fix: Only filter out audio content when sending history

audio is transcribed thus no need to be sent, but other blob(e.g. image) should still be sent.

Co-authored-by: Xiang (Sean) Zhou <seanzhougoogle@google.com>
PiperOrigin-RevId: 856422986

Author: seanzhougoogle

src/google/adk/models/gemini_llm_connection.py

@@ -20,6 +20,7 @@
 
 from google.genai import types
 
+from ..utils.content_utils import filter_audio_parts
 from ..utils.context_utils import Aclosing
 from ..utils.variant_utils import GoogleLLMVariant
 from .base_llm_connection import BaseLlmConnection
@@ -63,15 +64,22 @@ async def send_history(self, history: list[types.Content]):
     # TODO: Remove this filter and translate unary contents to streaming
     # contents properly.
 
-    # We ignore any audio from user during the agent transfer phase
+    # Filter out audio parts from history because:
+    # 1. audio has already been transcribed.
+    # 2. sending audio via connection.send or connection.send_live_content is
+    # not supported by LIVE API (session will be corrupted).
+    # This method is called when:
+    # 1. Agent transfer to a new agent
+    # 2. Establishing a new live connection with previous ADK session history
+
     contents = [
-        content
+        filtered
         for content in history
-        if content.parts and content.parts[0].text
+        if (filtered := filter_audio_parts(content)) is not None
     ]
-    logger.debug('Sending history to live connection: %s', contents)
 
     if contents:
+      logger.debug('Sending history to live connection: %s', contents)
       await self._gemini_session.send(
           input=types.LiveClientContent(
               turns=contents,

src/google/adk/utils/content_utils.py

@@ -0,0 +1,38 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from google.genai import types
+
+
+def is_audio_part(part: types.Part) -> bool:
+  return (
+      part.inline_data
+      and part.inline_data.mime_type
+      and part.inline_data.mime_type.startswith('audio/')
+  ) or (
+      part.file_data
+      and part.file_data.mime_type
+      and part.file_data.mime_type.startswith('audio/')
+  )
+
+
+def filter_audio_parts(content: types.Content) -> types.Content | None:
+  if not content.parts:
+    return None
+  filtered_parts = [part for part in content.parts if not is_audio_part(part)]
+  if not filtered_parts:
+    return None
+  return types.Content(role=content.role, parts=filtered_parts)

