fix(responses): tolerate dict response payloads in streaming events

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -1311,8 +1311,9 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
 
             # Check if response contains function_call items in output
             # to determine correct finish_reason
-            response_data = parsed_chunk.get("response", {})
-            output_items = response_data.get("output", []) if response_data else []
+            response_payload = parsed_chunk.get("response")
+            response_data: dict[str, Any] = response_payload if isinstance(response_payload, dict) else {}
+            output_items = response_data.get("output") or []
 
             has_function_calls = any(
                 item.get("type") == "function_call" for item in output_items if isinstance(item, dict)

litellm/litellm_core_utils/litellm_logging.py

@@ -78,7 +78,10 @@ from litellm.litellm_core_utils.redact_messages import (
 )
 from litellm.llms.base_llm.ocr.transformation import OCRResponse
 from litellm.llms.base_llm.search.transformation import SearchResponse
-from litellm.responses.utils import ResponseAPILoggingUtils
+from litellm.responses.utils import (
+    ResponseAPILoggingUtils,
+    normalize_response_api_usage,
+)
 from litellm.types.agents import LiteLLMSendMessageResponse
 from litellm.types.containers.main import ContainerObject
 from litellm.types.llms.openai import (
@@ -3224,13 +3227,17 @@ class Logging(LiteLLMLoggingBaseClass):
             (ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent),
         ):
             ## return unified Usage object
-            if isinstance(result.response.usage, ResponseAPIUsage):
+            response = result.response
+            if isinstance(response, dict):
+                response = ResponsesAPIResponse.model_construct(**response)
+            responses_api_usage = normalize_response_api_usage(response.usage)
+            if responses_api_usage is not None:
                 transformed_usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(
-                    result.response.usage
+                    responses_api_usage
                 )
                 # Set as dict instead of Usage object so model_dump() serializes it correctly
                 setattr(
-                    result.response,
+                    response,
                     "usage",
                     (
                         transformed_usage.model_dump()
@@ -3238,7 +3245,7 @@ class Logging(LiteLLMLoggingBaseClass):
                         else dict(transformed_usage)
                     ),
                 )
-            return result.response
+            return response
         else:
             return None
 

litellm/llms/openai/responses/transformation.py

@@ -30,6 +30,23 @@ else:
     LiteLLMLoggingObj = Any
 
 
+def _with_typed_response_payload(parsed_chunk: dict[str, Any]) -> dict[str, Any]:
+    """Build the nested ``ResponsesAPIResponse`` so ``model_construct`` doesn't leave it a raw dict."""
+    from litellm.responses.utils import normalize_response_api_usage
+
+    response_payload = parsed_chunk.get("response")
+    if not isinstance(response_payload, dict):
+        return parsed_chunk
+    try:
+        response = ResponsesAPIResponse(**response_payload)
+    except ValidationError:
+        response = ResponsesAPIResponse.model_construct(**response_payload)
+        normalized_usage = normalize_response_api_usage(response.usage)
+        if normalized_usage is not None:
+            response.usage = normalized_usage
+    return {**parsed_chunk, "response": response}
+
+
 class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
     @property
     def custom_llm_provider(self) -> LlmProviders:
@@ -349,7 +366,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
                 event_pydantic_model.__name__,
                 parsed_chunk,
             )
-            return event_pydantic_model.model_construct(**parsed_chunk)
+            return event_pydantic_model.model_construct(**_with_typed_response_payload(parsed_chunk))
 
     @staticmethod
     def get_event_model_class(event_type: str) -> Any:

litellm/responses/utils.py

@@ -6,6 +6,7 @@ from typing import (
     Iterable,
     List,
     Mapping,
+    MutableMapping,
     Optional,
     Type,
     Union,
@@ -14,7 +15,7 @@ from typing import (
     overload,
 )
 
-from pydantic import BaseModel
+from pydantic import BaseModel, ValidationError
 
 import litellm
 from litellm._logging import verbose_logger
@@ -1010,6 +1011,42 @@ class ResponsesAPIRequestUtils:
         )
 
