diff --git a/litellm/completion_extras/litellm_responses_transformation/transformation.py b/litellm/completion_extras/litellm_responses_transformation/transformation.py index a89efc4e82b..585454f944b 100644 --- a/litellm/completion_extras/litellm_responses_transformation/transformation.py +++ b/litellm/completion_extras/litellm_responses_transformation/transformation.py @@ -90,9 +90,14 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge): content_type = content_item.get("type") if content_type == "output_text": response_text = content_item.get("text", "") + # Extract annotations from content if present + annotations = LiteLLMResponsesTransformationHandler._convert_annotations_to_chat_format( + content_item.get("annotations", None) + ) msg = Message( role=item.get("role", "assistant"), content=response_text if response_text else "", + annotations=annotations, ) choice = Choices(message=msg, finish_reason="stop", index=index) return choice, index + 1 @@ -364,10 +369,16 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge): elif isinstance(item, ResponseOutputMessage): for content in item.content: response_text = getattr(content, "text", "") + # Extract annotations from content if present + raw_annotations = getattr(content, "annotations", None) + annotations = LiteLLMResponsesTransformationHandler._convert_annotations_to_chat_format( + raw_annotations + ) msg = Message( role=item.role, content=response_text if response_text else "", reasoning_content=reasoning_content, + annotations=annotations, ) choices.append( @@ -763,6 +774,42 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge): return {"format": {"type": "text"}} return None + + @staticmethod + def _convert_annotations_to_chat_format( + annotations: Optional[List[Any]], + ) -> Optional[List[Dict[str, Any]]]: + """ + Convert annotations from Responses API to Chat Completions format. + + Annotations are already in compatible format between both APIs, + so we just need to convert Pydantic models to dicts. + """ + if not annotations: + return None + + result: List[Dict[str, Any]] = [] + for annotation in annotations: + try: + # Convert Pydantic models to dicts (handles both v1 and v2) + if hasattr(annotation, "model_dump"): + annotation_dict = annotation.model_dump() + elif hasattr(annotation, "dict"): + annotation_dict = annotation.dict() + elif isinstance(annotation, dict): + annotation_dict = annotation + else: + # Skip unsupported annotation types + verbose_logger.debug(f"Skipping unsupported annotation type: {type(annotation)}") + continue + + result.append(annotation_dict) + except Exception as e: + # Skip malformed annotations + verbose_logger.debug(f"Skipping malformed annotation: {annotation}, error: {e}") + continue + + return result if result else None def _map_responses_status_to_finish_reason(self, status: Optional[str]) -> str: """Map responses API status to chat completion finish_reason""" diff --git a/litellm/proxy/hooks/litellm_skills/main.py b/litellm/proxy/hooks/litellm_skills/main.py index 26d4cbe1de7..c2ad1e29447 100644 --- a/litellm/proxy/hooks/litellm_skills/main.py +++ b/litellm/proxy/hooks/litellm_skills/main.py @@ -336,8 +336,8 @@ class SkillsInjectionHook(CustomLogger): ) # Check if code execution is enabled for this request - litellm_metadata = request_data.get("litellm_metadata", {}) - metadata = request_data.get("metadata", {}) + litellm_metadata = request_data.get("litellm_metadata") or {} + metadata = request_data.get("metadata") or {} code_exec_enabled = ( litellm_metadata.get("_litellm_code_execution_enabled") or diff --git a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py index 596398e639f..f8a082ee30c 100644 --- a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py +++ b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py @@ -1095,3 +1095,186 @@ def test_map_reasoning_effort_adds_summary_detailed(): os.environ["LITELLM_REASONING_AUTO_SUMMARY"] = original_env elif "LITELLM_REASONING_AUTO_SUMMARY" in os.environ: del os.environ["LITELLM_REASONING_AUTO_SUMMARY"] + + +def test_transform_response_preserves_annotations(): + """ + Test that annotations from Responses API are preserved when transforming to Chat Completions format. + + This is a regression test for the bug where annotations (like url_citation) were being + dropped during the transformation from ResponsesAPIResponse to ModelResponse. + + The fix ensures annotations are extracted from ResponseOutputText content items and + passed through to the Message object in the Chat Completions response. + """ + from unittest.mock import Mock + + from openai.types.responses import ResponseOutputMessage, ResponseOutputText + + from litellm.completion_extras.litellm_responses_transformation.transformation import ( + LiteLLMResponsesTransformationHandler, + ) + from litellm.types.llms.openai import ( + InputTokensDetails, + OutputTokensDetails, + ResponseAPIUsage, + ResponsesAPIResponse, + ) + from litellm.types.utils import ModelResponse, Usage + + handler = LiteLLMResponsesTransformationHandler() + + # Create annotations similar to what OpenAI Responses API returns + annotations = [ + { + "type": "url_citation", + "start_index": 0, + "end_index": 100, + "title": "Example Article", + "url": "https://example.com/article", + }, + { + "type": "url_citation", + "start_index": 101, + "end_index": 200, + "title": "Another Source", + "url": "https://example.com/source", + }, + ] + + # Create output text with annotations + output_text = ResponseOutputText( + annotations=annotations, + text="Here is some information with citations.", + type="output_text", + logprobs=[], + ) + + # Create output message + output_message = ResponseOutputMessage( + id="msg_test123", + content=[output_text], + role="assistant", + status="completed", + type="message", + ) + + # Create usage information + usage = ResponseAPIUsage( + input_tokens=10, + input_tokens_details=InputTokensDetails( + audio_tokens=None, cached_tokens=0, text_tokens=None + ), + output_tokens=20, + output_tokens_details=OutputTokensDetails( + reasoning_tokens=0, text_tokens=None + ), + total_tokens=30, + cost=None, + ) + + # Create the full ResponsesAPIResponse + raw_response = ResponsesAPIResponse( + id="resp_test123", + created_at=1234567890, + error=None, + incomplete_details=None, + instructions=None, + metadata={}, + model="gpt-5.1", + object="response", + output=[output_message], + parallel_tool_calls=True, + temperature=1.0, + tool_choice="auto", + tools=[], + top_p=1.0, + max_output_tokens=None, + previous_response_id=None, + reasoning=None, + status="completed", + text={"format": {"type": "text"}, "verbosity": "medium"}, + truncation="disabled", + usage=usage, + user=None, + store=True, + background=False, + billing={"payer": "openai"}, + max_tool_calls=None, + prompt_cache_key=None, + safety_identifier=None, + service_tier="default", + top_logprobs=0, + ) + + # Create empty model_response + model_response = ModelResponse( + id="chatcmpl-test123", + created=1234567890, + model=None, + object="chat.completion", + system_fingerprint=None, + choices=[], + usage=Usage(completion_tokens=0, prompt_tokens=0, total_tokens=0), + ) + + # Create mock objects for required parameters + logging_obj = Mock() + messages = [{"role": "user", "content": "Tell me about AI"}] + request_data = {"model": "gpt-5.1"} + optional_params = {} + litellm_params = {"acompletion": False, "api_key": None} + encoding = Mock() + + # Call transform_response + result = handler.transform_response( + model="gpt-5.1", + raw_response=raw_response, + model_response=model_response, + logging_obj=logging_obj, + request_data=request_data, + messages=messages, + optional_params=optional_params, + litellm_params=litellm_params, + encoding=encoding, + api_key=None, + json_mode=None, + ) + + # Assertions + assert result.model == "gpt-5.1" + assert len(result.choices) == 1 + + # Check the choice + choice = result.choices[0] + assert choice.finish_reason == "stop" + assert choice.index == 0 + assert choice.message.role == "assistant" + assert choice.message.content == "Here is some information with citations." + + # Check that annotations are preserved + assert hasattr(choice.message, "annotations"), "Message should have annotations attribute" + assert choice.message.annotations is not None, "Annotations should not be None" + assert len(choice.message.annotations) == 2, f"Expected 2 annotations, got {len(choice.message.annotations)}" + + # Verify annotation content + annotation1 = choice.message.annotations[0] + assert annotation1["type"] == "url_citation" + assert annotation1["title"] == "Example Article" + assert annotation1["url"] == "https://example.com/article" + assert annotation1["start_index"] == 0 + assert annotation1["end_index"] == 100 + + annotation2 = choice.message.annotations[1] + assert annotation2["type"] == "url_citation" + assert annotation2["title"] == "Another Source" + assert annotation2["url"] == "https://example.com/source" + assert annotation2["start_index"] == 101 + assert annotation2["end_index"] == 200 + + # Check usage + assert result.usage.prompt_tokens == 10 + assert result.usage.completion_tokens == 20 + assert result.usage.total_tokens == 30 + + print("✓ Annotations from Responses API are correctly preserved in Chat Completions format")