tests/unittests/models/test_gemini_llm_connection.py

@@ -600,3 +600,177 @@ async def mock_receive_generator():
   assert responses[2].output_transcription.text == 'How can I help?'
   assert responses[2].output_transcription.finished is True
   assert responses[2].partial is False
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    'audio_part',
+    [
+        types.Part(
+            inline_data=types.Blob(data=b'\x00\xFF', mime_type='audio/pcm')
+        ),
+        types.Part(
+            file_data=types.FileData(
+                file_uri='artifact://app/user/session/_adk_live/audio.pcm#1',
+                mime_type='audio/pcm',
+            )
+        ),
+    ],
+)
+async def test_send_history_filters_audio(mock_gemini_session, audio_part):
+  """Test that audio parts (inline or file_data) are filtered out."""
+  connection = GeminiLlmConnection(
+      mock_gemini_session, api_backend=GoogleLLMVariant.VERTEX_AI
+  )
+  history = [
+      types.Content(
+          role='user',
+          parts=[audio_part],
+      ),
+      types.Content(
+          role='model', parts=[types.Part.from_text(text='I heard you')]
+      ),
+  ]
+
+  await connection.send_history(history)
+
+  mock_gemini_session.send.assert_called_once()
+  call_args = mock_gemini_session.send.call_args[1]
+  sent_contents = call_args['input'].turns
+  # Only the model response should be sent (user audio filtered out)
+  assert len(sent_contents) == 1
+  assert sent_contents[0].role == 'model'
+  assert sent_contents[0].parts == [types.Part.from_text(text='I heard you')]
+
+
+@pytest.mark.asyncio
+async def test_send_history_keeps_image_data(mock_gemini_session):
+  """Test that image data is NOT filtered out."""
+  connection = GeminiLlmConnection(
+      mock_gemini_session, api_backend=GoogleLLMVariant.VERTEX_AI
+  )
+  image_blob = types.Blob(data=b'\x89PNG\r\n', mime_type='image/png')
+  history = [
+      types.Content(
+          role='user',
+          parts=[types.Part(inline_data=image_blob)],
+      ),
+      types.Content(
+          role='model', parts=[types.Part.from_text(text='Nice image!')]
+      ),
+  ]
+
+  await connection.send_history(history)
+
+  mock_gemini_session.send.assert_called_once()
+  call_args = mock_gemini_session.send.call_args[1]
+  sent_contents = call_args['input'].turns
+  # Both contents should be sent (image is not filtered)
+  assert len(sent_contents) == 2
+  assert sent_contents[0].parts[0].inline_data == image_blob
+
+
+@pytest.mark.asyncio
+async def test_send_history_mixed_content_filters_only_audio(
+    mock_gemini_session,
+):
+  """Test that mixed content keeps non-audio parts."""
+  connection = GeminiLlmConnection(
+      mock_gemini_session, api_backend=GoogleLLMVariant.VERTEX_AI
+  )
+  history = [
+      types.Content(
+          role='user',
+          parts=[
+              types.Part(
+                  inline_data=types.Blob(
+                      data=b'\x00\xFF', mime_type='audio/wav'
+                  )
+              ),
+              types.Part.from_text(text='transcribed text'),
+          ],
+      ),
+  ]
+
+  await connection.send_history(history)
+
+  mock_gemini_session.send.assert_called_once()
+  call_args = mock_gemini_session.send.call_args[1]
+  sent_contents = call_args['input'].turns
+  # Content should be sent but only with the text part
+  assert len(sent_contents) == 1
+  assert len(sent_contents[0].parts) == 1
+  assert sent_contents[0].parts[0].text == 'transcribed text'
+
+
+@pytest.mark.asyncio
+async def test_send_history_all_audio_content_not_sent(mock_gemini_session):
+  """Test that content with only audio parts is completely removed."""
+  connection = GeminiLlmConnection(
+      mock_gemini_session, api_backend=GoogleLLMVariant.VERTEX_AI
+  )
+  history = [
+      types.Content(
+          role='user',
+          parts=[
+              types.Part(
+                  inline_data=types.Blob(
+                      data=b'\x00\xFF', mime_type='audio/pcm'
+                  )
+              ),
+              types.Part(
+                  file_data=types.FileData(
+                      file_uri='artifact://audio.pcm#1',
+                      mime_type='audio/wav',
+                  )
+              ),
+          ],
+      ),
+  ]
+
+  await connection.send_history(history)
+
+  # No content should be sent since all parts are audio
+  mock_gemini_session.send.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_send_history_empty_history_not_sent(mock_gemini_session):
+  """Test that empty history does not call send."""
+  connection = GeminiLlmConnection(
+      mock_gemini_session, api_backend=GoogleLLMVariant.VERTEX_AI
+  )
+
+  await connection.send_history([])
+
+  mock_gemini_session.send.assert_not_called()
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    'audio_mime_type',
+    ['audio/pcm', 'audio/wav', 'audio/mp3', 'audio/ogg'],
+)
+async def test_send_history_filters_various_audio_mime_types(
+    mock_gemini_session,
+    audio_mime_type,
+):
+  """Test that various audio mime types are all filtered."""
+  connection = GeminiLlmConnection(
+      mock_gemini_session, api_backend=GoogleLLMVariant.VERTEX_AI
+  )
+  history = [
+      types.Content(
+          role='user',
+          parts=[
+              types.Part(
+                  inline_data=types.Blob(data=b'', mime_type=audio_mime_type)
+              )
+          ],
+      ),
+  ]
+
+  await connection.send_history(history)
+
+  # No content should be sent since the only part is audio
+  mock_gemini_session.send.assert_not_called()