diff --git a/litellm/llms/bedrock_mantle/responses/transformation.py b/litellm/llms/bedrock_mantle/responses/transformation.py index 53a3e634adf..95375399033 100644 --- a/litellm/llms/bedrock_mantle/responses/transformation.py +++ b/litellm/llms/bedrock_mantle/responses/transformation.py @@ -17,7 +17,7 @@ BaseAWSLLM._sign_request after the request body is finalized. import json from collections.abc import Mapping -from typing import Any, Final +from typing import Any, Final, cast # noqa: TID251 # map_openai_params returns the filtered params as a bare dict import httpx from typing_extensions import ReadOnly, TypedDict @@ -32,6 +32,7 @@ from litellm.llms.bedrock_mantle.common_utils import ( BedrockMantleAuthMixin, ) from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig +from litellm.responses.additional_tools import HoistedAdditionalTools, hoist_additional_tools from litellm.secret_managers.main import get_secret_str from litellm.types.llms.openai import ( ResponseInputParam, @@ -58,8 +59,6 @@ _BEDROCK_MANTLE_SUPPORTED_RESPONSE_TOOL_TYPES: Final = frozenset( _BEDROCK_MANTLE_SUPPORTED_SERVICE_TIERS: Final = frozenset({"auto", "default"}) _BEDROCK_MANTLE_OPENAI_PATH_SUPPORTED_REASONING_SUMMARIES: Final = frozenset({"auto"}) -_CODEX_ADDITIONAL_TOOLS_INPUT_ITEM_TYPE: Final = "additional_tools" - _CODEX_AGENT_MESSAGE_INPUT_ITEM_TYPE: Final = "agent_message" _CODEX_CONTEXT_COMPACTION_INPUT_ITEM_TYPE: Final = "context_compaction" _CODEX_LOCAL_SHELL_CALL_INPUT_ITEM_TYPE: Final = "local_shell_call" @@ -233,62 +232,29 @@ class BedrockMantleResponsesAPIConfig(BedrockMantleAuthMixin, OpenAIResponsesAPI litellm_params: GenericLiteLLMParams, headers: dict, ) -> dict: - remaining_input, hoisted_tools = self._hoist_codex_additional_tools(input) - normalized_input: Final = self._normalize_codex_input_items(remaining_input) - request_params: Final = ( - { - **response_api_optional_request_params, - "tools": [ - *(response_api_optional_request_params.get("tools") or []), - *hoisted_tools, - ], - } - if hoisted_tools - else response_api_optional_request_params + params: Final = cast( # cast-ok: the base signature leaves the params dict untyped + "ResponsesAPIOptionalRequestParams", response_api_optional_request_params ) + hoisted: Final = hoist_additional_tools(input, params.get("tools")) + normalized_input: Final = self._normalize_codex_input_items(hoisted.input) return super().transform_responses_api_request( model=model, input=normalized_input, - response_api_optional_request_params=request_params, + response_api_optional_request_params=self._params_with_hoisted_tools(params, hoisted), litellm_params=litellm_params, headers=headers, ) - @staticmethod - def _is_codex_additional_tools_item(item: Any) -> bool: - return isinstance(item, dict) and item.get("type") == _CODEX_ADDITIONAL_TOOLS_INPUT_ITEM_TYPE - - @staticmethod - def _tools_of_additional_tools_item(item: "dict[str, Any]") -> "list[Any]": - tools: Final = item.get("tools") - return tools if isinstance(tools, list) else [] - @classmethod - def _hoist_codex_additional_tools( - cls, - input: "str | ResponseInputParam", - ) -> "tuple[str | ResponseInputParam, list[Any]]": - """Codex's "responses lite" wire mode ships tool definitions inside - `input` as {"type": "additional_tools", "role": "developer", - "tools": [...]