fix(responses): log completed response, not the event, for non-streaming callers

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/responses/streaming_iterator.py

@@ -368,18 +368,7 @@ class BaseResponsesAPIStreamingIterator:
         if self._persist_completed_response_before_logging:
             self._persist_completed_response_to_cache(is_async=is_async)
 
-        # Create a copy for logging to avoid modifying the response object that will be returned to the user
-        # The logging handlers may transform usage from Responses API format (input_tokens/output_tokens)
-        # to chat completion format (prompt_tokens/completion_tokens) for internal logging
-        # Use model_dump + model_validate instead of deepcopy to avoid pickle errors with
-        # Pydantic ValidatorIterator when response contains tool_choice with allowed_tools (fixes #17192)
-        logging_response = self.completed_response
-        if self.completed_response is not None and hasattr(self.completed_response, "model_dump"):
-            try:
-                logging_response = type(self.completed_response).model_validate(self.completed_response.model_dump())
-            except Exception:
-                # Fallback to original if serialization fails
-                pass
+        logging_response: Final = self._response_for_success_logging()
 
         end_time: Final = datetime.now()
         if is_async:
@@ -409,6 +398,32 @@ class BaseResponsesAPIStreamingIterator:
             )
         self._run_post_success_hooks(end_time=end_time)
 
+    def _response_for_success_logging(self) -> ResponsesAPIStreamingResponse | ResponsesAPIResponse | None:
+        """A copy of the completed response, so the handlers' usage transforms (Responses API
+        input_tokens/output_tokens to chat completion prompt_tokens/completion_tokens) don't touch
+        what the caller gets back. model_dump + model_validate instead of deepcopy avoids pickle
+        errors with Pydantic ValidatorIterator on tool_choice with allowed_tools (fixes #17192).
+
+        The handlers only unwrap the completion event in their assembled-stream branch, which a
+        non-streaming caller draining this iterator never reaches, so unwrap it here instead."""
+        completed: Final = self.completed_response
+        copied: Final = self._model_copy_for_logging(completed)
+        unwrapped: Final = getattr(copied, "response", None)
+        if getattr(self.logging_obj, "stream", None) is not True and isinstance(unwrapped, ResponsesAPIResponse):
+            return unwrapped
+        return copied
+
+    @staticmethod
+    def _model_copy_for_logging(
+        completed: ResponsesAPIStreamingResponse | None,
+    ) -> ResponsesAPIStreamingResponse | None:
+        if completed is None or not hasattr(completed, "model_dump"):
+            return completed
+        try:
+            return type(completed).model_validate(completed.model_dump())
+        except Exception:
+            return completed
+
     def _handle_logging_completed_response(self):
         """Base implementation - should be overridden by subclasses"""
 

ruff-strict-budget.json

@@ -192,7 +192,7 @@
     "limit": 0
   },
   "S110": {
-    "limit": 218
+    "limit": 217
   },
   "S112": {
     "limit": 22

tests/test_litellm/responses/test_streaming_iterator.py

@@ -235,3 +235,24 @@ def test_sync_transport_error_before_completed_event_raises():
     with pytest.raises(httpx.ReadError):
         for _ in iterator:
             pass
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("stream", [False, True])
+async def test_completed_event_is_unwrapped_only_for_non_streaming_callers(stream: bool):
+    """A non-streaming caller can still drain this iterator (chat -> responses bridge
+    against a provider that always answers with SSE). The success handlers only unwrap
+    the completion event when logging in stream mode, so a wrapped event there yields no
+    standard_logging_object and the SpendLogs row is dropped (#36426)."""
+    logging_obj = _logging_obj_stub()
+    logging_obj.stream = stream
+
+    iterator = _make_iterator(sse_events=_COMPLETE_STREAM_EVENTS, logging_obj=logging_obj)
+    async for _ in iterator:
+        pass
+
+    logged = logging_obj.dispatch_success_handlers.call_args.args[0]
+    if stream:
+        assert logged.type == ResponsesAPIStreamEvents.RESPONSE_COMPLETED
+    else:
+        assert isinstance(logged, ResponsesAPIResponse)

type-discipline-budget.json

@@ -27,7 +27,7 @@
     "limit": 0
   },
   "LIT010": {
-    "limit": 16744
+    "limit": 16742
   },
   "LIT011": {
     "limit": 5596