(fix) litellm_logging.py: fix mcp tool call response logging + (fix) responses_bridge: remove unmapped param error mid-stream - allows gpt-5 web search to work via responses api in .completion() (#16946)

* fix: fix getting mcp servers

* fix(litellm_logging.py): handle list objects for final response in standard logging payload

Fixes issue where mcp tool call response wouldn't show up

* fix(litellm_responses_transformation/): remove invalid item error for unmapped objects - breaks stream and there's no real value to this as outside of a few of them, not all can be mapped to chat completions

resolves error for web search calls via chat completions to responses api

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -32,6 +32,12 @@ from litellm.types.llms.openai import (
     ResponsesAPIOptionalRequestParams,
     ResponsesAPIStreamEvents,
 )
+from litellm.types.utils import (
+    Delta,
+    GenericStreamingChunk,
+    ModelResponseStream,
+    StreamingChoices,
+)
 
 if TYPE_CHECKING:
     from openai.types.responses import ResponseInputImageParam
@@ -46,7 +52,6 @@ if TYPE_CHECKING:
         ChatCompletionThinkingBlock,
         OpenAIMessageContentListBlock,
     )
-    from litellm.types.utils import GenericStreamingChunk, ModelResponseStream
 
 
 class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
@@ -606,11 +611,13 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
             ResponsesAPIOptionalRequestParams.__annotations__.keys()
         )
         # Also include params we handle specially
-        supported_responses_api_params.update({
-            "previous_response_id",
-            "reasoning_effort",  # We map this to "reasoning"
-        })
-        
+        supported_responses_api_params.update(
+            {
+                "previous_response_id",
+                "reasoning_effort",  # We map this to "reasoning"
+            }
+        )
+
         # Extract supported params from extra_body and merge into optional_params
         extra_body_copy = extra_body.copy()
         for key, value in extra_body_copy.items():
@@ -620,14 +627,16 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
 
         return optional_params
 
-    def _map_reasoning_effort(self, reasoning_effort: Union[str, Dict[str, Any]]) -> Optional[Reasoning]:
+    def _map_reasoning_effort(
+        self, reasoning_effort: Union[str, Dict[str, Any]]
+    ) -> Optional[Reasoning]:
         # If dict is passed, convert it directly to Reasoning object
         if isinstance(reasoning_effort, dict):
             return Reasoning(**reasoning_effort)  # type: ignore[typeddict-item]
 
         # If string is passed, map without summary (default)
         if reasoning_effort == "none":
-            return Reasoning(effort="none") # type: ignore
+            return Reasoning(effort="none")  # type: ignore
         elif reasoning_effort == "high":
             return Reasoning(effort="high")
         elif reasoning_effort == "medium":
@@ -746,12 +755,6 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
                     finish_reason="",
                     usage=None,
                 )
-            elif output_item.get("type") == "message":
-                pass
-            elif output_item.get("type") == "reasoning":
-                pass
-            else:
-                raise ValueError(f"Chat provider: Invalid output_item  {output_item}")
         elif event_type == "response.function_call_arguments.delta":
             content_part: Optional[str] = parsed_chunk.get("delta", None)
             if content_part:
@@ -813,10 +816,6 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
                 return GenericStreamingChunk(
                     finish_reason="stop", is_finished=True, usage=None, text=""
                 )
-            elif output_item.get("type") == "reasoning":
-                pass
-            else:
-                raise ValueError(f"Chat provider: Invalid output_item  {output_item}")
 
         elif event_type == "response.output_text.delta":
             # Content part added to output

litellm/litellm_core_utils/litellm_logging.py

@@ -4358,12 +4358,12 @@ class StandardLoggingPayloadSetup:
         """
         Get final response object after redacting the message input/output from logging
         """
-        if response_obj is not None:
+        if response_obj:
             final_response_obj: Optional[Union[dict, str, list]] = response_obj
         elif isinstance(init_response_obj, list) or isinstance(init_response_obj, str):
             final_response_obj = init_response_obj
         else:
-            final_response_obj = None
+            final_response_obj = {}
 
         modified_final_response_obj = redact_message_input_output_from_logging(
             model_call_details=kwargs,

