Merge pull request #39981 from BerriAI/litellm_lit_7079_bridge_preserve_provider_metadata

fix: keep provider id and metadata on Responses API bridged chat completions

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -5,8 +5,11 @@ Handler for transforming /chat/completions api requests to litellm.responses req
 import json
 import os
 from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping, Sequence
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal, TypedDict, Union, cast, get_args
 
+from openai.types.chat import ChatCompletion
+from openai.types.responses import Response
 from openai.types.responses.custom_tool_param import CustomToolParam
 from openai.types.responses.response_input_param import (
     FunctionCallOutput,
@@ -33,7 +36,7 @@ from litellm.responses.sse_output_recovery import (
     record_output_item_chunk,
     record_output_text_chunk,
 )
-from litellm.responses.utils import normalize_responses_api_stream_options
+from litellm.responses.utils import ResponsesAPIRequestUtils, normalize_responses_api_stream_options
 from litellm.types.llms.openai import (
     REASONING_EFFORT,
     ChatCompletionAnnotation,
@@ -43,6 +46,7 @@ from litellm.types.llms.openai import (
     ChatCompletionToolParamFunctionChunk,
     Reasoning,
     ResponsesAPIOptionalRequestParams,
+    ResponsesAPIResponse,
     ResponsesAPIStreamEvents,
 )
 from litellm.types.utils import GenericStreamingChunk, ModelResponseStream
@@ -54,7 +58,7 @@ if TYPE_CHECKING:
     )
     from pydantic import BaseModel
 
-    from litellm import LiteLLMLoggingObj, ModelResponse
+    from litellm import LiteLLMLoggingObj
     from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
     from litellm.types.llms.openai import (
         ALL_RESPONSES_API_TOOL_PARAMS,
@@ -69,6 +73,28 @@ if TYPE_CHECKING:
     from litellm.types.utils import Choices
 
 
+_CHAT_COMPLETION_FIELDS: Final = frozenset((*ModelResponse.model_fields, "usage"))
+_RESPONSES_API_ONLY_FIELDS: Final = frozenset((*Response.model_fields, *ResponsesAPIResponse.model_fields)) - frozenset(
+    ChatCompletion.model_fields
+)
+
+
+def _provider_metadata(response_fields: Mapping[str, object] | None) -> Mapping[str, object]:
+    return MappingProxyType(
+        {
+            key: value
+            for key, value in (response_fields.items() if response_fields else ())
+            if value is not None and key not in _CHAT_COMPLETION_FIELDS and key not in _RESPONSES_API_ONLY_FIELDS
+        }
+    )
+
+
+def _upstream_response_id(response_id: str | None) -> str | None:
+    if response_id is None:
+        return None
+    return ResponsesAPIRequestUtils.decode_previous_response_id_to_original_previous_response_id(response_id)
+
+
 class _ReasoningSummaryText(TypedDict):
     type: str
     text: str
@@ -904,6 +930,10 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
             ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(raw_response.usage),
         )
 
+        model_response.id = _upstream_response_id(raw_response.id) or raw_response.id
+        for key, value in _provider_metadata(raw_response.model_extra).items():
+            setattr(model_response, key, value)
+
         # Preserve hidden params from the ResponsesAPIResponse, especially the headers
         # which contain important provider information like x-request-id
         raw_response_hidden_params: Final = getattr(raw_response, "_hidden_params", {})
@@ -1359,14 +1389,16 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
         if event_type == "response.created":
             # Initial response creation event
             verbose_logger.debug("Chat provider: response.created -> %s", parsed_chunk)
+            created_response: Final = parsed_chunk.get("response")
             return ModelResponseStream(
+                id=_upstream_response_id(created_response.get("id")) if created_response else None,
                 choices=[
                     StreamingChoices(
                         index=0,
                         delta=Delta(content=""),
                         finish_reason=None,
                     )
-                ]
+                ],
             )
         elif event_type == "response.output_item.added":
             # New output item added
@@ -1534,6 +1566,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
                 from litellm.responses.utils import ResponseAPILoggingUtils
 
                 usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(response_data.get("usage"))
+            provider_metadata: Final = _provider_metadata(response_data)
             return ModelResponseStream(
                 choices=[
                     StreamingChoices(
@@ -1546,6 +1579,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
                     )
                 ],
                 usage=usage,
+                provider_specific_fields=dict(provider_metadata) or None,  # mutable-ok: field is typed dict
             )
         else:
             pass

litellm/litellm_core_utils/litellm_logging.py

@@ -3909,11 +3909,12 @@ class Logging(LiteLLMLoggingBaseClass):
             LiteLLMResponsesTransformationHandler,
         )
 