} items. api.openai.com accepts that item type; Mantle - rejects the whole request with 400 "Invalid 'input': value did not - match any expected variant" but accepts the same tools at the top - level, so move them there and strip the items from `input`. - """ - if not isinstance(input, list): - return input, [] - additional_tools_items: Final = [item for item in input if cls._is_codex_additional_tools_item(item)] - if not additional_tools_items: - return input, [] - remaining_input: Final = [item for item in input if not cls._is_codex_additional_tools_item(item)] - hoisted_tools = [tool for item in additional_tools_items for tool in cls._tools_of_additional_tools_item(item)] - verbose_logger.debug( - "Bedrock Mantle Responses API: hoisting %d tool(s) out of %d 'additional_tools' input item(s) " - "into the top-level tools param (Mantle rejects that input item type).", - len(hoisted_tools), - len(additional_tools_items), - ) - return remaining_input, cls._filter_unsupported_tools(hoisted_tools) + def _params_with_hoisted_tools( + cls, params: Mapping[str, object], hoisted: HoistedAdditionalTools + ) -> dict[str, object]: + if not hoisted.hoisted: + return dict(params) + supported_tools: Final = cls._filter_unsupported_tools(list(hoisted.tools)) + if supported_tools: + return {**params, "tools": supported_tools} + return {key: value for key, value in params.items() if key != "tools"} @staticmethod def _agent_message_text(item: "Mapping[str, object]") -> str: diff --git a/litellm/responses/additional_tools.py b/litellm/responses/additional_tools.py new file mode 100644 index 00000000000..ea0d7af350c --- /dev/null +++ b/litellm/responses/additional_tools.py @@ -0,0 +1,65 @@ +from collections.abc import Sequence +from dataclasses import dataclass +from typing import Final, cast # noqa: TID251 # validating the openai tool union strips vendor keys from raw tools + +from pydantic import BaseModel, ValidationError + +from litellm._logging import verbose_logger +from litellm.types.llms.openai import ALL_RESPONSES_API_TOOL_PARAMS, ResponseInputParam + +ADDITIONAL_TOOLS_INPUT_ITEM_TYPE: Final = "additional_tools" + + +class _InputItemType(BaseModel): + type: str = "" + + +class _AdditionalToolsItem(BaseModel): + tools: tuple[dict[str, object], ...] = () + + +@dataclass(frozen=True, slots=True) +class HoistedAdditionalTools: + input: str | ResponseInputParam + tools: tuple[ALL_RESPONSES_API_TOOL_PARAMS, ...] + hoisted: tuple[ALL_RESPONSES_API_TOOL_PARAMS, ...] + + +def _is_additional_tools_item(item: object) -> bool: + try: + return _InputItemType.model_validate(item).type == ADDITIONAL_TOOLS_INPUT_ITEM_TYPE + except ValidationError: + return False + + +def _tools_of_item(item: object) -> tuple[ALL_RESPONSES_API_TOOL_PARAMS, ...]: + try: + parsed: Final = _AdditionalToolsItem.model_validate(item) + except ValidationError: + return () + return tuple( + cast( + "ALL_RESPONSES_API_TOOL_PARAMS", tool + ) # cast-ok: nested tools carry the same raw tool JSON as top-level tools + for tool in parsed.tools + ) + + +def hoist_additional_tools( + input: str | ResponseInputParam, + tools: Sequence[ALL_RESPONSES_API_TOOL_PARAMS] | None, +) -> HoistedAdditionalTools: + existing: Final = tuple(tools or ()) + if isinstance(input, str): + return HoistedAdditionalTools(input=input, tools=existing, hoisted=()) + items: Final = tuple(item for item in input if _is_additional_tools_item(item)) + if not items: + return HoistedAdditionalTools(input=input, tools=existing, hoisted=()) + hoisted: Final = tuple(tool for item in items for tool in _tools_of_item(item)) + verbose_logger.debug( + "Responses API: hoisting %d tool(s) out of %d 'additional_tools' input item(s) into the top-level tools param.", + len(hoisted), + len(items), + ) + remaining_input: Final = [item for item in input if not _is_additional_tools_item(item)] + return HoistedAdditionalTools(input=remaining_input, tools=(*existing, *hoisted), hoisted=hoisted) diff --git a/litellm/responses/litellm_completion_transformation/custom_tools.py b/litellm/responses/litellm_completion_transformation/custom_tools.py index 4aa489d9e50..038964055c3 100644 --- a/litellm/responses/litellm_completion_transformation/custom_tools.py +++ b/litellm/responses/litellm_completion_transformation/custom_tools.py @@ -39,15 +39,27 @@ def openai_shaped_tool_call_item_id(item_type: str, tool_id: str) -> str: return f"{prefix}_{tool_id}" +class _ToolNameFields(BaseModel): + type: str = "" + name: str = "" + tools: tuple[object, ...] = () + + +def _custom_tool_names_of(tool: object) -> tuple[str, ...]: + try: + parsed: Final = _ToolNameFields.model_validate(tool) + except ValidationError: + return () + if parsed.type == "custom": + return (parsed.name,) if parsed.name else () + if parsed.type != "namespace": + return () + return tuple(name for nested_tool in parsed.tools for name in _custom_tool_names_of(nested_tool)) + + def extract_custom_tool_names(tools: Sequence[object] | None) -> set[str]: - """Extract names of tools originally defined as ``type: "custom"``.""" - if not tools: - return set() - names: Final[set[str]] = set() - for tool in tools: - if isinstance(tool, dict) and tool.get("type") == "custom" and "name" in tool: - names.add(tool["name"]) - return names + """Extract names of tools defined as ``type: "custom"``, at the top level or inside a ``namespace`` tool.""" + return {name for tool in tools or () for name in _custom_tool_names_of(tool)} def is_custom_tool_call(tool_name: str, custom_tool_names: set[str]) -> bool: diff --git a/litellm/responses/litellm_completion_transformation/handler.py b/litellm/responses/litellm_completion_transformation/handler.py index a0e8cd278e6..505b5b09433 100644 --- a/litellm/responses/litellm_completion_transformation/handler.py +++ b/litellm/responses/litellm_completion_transformation/handler.py @@ -6,6 +6,7 @@ from collections.abc import Coroutine, Mapping from typing import Final import litellm +from litellm.responses.additional_tools import hoist_additional_tools from litellm.responses.litellm_completion_transformation.streaming_iterator import ( LiteLLMCompletionStreamingIterator, ) @@ -37,11 +38,16 @@ class LiteLLMCompletionTransformationHandler: | BaseResponsesAPIStreamingIterator | Coroutine[object, object, ResponsesAPIResponse | BaseResponsesAPIStreamingIterator] ): + hoisted: Final = hoist_additional_tools(input, responses_api_request.get("tools")) + bridged_input: Final = hoisted.input + bridged_request: Final[ResponsesAPIOptionalRequestParams] = ( + {**responses_api_request, "tools": list(hoisted.tools)} if hoisted.hoisted else responses_api_request + ) litellm_completion_request: Final[dict] = ( LiteLLMCompletionResponsesConfig.transform_responses_api_request_to_chat_completion_request( model=model, - input=input, - responses_api_request=responses_api_request, + input=bridged_input, + responses_api_request=bridged_request, custom_llm_provider=custom_llm_provider, stream=stream, extra_headers=extra_headers, @@ -52,8 +58,8 @@ class LiteLLMCompletionTransformationHandler: if _is_async: return self.async_response_api_handler( litellm_completion_request=litellm_completion_request, - request_input=input, - responses_api_request=responses_api_request, + request_input=bridged_input, + responses_api_request=bridged_request, **kwargs, ) @@ -70,8 +76,8 @@ class LiteLLMCompletionTransformationHandler: responses_api_response: Final[ResponsesAPIResponse] = ( LiteLLMCompletionResponsesConfig.transform_chat_completion_response_to_responses_api_response( chat_completion_response=litellm_completion_response, - request_input=input, - responses_api_request=responses_api_request, + request_input=bridged_input, + responses_api_request=bridged_request, ) ) @@ -81,8 +87,8 @@ class LiteLLMCompletionTransformationHandler: return