litellm/proxy/_experimental/mcp_server/server.py

@@ -1286,6 +1286,7 @@ if MCP_AVAILABLE:
         # Allow modifying the MCP tool call response before it is returned to the user
         #########################################################
         if litellm_logging_obj:
+            litellm_logging_obj.post_call(original_response=response)
             end_time = datetime.now()
             await litellm_logging_obj.async_post_mcp_tool_call_hook(
                 kwargs=litellm_logging_obj.model_call_details,

litellm/proxy/db/db_spend_update_writer.py

@@ -8,9 +8,9 @@ Module responsible for
 import asyncio
 import json
 import os
+import random
 import time
 import traceback
-import random
 from datetime import datetime, timedelta
 from typing import TYPE_CHECKING, Any, Dict, Literal, Optional, Union, cast, overload
 
@@ -342,7 +342,7 @@ class DBSpendUpdateWriter:
     ):
         """
         Update spend for all tags in the request.
-        
+
         Args:
             response_cost: Cost of the request
             request_tags: JSON string of tags list e.g. '["prod-tag", "test-tag"]'
@@ -863,7 +863,7 @@ class DBSpendUpdateWriter:
     ):
         """
         Helper function to update spend for any entity type (team, org, tag, etc).
-        
+
         Args:
             entity_name: Name of entity for logging (e.g., "Team", "Org", "Tag")
             transactions: Dictionary of {entity_id: response_cost}
@@ -875,9 +875,7 @@ class DBSpendUpdateWriter:
         """
         from litellm.proxy.utils import _raise_failed_update_spend_exception
 
-        verbose_proxy_logger.debug(
-            f"{entity_name} Spend transactions: {transactions}"
-        )
+        verbose_proxy_logger.debug(f"{entity_name} Spend transactions: {transactions}")
         if transactions is not None and len(transactions.keys()) > 0:
             for i in range(n_retry_times + 1):
                 start_time = time.time()
@@ -1062,17 +1060,19 @@ class DBSpendUpdateWriter:
 
                             # Add cache-related fields if they exist
                             if "cache_read_input_tokens" in transaction:
-                                common_data[
-                                    "cache_read_input_tokens"
-                                ] = transaction.get("cache_read_input_tokens", 0)
+                                common_data["cache_read_input_tokens"] = (
+                                    transaction.get("cache_read_input_tokens", 0)
+                                )
                             if "cache_creation_input_tokens" in transaction:
-                                common_data[
-                                    "cache_creation_input_tokens"
-                                ] = transaction.get("cache_creation_input_tokens", 0)
+                                common_data["cache_creation_input_tokens"] = (
+                                    transaction.get("cache_creation_input_tokens", 0)
+                                )
 
                             if entity_type == "tag" and "request_id" in transaction:
-                                common_data["request_id"] = transaction.get("request_id")
-                            
+                                common_data["request_id"] = transaction.get(
+                                    "request_id"
+                                )
+
                             # Create update data structure
                             update_data = {
                                 "prompt_tokens": {

litellm/proxy/spend_tracking/spend_tracking_utils.py

@@ -95,9 +95,9 @@ def _get_spend_logs_metadata(
     clean_metadata["applied_guardrails"] = applied_guardrails
     clean_metadata["batch_models"] = batch_models
     clean_metadata["mcp_tool_call_metadata"] = mcp_tool_call_metadata
-    clean_metadata[
-        "vector_store_request_metadata"
-    ] = _get_vector_store_request_for_spend_logs_payload(vector_store_request_metadata)
+    clean_metadata["vector_store_request_metadata"] = (
+        _get_vector_store_request_for_spend_logs_payload(vector_store_request_metadata)
+    )
     clean_metadata["guardrail_information"] = guardrail_information
     clean_metadata["usage_object"] = usage_object
     clean_metadata["model_map_information"] = model_map_information
@@ -142,28 +142,26 @@ def get_spend_logs_id(
     return id
 