+        served_id: Final = _provider_response_id(result)
         try:
-            return LiteLLMResponsesTransformationHandler().transform_response(
+            translated: Final = LiteLLMResponsesTransformationHandler().transform_response(
                 model=self.model,
                 raw_response=result,
-                model_response=litellm.ModelResponse(id=_provider_response_id(result)),
+                model_response=litellm.ModelResponse(id=served_id),
                 logging_obj=self,
                 request_data={},
                 messages=[],
@@ -3921,6 +3922,8 @@ class Logging(LiteLLMLoggingBaseClass):
                 litellm_params={},
                 encoding=litellm.encoding,
             )
+            translated.id = served_id or translated.id
+            return translated
         except Exception as e:
             verbose_logger.debug(
                 "Responses API -> ModelResponse translation failed for "
@@ -3928,7 +3931,7 @@ class Logging(LiteLLMLoggingBaseClass):
                 "usage-only ModelResponse to keep the spend_logs row.",
                 str(e),
             )
-            model_response: Final = litellm.ModelResponse(id=_provider_response_id(result))
+            model_response: Final = litellm.ModelResponse(id=served_id)
             model_response.model = self.model
             usage: Final = getattr(result, "usage", None)
             if usage is not None and ResponseAPILoggingUtils._is_response_api_usage(usage):

tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py

@@ -3952,3 +3952,198 @@ def test_convert_chat_completion_messages_to_responses_api_tool_result_with_tool
 
     function_call_output = next(item for item in response if item.get("type") == "function_call_output")
     assert function_call_output["output"] == [{"type": "input_text", "text": "1 tool found"}]
+
+
+def _litellm_encoded_response_id(upstream_id: str) -> str:
+    from litellm.responses.utils import ResponsesAPIRequestUtils
+
+    return ResponsesAPIRequestUtils._build_responses_api_response_id(
+        custom_llm_provider="azure", model_id="deployment-1", response_id=upstream_id
+    )
+
+
+def test_transform_response_keeps_upstream_id_and_provider_extras():
+    from unittest.mock import Mock
+
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        LiteLLMResponsesTransformationHandler,
+    )
+    from litellm.types.llms.openai import ResponsesAPIResponse
+    from litellm.types.utils import ModelResponse, Usage
+
+    content_filters = [
+        {"blocked": False, "source_type": "prompt", "content_filter_results": {"hate": {"filtered": False}}}
+    ]
+    raw_response = ResponsesAPIResponse.model_validate(
+        {
+            "id": _litellm_encoded_response_id("resp_azure_123"),
+            "created_at": 1734366691,
+            "object": "response",
+            "model": "gpt-5.6",
+            "status": "completed",
+            "output": [
+                {
+                    "type": "function_call",
+                    "id": "fc_1",
+                    "call_id": "call_1",
+                    "name": "lookup_weather",
+                    "arguments": '{"city": "Seattle"}',
+                    "status": "completed",
+                }
+            ],
+            "parallel_tool_calls": True,
+            "tool_choice": "auto",
+            "tools": [],
+            "usage": {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15},
+            "service_tier": "default",
+            "content_filters": content_filters,
+            "max_tool_calls": None,
+            "background": False,
+            "top_logprobs": 0,
+            "store": True,
+        }
+    )
+    model_response = ModelResponse(
+        id="chatcmpl-local",
+        created=1734366691,
+        model=None,
+        object="chat.completion",
+        choices=[],
+        usage=Usage(completion_tokens=0, prompt_tokens=0, total_tokens=0),
+    )
+
+    result = LiteLLMResponsesTransformationHandler().transform_response(
+        model="gpt-5.6",
+        raw_response=raw_response,
+        model_response=model_response,
+        logging_obj=Mock(),
+        request_data={"model": "gpt-5.6"},
+        messages=[{"role": "user", "content": "What is the weather in Seattle?"}],
+        optional_params={},
+        litellm_params={},
+        encoding=Mock(),
+    )
+    dumped = result.model_dump()
+
+    assert dumped["id"] == "resp_azure_123"
+    assert dumped["object"] == "chat.completion"
+    assert dumped["service_tier"] == "default"
+    assert dumped["content_filters"] == content_filters
+    assert "max_tool_calls" not in dumped, "a null provider field must not appear as a null top-level key"
+    assert "output" not in dumped and "status" not in dumped, (
+        "Responses schema fields must not leak into the chat response"
+    )
+    assert not {"background", "top_logprobs", "store"} & dumped.keys(), (
+        "Responses API bookkeeping must not ride along as chat metadata"
+    )
+    assert dumped["choices"][0]["message"]["tool_calls"][0]["function"]["name"] == "lookup_weather"
+
+
+def test_bridged_response_is_priced_by_the_reported_service_tier():
+    from unittest.mock import Mock
+
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        LiteLLMResponsesTransformationHandler,
+    )
+    from litellm.types.llms.openai import ResponsesAPIResponse
+    from litellm.types.utils import ModelResponse
+
+    raw_response = ResponsesAPIResponse.model_validate(
+        {
+            "id": "resp_flex",
+            "created_at": 1734366691,
+            "object": "response",
+            "model": "gpt-5.4",
+            "status": "completed",
+            "output": [
+                {
+                    "type": "message",
+                    "id": "msg_1",
+                    "role": "assistant",
+                    "status": "completed",
+                    "content": [{"type": "output_text", "text": "hi", "annotations": []}],
+                }
+            ],
+            "usage": {"input_tokens": 1000, "output_tokens": 100, "total_tokens": 1100},
+            "service_tier": "flex",
+        }
+    )
+
+    result = LiteLLMResponsesTransformationHandler().transform_response(
+        model="gpt-5.4",
+        raw_response=raw_response,
+        model_response=ModelResponse(),
+        logging_obj=Mock(),
+        request_data={"model": "gpt-5.4"},
+        messages=[{"role": "user", "content": "hi"}],
+        optional_params={},
+        litellm_params={},
+        encoding=Mock(),
+    )
+    pricing = litellm.model_cost["gpt-5.4"]
+    flex_cost = 1000 * pricing["input_cost_per_token_flex"] + 100 * pricing["output_cost_per_token_flex"]
+    standard_cost = 1000 * pricing["input_cost_per_token"] + 100 * pricing["output_cost_per_token"]
+
+    cost = litellm.completion_cost(completion_response=result, custom_llm_provider="openai")
+
+    assert cost == pytest.approx(flex_cost)
+    assert cost < standard_cost
+
+
+def test_streaming_chunks_carry_the_upstream_response_id():
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        OpenAiResponsesToChatCompletionStreamIterator,
+    )
+
+    iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
+    encoded_id = _litellm_encoded_response_id("resp_azure_stream")
+    events = [
+        {"type": "response.created", "response": {"id": encoded_id, "output": []}},
+        {"type": "response.output_text.delta", "delta": "Hel"},
+        {"type": "response.completed", "response": {"id": encoded_id, "output": [{"type": "message"}]}},
+    ]
+
+    ids = [iterator.chunk_parser(event).id for event in events]
+
+    assert ids == ["resp_azure_stream"] * len(events), f"streamed chunks did not carry the upstream id: {ids}"
+
+
+def test_streaming_final_chunk_carries_provider_metadata():
+    from unittest.mock import MagicMock
+
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        OpenAiResponsesToChatCompletionStreamIterator,
+    )
+    from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+
+    iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
+    content_filters = [{"blocked": False, "source_type": "completion", "content_filter_results": {}}]
+    events = [
+        {"type": "response.created", "response": {"id": "resp_azure_stream", "output": []}},
+        {"type": "response.output_text.delta", "delta": "Hello"},
+        {
+            "type": "response.completed",
+            "response": {
+                "id": "resp_azure_stream",
+                "output": [{"type": "message"}],
+                "usage": {"input_tokens": 3, "output_tokens": 1, "total_tokens": 4},
+                "service_tier": "default",
+                "content_filters": content_filters,
+                "background": False,
+            },
+        },
+    ]
+    stream = CustomStreamWrapper(
+        completion_stream=iter([iterator.chunk_parser(event) for event in events]),
+        model="gpt-5.6",
+        custom_llm_provider="azure",
+        logging_obj=MagicMock(),
+    )
+
+    chunks = [chunk.model_dump() for chunk in stream]
+
+    assert chunks[-1]["choices"][0]["finish_reason"] == "stop"
+    assert chunks[-1]["service_tier"] == "default"
+    assert chunks[-1]["content_filters"] == content_filters
+    assert "background" not in chunks[-1]
+    assert all("service_tier" not in chunk for chunk in chunks[:-1])

