fix: preserve video content parts in chat history

atomic-agents/atomic_agents/agents/atomic_agent.py

@@ -429,9 +429,9 @@ def _serialize_history_for_token_count(self) -> List[Dict[str, Any]]:
         """
         Serialize conversation history for token counting, handling multimodal content.
 
-        This method converts instructor multimodal objects (Image, Audio, PDF) to the
-        OpenAI format that LiteLLM's token counter expects. Text content is also
-        converted to the proper multimodal text format when mixed with media.
+        This method converts Instructor multimodal objects (Image, Audio, PDF) to the
+        format that LiteLLM's token counter expects. Native content-part dictionaries
+        are preserved, and text content is wrapped when mixed with media.
 
         Returns:
             List[Dict[str, Any]]: History messages in LiteLLM-compatible format.
@@ -449,6 +449,8 @@ def _serialize_history_for_token_count(self) -> List[Dict[str, Any]]:
                     if isinstance(item, str):
                         # Text content - wrap in OpenAI text format
                         serialized_content.append({"type": "text", "text": item})
+                    elif isinstance(item, dict):
+                        serialized_content.append(item)
                     elif isinstance(item, (Image, Audio, PDF)):
                         # Multimodal object - use instructor's to_openai method
                         try:

atomic-agents/atomic_agents/context/chat_history.py

@@ -14,6 +14,14 @@
 INSTRUCTOR_MULTIMODAL_TYPES = (Image, Audio, PDF)
 
 
+def _is_video_content_part(obj) -> bool:
+    if not isinstance(obj, dict) or obj.get("type") != "video_url":
+        return False
+
+    video_url = obj.get("video_url")
+    return isinstance(video_url, dict) and isinstance(video_url.get("url"), str)
+
+
 class Message(BaseModel):
     """
     Represents a message in the chat history.
@@ -126,8 +134,9 @@ def _extract_multimodal_info(obj):
         Recursively extract multimodal objects and build a Pydantic-compatible exclude spec.
 
         Walks the object tree to find all Instructor multimodal types (Image, Audio, PDF)
-        at any nesting depth, collecting them into a flat list and building an exclude
-        specification that can be passed to model_dump_json(exclude=...).
+        and native video content parts at any nesting depth, collecting them into a flat
+        list and building an exclude specification that can be passed to
+        model_dump_json(exclude=...).
 
         Args:
             obj: The object to inspect (BaseIOSchema, list, dict, or primitive).
@@ -137,7 +146,7 @@ def _extract_multimodal_info(obj):
                 - multimodal_objects: flat list of all multimodal objects found
                 - exclude_spec: Pydantic exclude dict, True (exclude entirely), or None
         """
-        if isinstance(obj, INSTRUCTOR_MULTIMODAL_TYPES):
+        if isinstance(obj, INSTRUCTOR_MULTIMODAL_TYPES) or _is_video_content_part(obj):
             return [obj], True
 
         if hasattr(obj, "__class__") and hasattr(obj.__class__, "model_fields"):

atomic-agents/tests/agents/test_atomic_agent.py

@@ -1120,6 +1120,29 @@ class MultimodalInputSchema(BaseIOSchema):
     assert image_entry["image_url"]["url"] == "https://example.com/test.png"
 
 
+def test_serialize_history_for_token_count_preserves_content_part(agent, mock_history):
+    video_part = {
+        "type": "video_url",
+        "video_url": {
+            "url": "mm_file://video-file",
+            "detail": "default",
+        },
+    }
+    mock_history.get_history.return_value = [
+        {
+            "role": "user",
+            "content": ['{"prompt":"Summarize the video"}', video_part],
+        }
+    ]
+
+    serialized = agent._serialize_history_for_token_count()
+
+    assert serialized[0]["content"] == [
+        {"type": "text", "text": '{"prompt":"Summarize the video"}'},
+        video_part,
+    ]
+
+
 # --- Tests for tool_result_role and Gemini system message remapping (issue #221) ---
 
 

atomic-agents/tests/context/test_chat_history.py

@@ -646,6 +646,33 @@ class MessageInput(BaseIOSchema):
     assert img2 in result[0]["content"]
 
 
+def test_get_history_video_content_part(history):
+    """Native video content parts are separated from structured text."""
+
+    class VideoInput(BaseIOSchema):
+        """Input containing a video content part."""
+
+        prompt: str = Field(..., description="Instruction for the video")
+        video: Dict[str, object] = Field(..., description="Video content part")
+
+    video_part = {
+        "type": "video_url",
+        "video_url": {
+            "url": "mm_file://video-file",
+            "detail": "default",
+            "fps": 1,
+        },
+    }
+    history.add_message("user", VideoInput(prompt="Summarize the video", video=video_part))
+
+    result = history.get_history()
+
+    assert len(result) == 1
+    assert isinstance(result[0]["content"], list)
+    assert json.loads(result[0]["content"][0]) == {"prompt": "Summarize the video"}
+    assert result[0]["content"][1] == video_part
+
+
 def test_get_history_list_of_nested_schemas_with_multimodal(history):
     """Multiple nested schemas each containing multimodal objects"""