 
-def _extract_usage_for_ocr_call(
-    response_obj: Any, response_obj_dict: dict
-) -> dict:
+def _extract_usage_for_ocr_call(response_obj: Any, response_obj_dict: dict) -> dict:
     """
     Extract usage information for OCR/AOCR calls.
-    
+
     OCR responses use usage_info (with pages_processed) instead of token-based usage.
-    
+
     Args:
         response_obj: The raw response object (can be dict, BaseModel, or other)
         response_obj_dict: Dictionary representation of the response object
-    
+
     Returns:
         A dict with prompt_tokens=0, completion_tokens=0, total_tokens=0,
         and pages_processed from usage_info.
     """
     usage_info = None
-    
+
     # Try to extract usage_info from dict
     if isinstance(response_obj_dict, dict) and "usage_info" in response_obj_dict:
         usage_info = response_obj_dict.get("usage_info")
-    
+
     # Try to extract usage_info from object attributes if not found in dict
     if not usage_info and hasattr(response_obj, "usage_info"):
         usage_info = response_obj.usage_info
@@ -171,7 +169,7 @@ def _extract_usage_for_ocr_call(
             usage_info = usage_info.model_dump()
         elif hasattr(usage_info, "__dict__"):
             usage_info = vars(usage_info)
-    
+
     # For OCR, we track pages instead of tokens
     if usage_info is not None:
         # Handle dict or object with attributes
@@ -193,7 +191,7 @@ def _extract_usage_for_ocr_call(
                 "prompt_tokens": 0,
                 "completion_tokens": 0,
                 "total_tokens": 0,
-                "pages_processed": 0
+                "pages_processed": 0,
             }
     else:
         return {}
@@ -206,17 +204,18 @@ def get_logging_payload(  # noqa: PLR0915
 
     if kwargs is None:
         kwargs = {}
-    if response_obj is None or (
-        not isinstance(response_obj, BaseModel) and not isinstance(response_obj, dict)
-    ):
+
+    if response_obj is None:
         response_obj = {}
+    elif not isinstance(response_obj, BaseModel) and not isinstance(response_obj, dict):
+        response_obj = {"result": str(response_obj)}
     # standardize this function to be used across, s3, dynamoDB, langfuse logging
     litellm_params = kwargs.get("litellm_params", {})
     metadata = get_litellm_metadata_from_kwargs(kwargs)
     completion_start_time = kwargs.get("completion_start_time", end_time)
     call_type = kwargs.get("call_type")
     cache_hit = kwargs.get("cache_hit", False)
-    
+
     # Convert response_obj to dict first
     if isinstance(response_obj, dict):
         response_obj_dict = response_obj
@@ -224,7 +223,7 @@ def get_logging_payload(  # noqa: PLR0915
         response_obj_dict = response_obj.model_dump()
     else:
         response_obj_dict = {}
-    
+
     # Handle OCR responses which use usage_info instead of usage
     if call_type in ["ocr", "aocr"]:
         usage = _extract_usage_for_ocr_call(response_obj, response_obj_dict)
@@ -658,6 +657,7 @@ def _get_response_for_spend_logs_payload(
         sanitized_wrapper = _sanitize_request_body_for_spend_logs_payload(
             {"response": response_obj}
         )
+
         sanitized_response = sanitized_wrapper.get("response", response_obj)
 
         if sanitized_response is None:

tests/llm_translation/test_openai.py

@@ -348,7 +348,11 @@ def test_openai_image_generation_forwards_organization(mock_get_openai_client):
                     return {
                         "created": 123,
                         "data": [{"url": "http://example.com/image.png"}],
-                        "usage": {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0},
+                        "usage": {
+                            "input_tokens": 0,
+                            "output_tokens": 0,
+                            "total_tokens": 0,
+                        },
                     }
 
             return _Resp()
@@ -797,7 +801,9 @@ async def test_openai_service_tier_parameter():
         # Verify the request contains the service_tier parameter
         assert "service_tier" in request_body, "service_tier should be in request body"
         # Verify service_tier is correctly sent to the API
-        assert request_body["service_tier"] == "priority", "service_tier should be 'priority'"
+        assert (
+            request_body["service_tier"] == "priority"
+        ), "service_tier should be 'priority'"
 
 
 def test_openai_service_tier_parameter_sync():
@@ -826,7 +832,9 @@ def test_openai_service_tier_parameter_sync():
         # Verify the request contains the service_tier parameter
         assert "service_tier" in request_body, "service_tier should be in request body"
         # Verify service_tier is correctly sent to the API
-        assert request_body["service_tier"] == "priority", "service_tier should be 'priority'"
+        assert (
+            request_body["service_tier"] == "priority"
+        ), "service_tier should be 'priority'"
 
