fix: Filter out thought parts in lite_llm._get_content

GWeale · copybara-github · commit 1ace8fc6780b · 2026-01-06T15:44:33.000-08:00
Thought parts represent internal model reasoning and should not be included in the content sent back to the model in subsequent turns Close #3948 Co-authored-by: George Weale <gweale@google.com> PiperOrigin-RevId: 852965417
diff --git a/src/google/adk/models/lite_llm.py b/src/google/adk/models/lite_llm.py
@@ -584,9 +584,12 @@ async def _get_content(
     parts: Iterable[types.Part],
     *,
     provider: str = "",
-) -> Union[OpenAIMessageContent, str]:
+) -> OpenAIMessageContent:
   """Converts a list of parts to litellm content.
 
+  Thought parts represent internal model reasoning and are always dropped so
+  they are not replayed back to the model in subsequent turns.
+
   Args:
     parts: The parts to convert.
     provider: The LLM provider name (e.g., "openai", "azure").
@@ -595,11 +598,25 @@ async def _get_content(
     The litellm content.
   """
 
+  parts_without_thought = [part for part in parts if not part.thought]
+  if len(parts_without_thought) == 1:
+    part = parts_without_thought[0]
+    if part.text:
+      return part.text
+    if (
+        part.inline_data
+        and part.inline_data.data
+        and part.inline_data.mime_type
+        and part.inline_data.mime_type.startswith("text/")
+    ):
+      return _decode_inline_text_data(part.inline_data.data)
+
   content_objects = []
-  for part in parts:
+  for part in parts_without_thought:
+    # Skip thought parts to prevent reasoning from being replayed in subsequent
+    # turns. Thought parts are internal model reasoning and should not be sent
+    # back to the model.
     if part.text:
-      if len(parts) == 1:
-        return part.text
       content_objects.append({
           "type": "text",
           "text": part.text,
@@ -611,8 +628,6 @@ async def _get_content(
     ):
       if part.inline_data.mime_type.startswith("text/"):
         decoded_text = _decode_inline_text_data(part.inline_data.data)
-        if len(parts) == 1:
-          return decoded_text
         content_objects.append({
             "type": "text",
             "text": decoded_text,
diff --git a/tests/unittests/models/test_litellm.py b/tests/unittests/models/test_litellm.py
@@ -2086,6 +2086,45 @@ def test_split_message_content_prefers_existing_structured_calls():
   assert tool_calls == [tool_call]
 
 
+@pytest.mark.asyncio
+async def test_get_content_filters_thought_parts():
+  """Test that thought parts are filtered from content.
+
+  Thought parts contain model reasoning that should not be sent back to
+  the model in subsequent turns. This test verifies that _get_content
+  skips parts with thought=True.
+
+  See: https://github.com/google/adk-python/issues/3948
+  """
+  # Create a thought part (reasoning) and a regular text part
+  thought_part = types.Part(text="Internal reasoning...", thought=True)
+  regular_part = types.Part.from_text(text="Visible response")
+  parts = [thought_part, regular_part]
+
+  content = await _get_content(parts)
+
+  # The thought part should be filtered out, leaving only the regular text
+  assert content == "Visible response"
+
+
+@pytest.mark.asyncio
+async def test_get_content_filters_all_thought_parts():
+  """Test that all thought parts are filtered when only thoughts present.
+
+  When all parts are thought parts, _get_content should return an empty list.
+
+  See: https://github.com/google/adk-python/issues/3948
+  """
+  thought_part1 = types.Part(text="First reasoning...", thought=True)
+  thought_part2 = types.Part(text="Second reasoning...", thought=True)
+  parts = [thought_part1, thought_part2]
+
+  content = await _get_content(parts)
+
+  # All thought parts should be filtered out
+  assert content == []
+
+
 @pytest.mark.asyncio
 async def test_get_content_text():
   parts = [types.Part.from_text(text="Test text")]
