_ack_caches_setup_to_prevent_duplicate_session_update_setup():
+    websocket = MagicMock()
+    backend_ws = MagicMock()
+    logging_obj = MagicMock()
+    logging_obj.pre_call = MagicMock()
+
+    # Two session.update messages arrive before setupComplete round-trip.
+    websocket.receive_text = AsyncMock(
+        side_effect=[
+            json.dumps({"type": "session.update", "session": {"tools": []}}),
+            json.dumps({"type": "session.update", "session": {"tools": []}}),
+            Exception("client done"),
+        ]
+    )
+
+    provider_config = MagicMock()
+
+    def _transform(message: str, model: str, session_configuration_request=None):
+        if session_configuration_request is None:
+            return [json.dumps({"setup": {"model": "models/gemini-2.5-flash"}})]
+        return []
+
+    provider_config.transform_realtime_request = MagicMock(side_effect=_transform)
+
+    backend_ws.send = AsyncMock()
+
+    streaming = RealTimeStreaming(
+        websocket=websocket,
+        backend_ws=backend_ws,
+        logging_obj=logging_obj,
+        provider_config=provider_config,
+        model="gemini-2.5-flash",
+    )
+
+    await streaming.client_ack_messages()
+
+    # Setup should be forwarded exactly once even with repeated session.update.
+    assert backend_ws.send.await_count == 1
+    assert streaming.session_configuration_request is not None
+    sent_payload = json.loads(backend_ws.send.await_args_list[0].args[0])
+    assert "setup" in sent_payload
+
+
 def test_collect_session_tools_from_session_update():
     """
     Test that tools from session.update events are collected.

tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py

@@ -297,6 +297,9 @@ def test_gemini_realtime_tool_call_transformation():
     assert function_call_event is not None, "Expected function_call_arguments.done event"
     assert function_call_event["call_id"] == "call_123"
     assert function_call_event["name"] == "get_weather"
+    assert function_call_event["response_id"] == "resp_123"
+    assert function_call_event["item_id"] == "item_123"
+    assert function_call_event["output_index"] == 0
     
     # Verify arguments are properly serialized as JSON string
     args = json.loads(function_call_event["arguments"])
@@ -454,3 +457,18 @@ def test_gemini_realtime_user_text_transformation():
     assert len(turn["parts"]) == 1
     assert turn["parts"][0]["text"] == "What's the weather in London?"
     assert client_content["turnComplete"] is True
+
+
+def test_return_new_content_delta_events_without_session_config_does_not_error():
+    config = GeminiRealtimeConfig()
+
+    events = config.return_new_content_delta_events(
+        response_id="resp_1",
+        output_item_id="item_1",
+        conversation_id="conv_1",
+        delta_type="text",
+        session_configuration_request=None,
+    )
+
+    assert len(events) >= 1
+    assert events[0]["type"] == "response.created"

tests/test_litellm/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py

@@ -97,6 +97,29 @@ def test_vertex_requires_session_configuration_feature_flag(monkeypatch):
     assert cfg.requires_session_configuration() is False
 
 
+def test_vertex_session_update_defaults_to_audio_modality():
+    cfg = VertexAIRealtimeConfig(
+        access_token="tok", project="my-proj", location="us-central1"
+    )
+
+    session_update = {
+        "type": "session.update",
+        "session": {
+            "instructions": "You are a helpful assistant.",
+            # No modalities provided on purpose
+        },
+    }
+
+    messages = cfg.transform_realtime_request(
+        json.dumps(session_update),
+        "gemini-live-2.5-flash-native-audio",
+        session_configuration_request=None,
+    )
+    assert len(messages) == 1
+    setup_payload = json.loads(messages[0])["setup"]
+    assert setup_payload["generationConfig"]["responseModalities"] == ["AUDIO"]
+
+
 # ---------------------------------------------------------------------------
 # Round-trip test: text-in / text-out via RealTimeStreaming
 # ---------------------------------------------------------------------------