From d56d875ea63ff5dd8b8ce64d45d2f5458dc047fa Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 22 May 2026 22:04:19 +0000
Subject: [PATCH] fix(vertex realtime): warn when dropping guardrail
 turn-detection update

In non-deferred mode the auto-setup is sent on connect, so the audio-transcription
guardrail's subsequent session.update carrying turn_detection.create_response=False
cannot be forwarded as a second setup (Vertex Live closes the WebSocket with 1007).
Surface a warning when this specific drop happens so operators know the model
will auto-respond before the guardrail can gate it, instead of failing silently
at debug level.
---
 .../llms/vertex_ai/realtime/transformation.py | 24 +++++++-
 .../test_vertex_ai_realtime_transformation.py | 58 +++++++++++++++++++
 2 files changed, 80 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/vertex_ai/realtime/transformation.py b/litellm/llms/vertex_ai/realtime/transformation.py
index d2d89a52c8d..ea4dbccc8c8 100644
--- a/litellm/llms/vertex_ai/realtime/transformation.py
+++ b/litellm/llms/vertex_ai/realtime/transformation.py
@@ -226,9 +226,29 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig):
                 )
                 return [gemini_setup_msg]
 
-            verbose_logger.debug(
-                "Vertex AI Realtime: Ignoring session.update (setup already sent)"
+            # A follow-up session.update can't be forwarded as a second setup
+            # (Vertex Live closes the WebSocket with 1007). If this drop is
+            # silencing the audio-transcription guardrail's create_response
+            # disable, surface a warning so operators know the model will
+            # auto-respond before the guardrail can gate it on Vertex AI.
+            client_turn_detection = GeminiRealtimeConfig._extract_turn_detection(
+                json_message.get("session") or {}
             )
+            if (
+                isinstance(client_turn_detection, dict)
+                and client_turn_detection.get("create_response") is False
+            ):
+                verbose_logger.warning(
+                    "Vertex AI Realtime: Dropping subsequent session.update "
+                    "(turn_detection.create_response=False) — Vertex Live "
+                    "rejects a second setup message. Audio-transcription "
+                    "guardrails cannot suppress the model's auto-response on "
+                    "Vertex AI in non-deferred mode."
+                )
+            else:
+                verbose_logger.debug(
+                    "Vertex AI Realtime: Ignoring session.update (setup already sent)"
+                )
             return []
 
         return super().transform_realtime_request(
diff --git a/tests/test_litellm/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py b/tests/test_litellm/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py
index 8917d460b84..0ad614099de 100644
--- a/tests/test_litellm/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py
+++ b/tests/test_litellm/llms/vertex_ai/realtime/test_vertex_ai_realtime_transformation.py
@@ -288,3 +288,61 @@ async def test_vertex_realtime_text_in_text_out():
     # response.done should have been forwarded
     done_msgs = [m for m in sent_to_client if '"response.done"' in m]
     assert done_msgs, "Expected response.done to be sent to client"
+
+
+def test_vertex_warns_when_dropping_guardrail_turn_detection_update(caplog):
+    """A subsequent session.update carrying the guardrail's
+    ``create_response: False`` cannot be forwarded as a follow-up setup on
+    Vertex AI (1007). Surface a warning so operators know the auto-response
+    suppression is being silently dropped."""
+    import logging
+
+    cfg = VertexAIRealtimeConfig(
+        access_token="tok", project="my-proj", location="us-central1"
+    )
+
+    session_update = {
+        "type": "session.update",
+        "session": {"turn_detection": {"create_response": False}},
+    }
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        result = cfg.transform_realtime_request(
+            json.dumps(session_update),
+            "gemini-live-2.5-flash-native-audio",
+            session_configuration_request=json.dumps({"setup": {"model": "x"}}),
+        )
+
+    assert result == []
+    assert any(
+        "Vertex AI Realtime" in record.message
+        and "create_response=False" in record.message
+        for record in caplog.records
+    )
+
+
+def test_vertex_does_not_warn_when_dropping_non_guardrail_session_update(caplog):
+    """A subsequent session.update without ``create_response: False`` is a
+    routine drop and should stay at debug level (no warning)."""
+    import logging
+
+    cfg = VertexAIRealtimeConfig(
+        access_token="tok", project="my-proj", location="us-central1"
+    )
+
+    session_update = {
+        "type": "session.update",
+        "session": {"instructions": "Be concise."},
+    }
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        cfg.transform_realtime_request(
+            json.dumps(session_update),
+            "gemini-live-2.5-flash-native-audio",
+            session_configuration_request=json.dumps({"setup": {"model": "x"}}),
+        )
+
+    assert not any(
+        "Vertex AI Realtime" in record.message and "session.update" in record.message
+        for record in caplog.records
+    )