fix(responses): normalize string input to canonical list before provider dispatch

OpenAI's Responses API documents input as accepting either a string or an
array. Provider transforms that only accept the array form (e.g. the chatgpt
deployment in #41963) turn a spec-compliant string request into a 400 that
looks like a client bug.

Add _normalize_responses_api_string_input() and call it once, centrally, in
responses() right before dispatch so every lane - MCP gateway, emulated file
search, chat-completions bridge, and native providers - sees the canonical
array shape. List input passes through untouched.

Closes #41963

Co-authored-by: arena-agent <agent@arena.ai>
This commit is contained in:
ravigohel142996 2026-09-19 16:51:18 +00:00
parent 209a780992
commit 8df99d83d3
7 changed files with 87 additions and 8 deletions

View file

@ -569,6 +569,15 @@ def _prompt_management_sees_a_provisional_message_list(
kwargs.pop(CARRY_UNMATCHED_MESSAGE_POINTS, None)
def _normalize_responses_api_string_input(input: str | ResponseInputParam) -> ResponseInputParam:
"""Return list-form ``input``; OpenAI's Responses API accepts a plain string, but provider transforms
that only handle the array shape turn it into a spurious client-facing 400 (#41963).
"""
if isinstance(input, str):
return cast(ResponseInputParam, [{"role": "user", "content": input}])
return input
@client
async def aresponses(
input: str | ResponseInputParam,
@ -1232,6 +1241,9 @@ def responses(
#########################################################
input, tools = _apply_managed_file_id_mapping(input=input, tools=tools, kwargs=kwargs, local_vars=local_vars)
input = _normalize_responses_api_string_input(input) # rebind-ok: dispatch reads this name
local_vars["input"] = input
#########################################################
# Native MCP Responses API
#########################################################

View file

@ -227,7 +227,8 @@ def _assert_native_responses_request(route, expected_url, expected_model):
assert f"{request.url.scheme}://{request.url.host}{request.url.path}" == expected_url
assert request.headers["api-key"] == "fake-key"
assert body["model"] == expected_model
assert body["input"] == "What is the weather in SF?"
# string input is normalized to the canonical list before provider dispatch
assert body["input"] == [{"role": "user", "content": "What is the weather in SF?"}]
assert "messages" not in body
assert body["reasoning"] == {"effort": "high"}
assert body["tools"] == [WEATHER_TOOL]

View file

@ -321,7 +321,7 @@ def test_responses_call_keeps_a_developer_item_with_non_text_parts_in_place_as_a
)
def test_responses_call_forwards_string_input_and_instructions_unchanged() -> None:
def test_responses_call_normalizes_string_input_and_forwards_instructions_unchanged() -> None:
client: Final = _mock_http_client(_fireworks_response("accounts/fireworks/models/kimi-k3"))
with patch(HTTPX_CLIENT_FACTORY, return_value=client):
litellm.responses(
@ -332,7 +332,8 @@ def test_responses_call_forwards_string_input_and_instructions_unchanged() -> No
)
_, _, body = _sent_request(client)
assert body["instructions"] == "Answer with exactly one word."
assert body["input"] == "What is the capital of France?"
# string input is normalized to the canonical list before provider dispatch
assert body["input"] == [{"role": "user", "content": "What is the capital of France?"}]
def test_transform_request_forwards_non_string_instructions_and_input_untouched() -> None:

View file

@ -1,12 +1,16 @@
import inspect
from collections.abc import Awaitable, Callable, Mapping
from typing import Final, cast # noqa: TID251 # narrows legacy callable signatures for inspect
from typing import Any, Final, cast # noqa: TID251 # narrows legacy callable signatures for inspect
from unittest.mock import MagicMock, patch
import pytest
import litellm
from litellm.responses import dispatch as responses_dispatch
from litellm.responses import main as python_responses
from litellm.responses.main import (
_normalize_responses_api_string_input, # pyright: ignore[reportPrivateUsage] # regression target
)
from litellm.responses.dispatch import (
_ADISPATCH, # pyright: ignore[reportPrivateUsage] # tests configured dispatch
_DISPATCH, # pyright: ignore[reportPrivateUsage] # tests configured dispatch
@ -320,3 +324,54 @@ def test_responses_with_retries_uses_the_dispatch_entrypoint(monkeypatch: pytest
assert result is expected
assert calls[0]["num_retries"] == 0
assert calls[0]["max_retries"] == 0
class TestNormalizeResponsesApiStringInputForIssue41963:
"""Regression tests for https://github.com/BerriAI/litellm/issues/41963."""
def test_string_input_becomes_canonical_list(self) -> None:
assert _normalize_responses_api_string_input("Reply with exactly ROUTE_OK.") == [
{"role": "user", "content": "Reply with exactly ROUTE_OK."}
]
def test_empty_string_is_normalized(self) -> None:
assert _normalize_responses_api_string_input("") == [{"role": "user", "content": ""}]
def test_list_input_passes_through_untouched(self) -> None:
original: Final = [{"role": "user", "content": "Reply with exactly ROUTE_OK."}]
assert _normalize_responses_api_string_input(original) == original
def test_native_provider_receives_normalized_list_for_string_input(self) -> None:
captured: Final[dict[str, Any]] = {}
def capture_handler(**kwargs: Any) -> None:
captured.update(kwargs)
with patch("litellm.responses.main.base_llm_http_handler") as mock_handler: # test-quality-ok: the dispatch seam is the behavior under test; no DI seam exposes what input the provider receives
mock_handler.response_api_handler.side_effect = capture_handler
python_responses.responses(
input="Reply with exactly ROUTE_OK.",
model="gpt-4o",
litellm_logging_obj=MagicMock(),
api_key="sk-test",
)
assert captured["input"] == [{"role": "user", "content": "Reply with exactly ROUTE_OK."}]
def test_native_provider_receives_list_content_intact_for_list_input(self) -> None:
original: Final = [{"role": "user", "content": "Reply with exactly ROUTE_OK."}]
captured: Final[dict[str, Any]] = {}
def capture_handler(**kwargs: Any) -> None:
captured.update(kwargs)
with patch("litellm.responses.main.base_llm_http_handler") as mock_handler: # test-quality-ok: same dispatch seam as the string-input case
mock_handler.response_api_handler.side_effect = capture_handler
python_responses.responses(
input=original,
model="gpt-4o",
litellm_logging_obj=MagicMock(),
api_key="sk-test",
)
assert isinstance(captured["input"], list)
assert captured["input"] == original

View file

@ -94,6 +94,9 @@ async def test_aresponses_context_management_and_shell_request_body_matches_expe
assert the httpx POST request body matches the expected JSON.
"""
expected_body = _load_expected_body("context_management_and_shell.json")
call_input = expected_body["input"]
# string input is normalized to the canonical list before dispatch
expected_body["input"] = [{"role": "user", "content": call_input}]
with patch(
"litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
@ -103,7 +106,7 @@ async def test_aresponses_context_management_and_shell_request_body_matches_expe
await litellm.aresponses(
model="openai/gpt-4o",
input=expected_body["input"],
input=call_input,
context_management=expected_body["context_management"],
tools=expected_body["tools"],
tool_choice=expected_body["tool_choice"],
@ -121,6 +124,9 @@ async def test_aresponses_azure_shell_tool_request_body_matches_expected():
assert the httpx POST request body carries the shell tool verbatim.
"""
expected_body = _load_expected_body("azure_shell_tool.json")
call_input = expected_body["input"]
# string input is normalized to the canonical list before dispatch
expected_body["input"] = [{"role": "user", "content": call_input}]
with patch(
"litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
@ -135,7 +141,7 @@ async def test_aresponses_azure_shell_tool_request_body_matches_expected():
api_base="https://fake-resource.openai.azure.com",
api_key="fake-api-key",
api_version="2025-03-01-preview",
input=expected_body["input"],
input=call_input,
tools=expected_body["tools"],
tool_choice=expected_body["tool_choice"],
max_output_tokens=expected_body["max_output_tokens"],

View file

@ -141,7 +141,8 @@ async def test_opt_in_forwards_responses_natively():
assert chat_route.call_count == 0
request = responses_route.calls.last.request
assert request.headers["authorization"] == "Bearer sk-backend"
assert json.loads(request.content)["input"] == "hi"
# string input is normalized to the canonical list before native forwarding
assert json.loads(request.content)["input"] == [{"role": "user", "content": "hi"}]
assert isinstance(result, ResponsesAPIResponse)
assert result.output[0].content[0].text == "native"

View file

@ -200,7 +200,10 @@ class TestTextFormatConversion:
), "schema should have confidence property"
# Validate other request parameters
assert captured_request["input"] == "What is the capital of France?"
# string input is normalized to the canonical list before dispatch
assert captured_request["input"] == [
{"role": "user", "content": "What is the capital of France?"}
]
# Validate the response
print("Response:", json.dumps(response, indent=4, default=str))