LiteLLMCompletionStreamingIterator( model=model, litellm_custom_stream_wrapper=litellm_completion_response, - request_input=input, - responses_api_request=responses_api_request, + request_input=bridged_input, + responses_api_request=bridged_request, custom_llm_provider=custom_llm_provider, litellm_metadata=kwargs.get("litellm_metadata", {}), ) diff --git a/litellm/responses/litellm_completion_transformation/transformation.py b/litellm/responses/litellm_completion_transformation/transformation.py index e8aacac9e67..27756b405ec 100644 --- a/litellm/responses/litellm_completion_transformation/transformation.py +++ b/litellm/responses/litellm_completion_transformation/transformation.py @@ -1890,9 +1890,21 @@ class LiteLLMCompletionResponsesConfig: namespace_tool: NamespaceTool, nested: bool, ) -> ChatCompletionToolParam | None: - if nested and namespace_tool.get("type") != "function": + tool_type: Final = namespace_tool.get("type") + if nested and tool_type not in ("function", "custom"): return None + raw_description: Final = str(namespace_tool.get("description") or "") + description: Final = ( + f"{namespace_description}{NAMESPACE_DESCRIPTION_SEPARATOR}{raw_description}" + if nested and namespace_description and raw_description + else namespace_description + if nested and namespace_description + else raw_description + ) + if nested and tool_type == "custom": + return convert_custom_tool_to_function_tool({**namespace_tool, "description": description}) + raw_parameters: Final = namespace_tool.get("parameters") parameters: Final = ( MappingProxyType(raw_parameters) if isinstance(raw_parameters, Mapping) else MappingProxyType({}) @@ -1901,14 +1913,6 @@ class LiteLLMCompletionResponsesConfig: parameters if parameters and "type" in parameters else MappingProxyType({**parameters, "type": "object"}) ) tool_name: Final = str(namespace_tool.get("name") or "") - raw_description: Final = str(namespace_tool.get("description") or "") - description: Final = ( - f"{namespace_description}{NAMESPACE_DESCRIPTION_SEPARATOR}{raw_description}" - if nested and namespace_description and raw_description - else namespace_description - if nested and namespace_description - else raw_description - ) chat_tool_name: Final = f"{namespace}__{tool_name}" if nested else tool_name function: Final = ChatCompletionToolParamFunctionChunk( name=chat_tool_name, diff --git a/tests/test_litellm/responses/litellm_completion_transformation/test_handler.py b/tests/test_litellm/responses/litellm_completion_transformation/test_handler.py index 2cfec6a1844..b78dabbfe48 100644 --- a/tests/test_litellm/responses/litellm_completion_transformation/test_handler.py +++ b/tests/test_litellm/responses/litellm_completion_transformation/test_handler.py @@ -68,3 +68,105 @@ async def test_async_fallback_tags_skip_responses_api_bridge(): await coro assert captured.get("_skip_responses_api_bridge") is True + + +_CODEX_ADDITIONAL_TOOLS_ITEM = { + "type": "additional_tools", + "id": "at_codex", + "role": "developer", + "tools": [ + { + "type": "namespace", + "name": "functions", + "description": "", + "tools": [ + { + "type": "custom", + "name": "exec", + "description": "Runs a shell command.", + "format": {"type": "grammar", "syntax": "lark", "definition": "start: /.+/"}, + }, + { + "type": "function", + "name": "wait", + "description": "Waits for a background command.", + "parameters": {"type": "object", "properties": {"id": {"type": "string"}}}, + }, + ], + } + ], +} +_CODEX_INPUT = [_CODEX_ADDITIONAL_TOOLS_ITEM, {"type": "message", "role": "user", "content": "Run ls"}] + + +def test_sync_fallback_hoists_additional_tools_input_items_into_chat_tools(): + handler = LiteLLMCompletionTransformationHandler() + captured: dict = {} + + def fake_completion(**kwargs): + captured.update(kwargs) + raise _StopForwarding() + + with patch("litellm.completion", fake_completion): # test-quality-ok: no