 
 def test_gpt_5_reasoning_streaming():
@@ -1358,7 +1366,7 @@ async def test_streaming_tool_calls_with_n_greater_than_1(model):
     """
     Test that the index field in a choice object is correctly populated
     when using streaming mode with n>1 and tool calls.
-    
+
     Regression test for: https://github.com/BerriAI/litellm/issues/8977
     """
     tools = [
@@ -1386,7 +1394,7 @@ async def test_streaming_tool_calls_with_n_greater_than_1(model):
             },
         }
     ]
-    
+
     response = litellm.completion(
         model=model,
         messages=[
@@ -1399,20 +1407,30 @@ async def test_streaming_tool_calls_with_n_greater_than_1(model):
         stream=True,
         n=3,
     )
-    
+
     # Collect all chunks and their indices
     indices_seen = []
     for chunk in response:
-        assert len(chunk.choices) == 1, "Each streaming chunk should have exactly 1 choice"
-        assert hasattr(chunk.choices[0], "index"), "Choice should have an index attribute"
+        assert (
+            len(chunk.choices) == 1
+        ), "Each streaming chunk should have exactly 1 choice"
+        assert hasattr(
+            chunk.choices[0], "index"
+        ), "Choice should have an index attribute"
         index = chunk.choices[0].index
         indices_seen.append(index)
-    
+
     # Verify that we got chunks with different indices (0, 1, 2 for n=3)
     unique_indices = set(indices_seen)
-    assert unique_indices == {0, 1, 2}, f"Should have indices 0, 1, 2 for n=3, got {unique_indices}"
-    
-    print(f"✓ Test passed: streaming with n=3 and tool calls correctly populates index field")
+    assert unique_indices == {
+        0,
+        1,
+        2,
+    }, f"Should have indices 0, 1, 2 for n=3, got {unique_indices}"
+
+    print(
+        f"✓ Test passed: streaming with n=3 and tool calls correctly populates index field"
+    )
     print(f"  Indices seen: {indices_seen}")
     print(f"  Unique indices: {unique_indices}")
 
@@ -1436,19 +1454,42 @@ async def test_streaming_content_with_n_greater_than_1(model):
         n=2,
         max_tokens=10,
     )
-    
+
     # Collect all chunks and their indices
     indices_seen = []
     for chunk in response:
-        assert len(chunk.choices) == 1, "Each streaming chunk should have exactly 1 choice"
-        assert hasattr(chunk.choices[0], "index"), "Choice should have an index attribute"
+        assert (
+            len(chunk.choices) == 1
+        ), "Each streaming chunk should have exactly 1 choice"
+        assert hasattr(
+            chunk.choices[0], "index"
+        ), "Choice should have an index attribute"
         index = chunk.choices[0].index
         indices_seen.append(index)
-    
+
     # Verify that we got chunks with different indices (0, 1 for n=2)
     unique_indices = set(indices_seen)
-    assert unique_indices == {0, 1}, f"Should have indices 0, 1 for n=2, got {unique_indices}"
-    
-    print(f"✓ Test passed: streaming with n=2 and regular content correctly populates index field")
+    assert unique_indices == {
+        0,
+        1,
+    }, f"Should have indices 0, 1 for n=2, got {unique_indices}"
+
+    print(
+        f"✓ Test passed: streaming with n=2 and regular content correctly populates index field"
+    )
     print(f"  Indices seen: {indices_seen}")
     print(f"  Unique indices: {unique_indices}")
+
+
+def test_gpt_5_web_search():
+    response = litellm.completion(
+        model="openai/responses/gpt-5",
+        messages=[{"role": "user", "content": "get price of nvda"}],
+        stream=True,
+        temperature=1,
+        max_tokens=8192,
+        tools=[{"type": "web_search"}],
+    )
+
+    for chunk in response:
+        print("chunk: ", chunk)