 
+def normalize_response_api_usage(usage: object) -> ResponseAPIUsage | None:
+    """Coerce a usage payload to ``ResponseAPIUsage``; returns None when it isn't Responses-API usage.
+
+    Streaming events built via ``model_construct`` (provider payloads that fail validation) carry
+    their usage as a plain dict, so attribute access on it would raise.
+    """
+    if isinstance(usage, ResponseAPIUsage):
+        return usage
+    if isinstance(usage, Mapping):
+        try:
+            return ResponseAPIUsage(**dict(usage))
+        except ValidationError:
+            return None
+    return None
+
+
+def get_response_api_usage(
+    response: ResponsesAPIResponse | Mapping[str, Any],
+) -> ResponseAPIUsage | None:
+    """Read usage off a Responses API response payload that may still be an unvalidated dict."""
+    if isinstance(response, Mapping):
+        return normalize_response_api_usage(response.get("usage"))
+    return normalize_response_api_usage(response.usage)
+
+
+def set_response_api_usage(
+    response: ResponsesAPIResponse | MutableMapping[str, Any],
+    usage: ResponseAPIUsage,
+) -> None:
+    """Write usage back onto a Responses API response payload that may still be an unvalidated dict."""
+    if isinstance(response, MutableMapping):
+        response["usage"] = usage
+    else:
+        response.usage = usage
+
+
 class ResponseAPILoggingUtils:
     @staticmethod
     def _is_response_api_usage(usage: Union[dict, ResponseAPIUsage]) -> bool:

litellm/router.py

@@ -2242,7 +2242,9 @@ class Router:
             completed,
             (ResponseCompletedEvent, ResponseFailedEvent, ResponseIncompleteEvent),
         ):
-            return completed.response.usage
+            from litellm.responses.utils import get_response_api_usage
+
+            return get_response_api_usage(completed.response)
         return None
 
     @staticmethod
@@ -2264,6 +2266,10 @@ class Router:
         and produce a clean ResponseAPIUsage — no token-naming split, no
         setattr bypass.
         """
+        from litellm.responses.utils import (
+            get_response_api_usage,
+            set_response_api_usage,
+        )
         from litellm.types.llms.openai import (
             ResponseAPIUsage,
             ResponseCompletedEvent,
@@ -2276,15 +2282,17 @@ class Router:
             (ResponseCompletedEvent, ResponseFailedEvent, ResponseIncompleteEvent),
         ):
             return
-        response = fallback_item.response
-        if response.usage is None:
+        fb = get_response_api_usage(fallback_item.response)
+        if fb is None:
             return
 
-        fb = response.usage
-        response.usage = ResponseAPIUsage(
-            input_tokens=(partial_usage.input_tokens or 0) + (fb.input_tokens or 0),
-            output_tokens=(partial_usage.output_tokens or 0) + (fb.output_tokens or 0),
-            total_tokens=(partial_usage.total_tokens or 0) + (fb.total_tokens or 0),
+        set_response_api_usage(
+            fallback_item.response,
+            ResponseAPIUsage(
+                input_tokens=(partial_usage.input_tokens or 0) + (fb.input_tokens or 0),
+                output_tokens=(partial_usage.output_tokens or 0) + (fb.output_tokens or 0),
+                total_tokens=(partial_usage.total_tokens or 0) + (fb.total_tokens or 0),
+            ),
         )
 
     @staticmethod

tests/router_unit_tests/test_router_aresponses_streaming_fallback.py

@@ -107,6 +107,48 @@ def test_combine_responses_fallback_usage_sums_completed_event():
     assert combined.total_tokens == 26
 
 
+def _make_completed_event_with_dict_response(
+    input_tokens: int, output_tokens: int, total_tokens: int
+) -> ResponseCompletedEvent:
+    """A provider payload that failed validation keeps `response` as a raw dict (model_construct)."""
+    return ResponseCompletedEvent.model_construct(
+        type=ResponsesAPIStreamEvents.RESPONSE_COMPLETED,
+        response={
+            "id": "resp_123",
+            "usage": {
+                "input_tokens": input_tokens,
+                "output_tokens": output_tokens,
+                "total_tokens": total_tokens,
+            },
+        },
+    )
+
+
+def test_extract_partial_responses_usage_dict_response():
+    """Regression test for https://github.com/BerriAI/litellm/issues/34754: dict response payload."""
+    source = MagicMock()
+    source.completed_response = _make_completed_event_with_dict_response(11, 7, 18)
+
+    usage = Router._extract_partial_responses_usage(source)
+    assert usage is not None
+    assert (usage.input_tokens, usage.output_tokens, usage.total_tokens) == (11, 7, 18)
+
+
+def test_combine_responses_fallback_usage_dict_response():
+    """Usage is summed into a fallback event whose response payload is a raw dict."""
+    fallback_event = _make_completed_event_with_dict_response(5, 3, 8)
+    partial = ResponseAPIUsage(input_tokens=11, output_tokens=7, total_tokens=18)
+
+    Router._combine_responses_fallback_usage(fallback_event, partial)
+
+    combined = fallback_event.response["usage"]
+    assert (combined.input_tokens, combined.output_tokens, combined.total_tokens) == (
+        16,
+        10,
+        26,
+    )
+
+
 def test_combine_responses_fallback_usage_passthrough_for_unknown_event():
     """Events that are not completed/failed/incomplete are not mutated."""
     other = MagicMock()  # not a ResponseCompletedEvent etc. → isinstance false

tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py

@@ -1075,6 +1075,40 @@ def test_response_completed_with_function_calls_emits_tool_calls_finish_reason()
     ), "response.completed with function_call output should emit finish_reason='tool_calls'"
 
