From b60cc950f3be4edc61e3a1f01f288192c4db811c Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 22 May 2026 21:44:43 +0000
Subject: [PATCH] fix(gemini realtime): mirror
 modalities/temperature/max_output_tokens on tool-call response.created

The audio/text response.created preamble includes modalities, temperature,
and max_output_tokens on the response object so spec-compliant clients can
initialise per-response state. The tool-call response.created was missing
these fields, leaving clients without consistent response metadata when a
response starts with a tool call instead of content. Read them from the
cached session_configuration_request the same way the audio/text path
does.
---
 .../llms/gemini/realtime/transformation.py    | 33 ++++++++++++++++++-
 .../test_gemini_realtime_transformation.py    |  6 ++++
 2 files changed, 38 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/gemini/realtime/transformation.py b/litellm/llms/gemini/realtime/transformation.py
index 84bdd6051a4..327eee5fd94 100644
--- a/litellm/llms/gemini/realtime/transformation.py
+++ b/litellm/llms/gemini/realtime/transformation.py
@@ -1251,7 +1251,31 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                     current_response_id = f"resp_{uuid.uuid4()}"
                     current_output_item_id = f"item_{uuid.uuid4()}"
 
-                    # Emit response.created
+                    # Mirror the audio/text path: include modalities,
+                    # temperature, and max_output_tokens on response.created so
+                    # spec-compliant clients see consistent response metadata
+                    # regardless of whether the response starts with content or
+                    # a tool call.
+                    session_setup: BidiGenerateContentSetup = {}
+                    if session_configuration_request is not None:
+                        try:
+                            session_setup = json.loads(
+                                session_configuration_request
+                            ).get("setup", {})
+                        except (json.JSONDecodeError, TypeError):
+                            session_setup = {}
+                    tool_call_generation_config = (
+                        session_setup.get("generationConfig", {}) or {}
+                    )
+                    tool_call_modalities = [
+                        modality.lower()
+                        for modality in cast(
+                            List[str],
+                            tool_call_generation_config.get(
+                                "responseModalities", ["AUDIO"]
+                            ),
+                        )
+                    ]
                     returned_message.append(
                         {
                             "type": "response.created",
@@ -1262,6 +1286,13 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
                                 "status": "in_progress",
                                 "output": [],
                                 "conversation_id": current_conversation_id,
+                                "modalities": tool_call_modalities,
+                                "temperature": tool_call_generation_config.get(
+                                    "temperature"
+                                ),
+                                "max_output_tokens": tool_call_generation_config.get(
+                                    "maxOutputTokens"
+                                ),
                             },
                         }
                     )
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 aac14588c95..3d4b95d7740 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
@@ -715,6 +715,12 @@ def test_gemini_tool_call_emits_response_created_preamble():
     assert responses[0]["type"] == "response.created"
     assert "response" in responses[0]
     assert responses[0]["response"]["status"] == "in_progress"
+    # response.created on the tool-call path mirrors the audio/text preamble:
+    # modalities/temperature/max_output_tokens are present so spec-compliant
+    # clients see consistent response metadata regardless of payload type.
+    assert "modalities" in responses[0]["response"]
+    assert "temperature" in responses[0]["response"]
+    assert "max_output_tokens" in responses[0]["response"]
     assert responses[1]["type"] == "response.output_item.added"
     assert responses[1]["item"]["type"] == "function_call"
     assert responses[1]["item"]["status"] == "in_progress"