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -568,10 +568,11 @@ async def test_e2e_generate_cold_storage_object_key_successful():
     start_time = datetime(2025, 1, 15, 10, 30, 45, 123456, timezone.utc)
     response_id = "chatcmpl-test-12345"
     team_alias = "test-team"
-    
-    with patch("litellm.cold_storage_custom_logger", return_value="s3"), \
-         patch("litellm.integrations.s3.get_s3_object_key") as mock_get_s3_key:
-        
+
+    with patch("litellm.cold_storage_custom_logger", return_value="s3"), patch(
+        "litellm.integrations.s3.get_s3_object_key"
+    ) as mock_get_s3_key:
+
         # Mock the S3 object key generation to return a predictable result
         mock_get_s3_key.return_value = (
             "2025-01-15/time-10-30-45-123456_chatcmpl-test-12345.json"
@@ -613,11 +614,13 @@ async def test_e2e_generate_cold_storage_object_key_with_custom_logger_s3_path()
     # Create mock custom logger with s3_path
     mock_custom_logger = MagicMock()
     mock_custom_logger.s3_path = "storage"
-    
-    with patch("litellm.cold_storage_custom_logger", "s3_v2"), \
-         patch("litellm.logging_callback_manager.get_active_custom_logger_for_callback_name") as mock_get_logger, \
-         patch("litellm.integrations.s3.get_s3_object_key") as mock_get_s3_key:
-        
+
+    with patch("litellm.cold_storage_custom_logger", "s3_v2"), patch(
+        "litellm.logging_callback_manager.get_active_custom_logger_for_callback_name"
+    ) as mock_get_logger, patch(
+        "litellm.integrations.s3.get_s3_object_key"
+    ) as mock_get_s3_key:
+
         # Setup mocks
         mock_get_logger.return_value = mock_custom_logger
         mock_get_s3_key.return_value = (
@@ -663,11 +666,13 @@ async def test_e2e_generate_cold_storage_object_key_with_logger_no_s3_path():
     # Create mock custom logger without s3_path
     mock_custom_logger = MagicMock()
     mock_custom_logger.s3_path = None  # or could be missing attribute
-    
-    with patch("litellm.cold_storage_custom_logger", "s3_v2"), \
-         patch("litellm.logging_callback_manager.get_active_custom_logger_for_callback_name") as mock_get_logger, \
-         patch("litellm.integrations.s3.get_s3_object_key") as mock_get_s3_key:
-        
+
+    with patch("litellm.cold_storage_custom_logger", "s3_v2"), patch(
+        "litellm.logging_callback_manager.get_active_custom_logger_for_callback_name"
+    ) as mock_get_logger, patch(
+        "litellm.integrations.s3.get_s3_object_key"
+    ) as mock_get_s3_key:
+
         # Setup mocks
         mock_get_logger.return_value = mock_custom_logger
         mock_get_s3_key.return_value = (
@@ -708,7 +713,7 @@ async def test_e2e_generate_cold_storage_object_key_not_configured():
     team_alias = "another-team"
 
     # Use patch to ensure test isolation
-    with patch.object(litellm, 'cold_storage_custom_logger', None):
+    with patch.object(litellm, "cold_storage_custom_logger", None):
         # Call the function
         result = StandardLoggingPayloadSetup._generate_cold_storage_object_key(
             start_time=start_time, response_id=response_id, team_alias=team_alias
@@ -716,3 +721,41 @@ async def test_e2e_generate_cold_storage_object_key_not_configured():
 
     # Verify the result is None when cold storage is not configured
     assert result is None
+
+
+def test_get_final_response_obj_with_empty_response_obj_and_list_init():
+    """
+    Test get_final_response_obj when response_obj is empty dict and init_response_obj is a list.
+
+    When response_obj is empty (falsy), the method should return init_response_obj if it's a list.
+    """
+    from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup
+
+    # Create test objects
+    class TestObject1:
+        def __init__(self):
+            self.name = "Object1"
+
+    class TestObject2:
+        def __init__(self):
+            self.name = "Object2"
+
+    obj1 = TestObject1()
+    obj2 = TestObject2()
+
+    # Test case: empty response_obj, list init_response_obj
+    response_obj = {}
+    init_response_obj = [obj1, obj2]
+    kwargs = {}
+
+    # Call the method
+    result = StandardLoggingPayloadSetup.get_final_response_obj(
+        response_obj=response_obj, init_response_obj=init_response_obj, kwargs=kwargs
+    )
+
+    # Verify the result
+    assert result == [obj1, obj2]
+    assert result is init_response_obj  # Should be the exact same list object
+    assert len(result) == 2
+    assert result[0].name == "Object1"
+    assert result[1].name == "Object2"