 
+def test_response_completed_with_null_output_emits_stop_finish_reason():
+    """
+    Regression test for https://github.com/BerriAI/litellm/issues/34754
+
+    Upstream providers may send `response.output: null` on response.completed; iterating that
+    None used to raise `TypeError`/`AttributeError` while bridging /responses back to
+    /chat/completions.
+    """
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        OpenAiResponsesToChatCompletionStreamIterator,
+    )
+
+    iterator = OpenAiResponsesToChatCompletionStreamIterator(
+        streaming_response=None, sync_stream=True
+    )
+
+    chunk = {
+        "type": "response.completed",
+        "response": {
+            "id": "resp_123",
+            "status": "completed",
+            "output": None,
+            "usage": {"input_tokens": 5, "output_tokens": 3, "total_tokens": 8},
+        },
+    }
+
+    result = iterator.chunk_parser(chunk)
+
+    assert result.choices[0].finish_reason == "stop"
+    assert result.usage is not None
+    assert result.usage.prompt_tokens == 5
+    assert result.usage.completion_tokens == 3
+
+
 def test_response_completed_with_message_only_emits_stop_finish_reason():
     """
     Test that response.completed with only message output (no function_call) emits finish_reason='stop'.

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -2678,6 +2678,44 @@ def test_streaming_success_handler_includes_vertex_ai_metadata_in_standard_loggi
     assert payload["response"]["vertex_ai_url_context_metadata"] == url_context_metadata
 
 
+def test_get_assembled_streaming_response_handles_dict_response_payload():
+    """Regression test for https://github.com/BerriAI/litellm/issues/34754
+
+    Completed events built via model_construct carry a raw dict response; usage must still be
+    assembled into chat-style usage instead of raising AttributeError and dropping the log.
+    """
+    import datetime
+
+    from litellm.types.llms.openai import (
+        ResponseCompletedEvent,
+        ResponsesAPIStreamEvents,
+    )
+
+    logging_obj = _make_logging_obj(stream=True)
+    event = ResponseCompletedEvent.model_construct(
+        type=ResponsesAPIStreamEvents.RESPONSE_COMPLETED,
+        response={
+            "id": "resp-1",
+            "status": "completed",
+            "output": [],
+            "usage": {"input_tokens": 9, "output_tokens": 6, "total_tokens": 15},
+        },
+    )
+
+    assembled = logging_obj._get_assembled_streaming_response(
+        result=event,
+        start_time=datetime.datetime.now(),
+        end_time=datetime.datetime.now(),
+        is_async=True,
+        streaming_chunks=[],
+    )
+
+    assert assembled is not None
+    assert assembled.usage["prompt_tokens"] == 9
+    assert assembled.usage["completion_tokens"] == 6
+    assert assembled.usage["total_tokens"] == 15
+
+
 def test_get_assembled_streaming_response_returns_none_for_non_streaming_text_completion():
     """Non-streaming TextCompletionResponse should also return None."""
     import datetime

tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py

@@ -288,6 +288,36 @@ class TestOpenAIResponsesAPIConfig:
             assert result.type == ResponsesAPIStreamEvents.RESPONSE_COMPLETED
             assert result.response.id == "resp_123"
 
+    def test_transform_streaming_response_invalid_payload_keeps_typed_response(self):
+        """Regression test for https://github.com/BerriAI/litellm/issues/34754
+
+        A response payload that fails validation falls back to model_construct; the nested
+        response must still be a ResponsesAPIResponse so `event.response.usage` works.
+        """
+        completed_chunk = {
+            "type": "response.completed",
+            "response": {
+                "id": "resp_123",
+                "status": "completed",
+                "output": [],
+                "usage": {
+                    "input_tokens": 12,
+                    "output_tokens": 4,
+                    "total_tokens": 16,
+                },
+            },
+        }
+
+        result = self.config.transform_streaming_response(
+            model=self.model,
+            parsed_chunk=completed_chunk,
+            logging_obj=self.logging_obj,
+        )
+
+        assert isinstance(result, ResponseCompletedEvent)
+        assert isinstance(result.response, ResponsesAPIResponse)
+        assert result.response.usage.input_tokens == 12
+
     @pytest.mark.serial
     def test_validate_environment(self):
         """Test that validate_environment correctly sets the Authorization header"""