From 154aeeba27cff0716c6a8b61808b55106eff0d6d Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 16 Jul 2026 15:24:55 -0700
Subject: [PATCH] fix(gemini): tolerate generationComplete with no content
 deltas in realtime bridge

A Gemini Live generationComplete arriving before any modelTurn deltas (toolCall-only
or empty generations) raised ValueError('Unexpected delta type: None'), and the
caller dropped the entire backend frame. This fired on every gemini/vertex realtime
tool-call turn and intermittently stalled the vertex server-VAD path (LIT-4482).
map_openai_event now returns None for that case and the transform marks the frame
handled instead of raising.
---
 .../llms/gemini/realtime/transformation.py    | 13 +++++++-
 .../test_gemini_realtime_transformation.py    | 31 +++++++++++++++++++
 2 files changed, 43 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/gemini/realtime/transformation.py b/litellm/llms/gemini/realtime/transformation.py
index bc2145fd832..ccf759930ec 100644
--- a/litellm/llms/gemini/realtime/transformation.py
+++ b/litellm/llms/gemini/realtime/transformation.py
@@ -1103,7 +1103,11 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
         key: str,
         value: Any,
         current_delta_type: Optional[ALL_DELTA_TYPES],
-    ) -> Union[OpenAIRealtimeEventTypes, ResponsesAPIStreamEvents]:
+    ) -> Union[OpenAIRealtimeEventTypes, ResponsesAPIStreamEvents, None]:
+        """Map a Gemini top-level key to the OpenAI event it produces. Returns None
+        for a ``generationComplete`` with no preceding content deltas (a toolCall-only
+        or empty generation): there is no text/audio "done" to emit for it, and
+        raising here used to make the caller drop the entire backend frame."""
         if isinstance(value, dict):
             model_turn_event = value.get("modelTurn")
             generation_complete_event = value.get("generationComplete")
@@ -1114,6 +1118,8 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
         if model_turn_event:  # check if model turn event
             openai_event = self.map_model_turn_event(model_turn_event)
         elif generation_complete_event:
+            if current_delta_type is None:
+                return None
             openai_event = self.map_generation_complete_event(delta_type=current_delta_type)
         else:
             # Check if this key or any nested key matches our mapping. Use a
@@ -1237,6 +1243,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
             server_content_handled = False
 
         tool_call_handled = False
+        generation_complete_handled = False
         for key, value in list(json_message.items()):  # snapshot: handlers may mutate json_message
             # Skip sibling metadata keys (e.g. ``usageMetadata``) that can
             # accompany a primary payload like ``toolCall`` or ``serverContent``.
@@ -1255,6 +1262,9 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                 value=value,
                 current_delta_type=current_delta_type,
             )
+            if openai_event is None:
+                generation_complete_handled = True
+                continue
 
             if openai_event == OpenAIRealtimeEventTypes.SESSION_CREATED:
                 transformed_message = self.transform_session_created_event(
@@ -1498,6 +1508,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                 for key in json_message
                 if key in _KNOWN_GEMINI_TOP_LEVEL_KEYS
                 and not (key == "serverContent" and server_content_handled)
+                and not (key == "serverContent" and generation_complete_handled)
                 and not (key == "toolCall" and tool_call_handled)
             ]
             standalone_usage_metadata = json_message.get("usageMetadata")
diff --git a/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py b/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py
index 02004d5c8a8..3c738dbc8b0 100644
--- a/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py
+++ b/tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py
@@ -311,6 +311,37 @@ def test_gemini_realtime_transformation_generation_complete():
     assert contains_audio_done_event, "Expected audio done event"
 
 
+def test_gemini_realtime_transformation_generation_complete_without_deltas():
+    config = GeminiRealtimeConfig()
+
+    session_configuration_request_str = json.dumps(
+        {
+            "setup": {
+                "model": "gemini-1.5-flash",
+                "generationConfig": {"responseModalities": ["AUDIO"]},
+            }
+        }
+    )
+
+    result = config.transform_realtime_response(
+        json.dumps({"serverContent": {"generationComplete": True}}),
+        "gemini-1.5-flash",
+        MagicMock(),
+        realtime_response_transform_input={
+            "session_configuration_request": session_configuration_request_str,
+            "current_output_item_id": None,
+            "current_response_id": None,
+            "current_conversation_id": None,
+            "current_delta_chunks": [],
+            "current_item_chunks": [],
+            "current_delta_type": None,
+        },
+    )
+
+    assert result["response"] == []
+    assert result["current_delta_type"] is None
+
+
 def test_gemini_3_1_flash_live_preview_model_cost_map_entry():
     for key in (
         "gemini-3.1-flash-live-preview",