From 73db7a3f78bc475bdb1c9b2c6ff84b9b772f2a33 Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Wed, 8 Jul 2026 16:07:41 +0000
Subject: [PATCH] fix(logging): normalize dict response on streaming responses
 success logger

---
 litellm/litellm_core_utils/litellm_logging.py | 13 +++---
 .../test_litellm_logging.py                   | 43 +++++++++++++++++++
 2 files changed, 50 insertions(+), 6 deletions(-)

diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py
index 936d79b22d6..474f45cd2d1 100644
--- a/litellm/litellm_core_utils/litellm_logging.py
+++ b/litellm/litellm_core_utils/litellm_logging.py
@@ -3161,13 +3161,14 @@ class Logging(LiteLLMLoggingBaseClass):
             (ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent),
         ):
             ## return unified Usage object
-            if isinstance(result.response.usage, ResponseAPIUsage):
-                transformed_usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(
-                    result.response.usage
-                )
+            response = result.response
+            if isinstance(response, dict):
+                response = ResponsesAPIResponse.model_validate(response)
+            if isinstance(response.usage, ResponseAPIUsage):
+                transformed_usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(response.usage)
                 # Set as dict instead of Usage object so model_dump() serializes it correctly
                 setattr(
-                    result.response,
+                    response,
                     "usage",
                     (
                         transformed_usage.model_dump()
@@ -3175,7 +3176,7 @@ class Logging(LiteLLMLoggingBaseClass):
                         else dict(transformed_usage)
                     ),
                 )
-            return result.response
+            return response
         else:
             return None
 
diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
index 0523ed7ecb1..53cb280c59b 100644
--- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
+++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
@@ -2331,6 +2331,49 @@ def test_get_assembled_streaming_response_returns_result_for_streaming():
     assert assembled is result
 
 
+def test_get_assembled_streaming_response_handles_dict_response_on_completed_event():
+    """
+    Regression for streaming /v1/responses success logging: the ResponseCompletedEvent
+    can carry a plain dict in its `.response` field at runtime (the models allow extra
+    and the event may be constructed without validation). Accessing `result.response.usage`
+    then raised `'dict' object has no attribute 'usage'`, which crashed the success handler
+    before the spend log was written. The assembler must normalize the dict and compute usage.
+    """
+    import datetime
+
+    from litellm.types.llms.openai import ResponseCompletedEvent, ResponsesAPIResponse
+
+    logging_obj = _make_logging_obj(stream=True)
+    response_dict = {
+        "id": "resp-dict-1",
+        "created_at": 1234567890,
+        "output": [],
+        "usage": {
+            "input_tokens": 11,
+            "output_tokens": 22,
+            "total_tokens": 33,
+        },
+    }
+    event = ResponseCompletedEvent.model_construct(
+        type="response.completed", response=response_dict
+    )
+    assert isinstance(event.response, dict)
+
+    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 isinstance(assembled, ResponsesAPIResponse)
+    assert assembled.id == "resp-dict-1"
+    assert assembled.usage["prompt_tokens"] == 11
+    assert assembled.usage["completion_tokens"] == 22
+    assert assembled.usage["total_tokens"] == 33
+
+
 def test_streaming_success_handler_includes_vertex_ai_metadata_in_standard_logging():
     """Assembled streaming responses should include Vertex AI metadata in logging payload."""
     import datetime