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -4471,6 +4471,39 @@ def test_handle_anthropic_messages_response_logging_translates_bare_responses_ap
     assert result.usage.total_tokens == 18  # type: ignore[attr-defined]
 
 
+def test_handle_anthropic_messages_response_logging_keeps_the_served_response_id():
+    from openai.types.responses import ResponseOutputMessage, ResponseOutputText
+
+    from litellm.responses.utils import ResponsesAPIRequestUtils
+    from litellm.types.llms.openai import ResponseAPIUsage, ResponsesAPIResponse
+
+    served_id = ResponsesAPIRequestUtils._build_responses_api_response_id(
+        custom_llm_provider="openai", model_id="deployment-1", response_id="resp_upstream"
+    )
+    logging_obj = _anthropic_messages_logging_obj()
+    result = logging_obj._handle_anthropic_messages_response_logging(
+        result=ResponsesAPIResponse(
+            id=served_id,
+            created_at=1700000000,
+            output=[
+                ResponseOutputMessage(
+                    id="msg-1",
+                    type="message",
+                    role="assistant",
+                    status="completed",
+                    content=[ResponseOutputText(annotations=[], text="hi", type="output_text")],
+                )
+            ],
+            usage=ResponseAPIUsage(input_tokens=2, output_tokens=1, total_tokens=3),
+            service_tier="flex",
+        )
+    )
+
+    assert isinstance(result, ModelResponse)
+    assert result.id == served_id, "the spend log row must keep the id the caller was served"
+    assert result.service_tier == "flex"
+
+
 def test_handle_anthropic_messages_response_logging_passes_model_response_through():
     """Anthropic-native path already yields a ModelResponse; it must be returned unchanged."""
     logging_obj = _anthropic_messages_logging_obj()