fix(responses): keep access_programs on chat completions on OpenAI SDK 3

The Responses-to-chat bridge hid every field the installed SDK's Response
declared, so SDK 3 dropped access_programs, prompt_cache_options, and
prompt_cache_diagnostics and let metadata through on streams. A fixed list
gives the same output on SDK 2 and SDK 3.
This commit is contained in:
mateo-berri 2026-10-30 06:08:28 -07:00
parent 02ef44c388
commit e102f181ac
3 changed files with 107 additions and 4 deletions

View file

@ -1382,6 +1382,8 @@ jobs:
tests/unit/types/test_router.py \
tests/unit/test_exception_header_preservation.py \
tests/unit/integrations/otel/test_otel_v2_sources_of_truth.py::test_speech_binary_response_still_streaming_reports_the_bytes_downloaded_so_far \
tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py::test_non_stream_response_keeps_provider_metadata_and_drops_request_echoes \
tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py::test_streaming_final_chunk_keeps_provider_metadata_and_drops_request_echoes \
--tb=short -q
- name: Run workload identity tests with SDK 3

View file

@ -8,8 +8,6 @@ from itertools import accumulate, chain
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, TypedDict, TypeVar, Union, cast
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,
@ -92,8 +90,19 @@ if TYPE_CHECKING:
_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
_RESPONSES_API_ONLY_FIELDS: Final = frozenset(
(
*ResponsesAPIResponse.model_fields,
"background",
"completed_at",
"conversation",
"max_tool_calls",
"prompt",
"prompt_cache_key",
"prompt_cache_retention",
"safety_identifier",
"top_logprobs",
)
)

View file

@ -4992,6 +4992,98 @@ def test_streaming_final_chunk_carries_provider_metadata():
assert all("service_tier" not in chunk for chunk in chunks[:-1])
_RESPONSES_BODY_METADATA: Final = {
"access_programs": {"cyber": "daybreak_blue"},
"prompt_cache_options": {"mode": "implicit", "ttl": "30m"},
"prompt_cache_diagnostics": {"type": "cache_hit"},
}
_RESPONSES_REQUEST_ECHOES: Final = {
"metadata": {"trace": "abc"},
"background": False,
"completed_at": 1760144905,
"conversation": {"id": "conv_metadata"},
"max_tool_calls": 3,
"prompt": {"id": "pmpt_metadata"},
"prompt_cache_key": "cache-key",
"prompt_cache_retention": "24h",
"safety_identifier": "user-hash",
"top_logprobs": 0,
}
def test_non_stream_response_keeps_provider_metadata_and_drops_request_echoes() -> None:
from openai.types.responses import ResponseOutputMessage, ResponseOutputText
from litellm.types.llms.openai import ResponsesAPIResponse
from litellm.types.utils import ModelResponse
handler: Final = LiteLLMResponsesTransformationHandler()
output_message: Final = ResponseOutputMessage(
id="msg_metadata",
content=[ResponseOutputText(annotations=[], text="hi", type="output_text", logprobs=[])],
role="assistant",
status="completed",
type="message",
)
raw_response: Final = ResponsesAPIResponse(
id="resp_metadata",
created_at=1760144904,
model="gpt-5.6-sol",
object="response",
output=[output_message],
status="completed",
parallel_tool_calls=True,
tool_choice="auto",
tools=[],
**_RESPONSES_BODY_METADATA,
**_RESPONSES_REQUEST_ECHOES,
)
result: Final = handler.transform_response(
model="gpt-5.6-sol",
raw_response=raw_response,
model_response=ModelResponse(),
logging_obj=Mock(),
request_data={"model": "gpt-5.6-sol"},
messages=[{"role": "user", "content": "hi"}],
optional_params={},
litellm_params={"acompletion": False, "api_key": None},
encoding=Mock(),
api_key=None,
json_mode=None,
)
dumped: Final = result.model_dump()
assert {key: dumped.get(key) for key in _RESPONSES_BODY_METADATA} == _RESPONSES_BODY_METADATA
assert not set(_RESPONSES_REQUEST_ECHOES) & set(dumped)
def test_streaming_final_chunk_keeps_provider_metadata_and_drops_request_echoes() -> None:
from litellm.completion_extras.litellm_responses_transformation.transformation import (
OpenAiResponsesToChatCompletionStreamIterator,
)
iterator: Final = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
iterator.chunk_parser({"type": "response.created", "response": {"id": "resp_metadata", "output": []}})
final_chunk: Final = iterator.chunk_parser(
{
"type": "response.completed",
"response": {
"id": "resp_metadata",
"output": [{"type": "message"}],
"usage": {"input_tokens": 3, "output_tokens": 1, "total_tokens": 4},
**_RESPONSES_BODY_METADATA,
**_RESPONSES_REQUEST_ECHOES,
},
}
)
provider_fields: Final = final_chunk.provider_specific_fields or {}
assert {key: provider_fields.get(key) for key in _RESPONSES_BODY_METADATA} == _RESPONSES_BODY_METADATA
assert not set(_RESPONSES_REQUEST_ECHOES) & set(provider_fields)
def _system_input_item(text: str) -> dict[str, object]:
return {"type": "message", "role": "system", "content": [{"type": "input_text", "text": text}]}