fix(realtime): address P1 issues - guardrail timing and inputAudioTranscription default

- Remove early guardrail turn-detection update that consumed first setup slot
- Add inputAudioTranscription default in Gemini deferred-mode setup
- Add tests for both fixes

Made-with: Cursor

litellm/llms/custom_httpx/llm_http_handler.py

@@ -5167,7 +5167,6 @@ class BaseLLMHTTPHandler:
                     if synthetic_session is not None:
                         await websocket.send_text(json.dumps(synthetic_session))
                         realtime_streaming._session_created_sent_to_client = True
-                        await realtime_streaming._maybe_send_guardrail_turn_detection_update()
                         verbose_logger.debug(
                             "Sent synthetic session.created to client to unblock connection"
                         )

litellm/llms/gemini/realtime/transformation.py

@@ -239,6 +239,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                 "generationConfig", {}
             )
             generation_config.setdefault("responseModalities", ["AUDIO"])
+            client_session_configuration_request.setdefault("inputAudioTranscription", {})
             client_session_configuration_request["model"] = f"models/{model}"
             gemini_setup_msg = json.dumps({"setup": client_session_configuration_request})
             verbose_logger.debug(

tests/test_litellm/litellm_core_utils/test_realtime_streaming.py

@@ -1080,3 +1080,40 @@ async def test_duplicate_session_created_still_triggers_guardrail_turn_detection
     sent_update = json.loads(streaming._send_to_backend.await_args_list[0].args[0])
     assert sent_update["type"] == "session.update"
     assert sent_update["session"]["turn_detection"]["create_response"] is False
+
+
+@pytest.mark.asyncio
+async def test_guardrail_update_respects_idempotency_flag():
+    """Verify guardrail turn-detection update uses idempotency flag correctly."""
+    client_ws = AsyncMock()
+    backend_ws = MagicMock()
+    backend_ws.send = AsyncMock()
+
+    logging_obj = MagicMock()
+    logging_obj.litellm_trace_id = "trace_1"
+    logging_obj.async_success_handler = AsyncMock()
+    logging_obj.success_handler = MagicMock()
+
+    provider_config = MagicMock()
+    provider_config.transform_realtime_request = MagicMock(
+        side_effect=lambda msg, model, session_config: [msg]
+    )
+
+    streaming = RealTimeStreaming(
+        websocket=client_ws,
+        backend_ws=backend_ws,
+        logging_obj=logging_obj,
+        provider_config=provider_config,
+        model="gemini-2.5-flash",
+    )
+    streaming._has_audio_transcription_guardrails = MagicMock(return_value=True)  # type: ignore[method-assign]
+    
+    # First call should send the update
+    assert streaming._guardrail_turn_detection_update_sent is False
+    await streaming._maybe_send_guardrail_turn_detection_update()
+    assert streaming._guardrail_turn_detection_update_sent is True
+    assert backend_ws.send.await_count == 1
+    
+    # Second call should be a no-op (idempotent)
+    await streaming._maybe_send_guardrail_turn_detection_update()
+    assert backend_ws.send.await_count == 1  # Still 1, not 2

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

@@ -584,3 +584,37 @@ def test_gemini_realtime_multi_tool_calls_have_unique_item_ids():
     assert responses[0]["item_id"] != responses[1]["item_id"]
     assert responses[0]["output_index"] == 0
     assert responses[1]["output_index"] == 1
+
+
+def test_gemini_session_update_includes_input_audio_transcription_default():
+    """Verify _handle_session_update includes inputAudioTranscription default."""
+    config = GeminiRealtimeConfig()
+    session_update = {
+        "type": "session.update",
+        "session": {
+            "modalities": ["text", "audio"],
+            "tools": [
+                {
+                    "type": "function",
+                    "name": "get_weather",
+                    "description": "Get weather",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"location": {"type": "string"}},
+                    },
+                }
+            ],
+        },
+    }
+
+    result = config.transform_realtime_request(
+        json.dumps(session_update),
+        "gemini-2.5-flash",
+        session_configuration_request=None,
+    )
+
+    assert len(result) == 1
+    setup = json.loads(result[0])
+    assert "setup" in setup
+    assert "inputAudioTranscription" in setup["setup"]
+    assert setup["setup"]["inputAudioTranscription"] == {}