Merge 67bcf90068 into 82ca10cc2d

litellm/llms/gemini/realtime/transformation.py

@@ -391,7 +391,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
         return None
 
     @staticmethod
-    def _normalize_session_payload_for_mapping(session: dict) -> dict:
+    def _normalize_session_payload_for_mapping(session: dict) -> dict[str, object]:
         """Normalize GA-remapped session fields back to their beta keys.
 
         ``map_openai_params`` only recognises the flat OpenAI-beta key names
@@ -472,6 +472,39 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                 )
         return setup
 
+    def _warn_session_update_dropped(self, session_payload: Mapping[str, object], *, provider: str) -> None:
+        """Report, once per session, that a later ``session.update`` reached nothing.
+
+        Live accepts ``setup`` as the first-and-only client message, so a second one
+        closes the socket with 1007 and the update has to be dropped. That is
+        unavoidable; being quiet about it is not. This used to log at debug, which on
+        a proxy means the fields never arrive and nothing says so -- the caller sees
+        the model running on provider defaults and no instructions, voice or persona
+        of theirs (#44825).
+
+        Warned once because a GA client sends several ``session.update`` messages
+        while configuring; later drops stay at debug.
+        """
+        dropped: Final = sorted(key for key in session_payload if key != "type")
+        if getattr(self, "_session_update_drop_warned", False):
+            verbose_logger.debug(
+                "%s: dropping a further session.update (setup already sent): %s",
+                provider,
+                dropped,
+            )
+            return
+        self._session_update_drop_warned = True
+        verbose_logger.warning(
+            "%s: dropping session.update because setup was already sent, so these "
+            "fields do not reach the model: %s. Live accepts setup as the "
+            "first-and-only client message. To have the client's first "
+            "session.update carry them instead, set litellm.gemini_live_defer_setup "
+            "= True (env LITELLM_GEMINI_LIVE_DEFER_SETUP=true), which defers setup "
+            "until that update arrives. Further drops in this session log at debug.",
+            provider,
+            dropped,
+        )
+
     def _handle_session_update(
         self,
         json_message: _OpenAIRealtimeClientEvent,
@@ -527,7 +560,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                 "suppress the model's auto-response mid-session."
             )
         else:
-            verbose_logger.debug("Gemini Realtime: Ignoring session.update (setup already sent)")
+            self._warn_session_update_dropped(session_payload, provider="Gemini Realtime")
         return []
 
     def _handle_conversation_item(self, json_message: _OpenAIRealtimeClientEvent) -> list[str]:

litellm/llms/vertex_ai/realtime/transformation.py

@@ -17,7 +17,7 @@ Auth: OAuth2 Bearer token (not an API key).
 """
 
 import json
-from collections.abc import Awaitable, Callable
+from collections.abc import Awaitable, Callable, Mapping
 from typing import Final
 
 from litellm import verbose_logger
@@ -238,7 +238,10 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig):
                     "Vertex AI in non-deferred mode."
                 )
             else:
-                verbose_logger.debug("Vertex AI Realtime: Ignoring session.update (setup already sent)")
+                # Annotated rather than passed straight from json.loads, whose Any
+                # would reach the helper's Mapping parameter as an unknown type.
+                dropped_session: Final[Mapping[str, object]] = json_message.get("session") or {}
+                self._warn_session_update_dropped(dropped_session, provider="Vertex AI Realtime")
             return []
 
         return super().transform_realtime_request(message, model, session_configuration_request)

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

@@ -2235,3 +2235,34 @@ def test_gemini_realtime_response_done_reports_no_grounding_when_none_ran():
 
     assert input_details.get("web_search_requests") is None
     assert input_details.get("google_maps_grounding_requests") is None
+
+
+def test_gemini_warns_once_when_dropping_a_session_update(caplog):
+    """The Gemini path reports a dropped session.update the same way Vertex does.
+
+    Both drop it for the same reason -- Live accepts setup as the first-and-only
+    client message -- so both should say so rather than log at debug and let the
+    caller discover their instructions never arrived (#44825).
+    """
+    import logging
+
+    cfg = GeminiRealtimeConfig()
+    session_update = json.dumps({"type": "session.update", "session": {"instructions": "Be concise."}})
+    already_set_up = json.dumps({"setup": {"model": "x"}})
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM"):
+        for _ in range(2):
+            cfg.transform_realtime_request(
+                session_update,
+                "gemini-live-2.5-flash-preview-native-audio-09-2025",
+                session_configuration_request=already_set_up,
+            )
+
+    warnings = [
+        record
+        for record in caplog.records
+        if "Gemini Realtime" in record.message and "session.update" in record.message
+    ]
+    assert len(warnings) == 1
+    assert "instructions" in warnings[0].message
+    assert "gemini_live_defer_setup" in warnings[0].message

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

@@ -418,9 +418,19 @@ def test_vertex_warns_when_dropping_guardrail_turn_detection_update(caplog):
     )
 
 
-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)."""
+def test_vertex_warns_once_when_dropping_a_session_update(caplog):
+    """A dropped session.update is reported, and reported once.
+
+    This inverts what the test here previously asserted. It held that dropping a
+    subsequent session.update was "a routine drop" that should stay at debug --
+    with ``{"instructions": "Be concise."}`` as its own fixture, which is exactly
+    the case reported in #44825: the caller's instructions never reach the model
+    and nothing on a default proxy log says so.
+
+    The drop itself is unavoidable (Live rejects a second setup with 1007). Only
+    the silence changes. It warns once because a GA client sends several
+    session.updates while configuring; the rest stay at debug.
+    """
     import logging
 
     cfg = VertexAIRealtimeConfig(
@@ -429,20 +439,28 @@ def test_vertex_does_not_warn_when_dropping_non_guardrail_session_update(caplog)
 
     session_update = {
         "type": "session.update",
-        "session": {"instructions": "Be concise."},
+        "session": {"instructions": "Be concise.", "voice": "Puck"},
     }
+    already_set_up = json.dumps({"setup": {"model": "x"}})
 
     with caplog.at_level(logging.WARNING, logger="LiteLLM"):
-        cfg.transform_realtime_request(
-            json.dumps(session_update),
-            "gemini-live-2.5-flash-preview-native-audio-09-2025",
-            session_configuration_request=json.dumps({"setup": {"model": "x"}}),
-        )
+        for _ in range(3):
+            cfg.transform_realtime_request(
+                json.dumps(session_update),
+                "gemini-live-2.5-flash-preview-native-audio-09-2025",
+                session_configuration_request=already_set_up,
+            )
 
-    assert not any(
-        "Vertex AI Realtime" in record.message and "session.update" in record.message
+    warnings = [
+        record
         for record in caplog.records
-    )
+        if "Vertex AI Realtime" in record.message and "session.update" in record.message
+    ]
+    assert len(warnings) == 1, "three drops, one warning"
+    # the operator needs to know what was lost and what to do about it
+    assert "instructions" in warnings[0].message
+    assert "voice" in warnings[0].message
+    assert "gemini_live_defer_setup" in warnings[0].message
 
 
 @pytest.mark.asyncio