DI seam; the file stubs this same boundary + with pytest.raises(_StopForwarding): + handler.response_api_handler( + model="bedrock/us.openai.gpt-5.6", + input=_CODEX_INPUT, + responses_api_request={}, + custom_llm_provider="bedrock", + _is_async=False, + ) + + assert [message["role"] for message in captured["messages"]] == ["user"] + functions_by_name = {tool["function"]["name"]: tool["function"] for tool in captured["tools"]} + assert set(functions_by_name) == {"exec", "functions__wait"} + assert set(functions_by_name["exec"]["parameters"]["properties"]) == {"content"} + + +@pytest.mark.asyncio +async def test_async_fallback_returns_hoisted_nested_custom_tool_call_as_custom_tool_call(): + from litellm.responses.litellm_completion_transformation.transformation import TOOL_CALLS_CACHE + from litellm.types.utils import ChatCompletionMessageToolCall, Choices, Function, Message, ModelResponse + + handler = LiteLLMCompletionTransformationHandler() + tool_call_id = "call_exec_hoisted" + + async def fake_acompletion(**kwargs): + return ModelResponse( + id="chatcmpl-exec", + created=1, + model="us.openai.gpt-5.6", + object="chat.completion", + choices=[ + Choices( + finish_reason="tool_calls", + index=0, + message=Message( + content=None, + role="assistant", + tool_calls=[ + ChatCompletionMessageToolCall( + id=tool_call_id, + type="function", + function=Function(name="exec", arguments='{"content": "ls"}'), + ) + ], + ), + ) + ], + ) + + try: + with patch("litellm.acompletion", fake_acompletion): # test-quality-ok: no DI seam; file stubs this boundary + response = await handler.response_api_handler( + model="bedrock/us.openai.gpt-5.6", + input=_CODEX_INPUT, + responses_api_request={}, + custom_llm_provider="bedrock", + _is_async=True, + ) + finally: + TOOL_CALLS_CACHE.delete_cache(key=tool_call_id) + + tool_calls = [(item.type, item.name, item.input) for item in response.output if item.type == "custom_tool_call"] + assert tool_calls == [("custom_tool_call", "exec", "ls")] diff --git a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py index 3c78bbf79d7..1f497597a11 100644 --- a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py +++ b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py @@ -2506,6 +2506,7 @@ class TestToolTransformation: "tools": [ "ignored", {"type": "namespace", "name": "ignored"}, + {"type": "web_search", "name": "ignored"}, { "type": "function", "name": "spawn_agent", @@ -2527,6 +2528,36 @@ class TestToolTransformation: "type": "object", } + def test_transform_nested_namespace_custom_tool_becomes_a_content_function_under_its_short_name(self): + namespace_tool = { + "type": "namespace", + "name": "functions", + "description": "Codex shell tools.", + "tools": [ + { + "type": "custom", + "name": "exec", + "description": "Runs a shell command.", + "format": {"type": "grammar", "syntax": "lark", "definition": "start: /.+/"}, + }, + ], + } + + result_tools, _ = ( + LiteLLMCompletionResponsesConfig.transform_responses_api_tools_to_chat_completion_tools( + tools=[namespace_tool] + ) + ) + + assert len(result_tools) == 1 + function = result_tools[0]["function"] + assert function["name"] == "exec" + assert function["description"].startswith("Codex shell tools.") + assert "Runs a shell command." in function["description"] + assert "start: /.+/" in function["description"] + assert function["parameters"]["required"] == ["content"] + assert function["parameters"]["properties"]["content"]["type"] == "string" + @pytest.mark.parametrize( "model, custom_llm_provider", [ @@ -3786,6 +3817,66 @@ class TestEnsureOutputItemContentPartAdded: assert added.item.name == "spawn_agent" assert added.item.namespace == "collaboration" + def test_streaming_nested_custom_tool_call_comes_back_as_custom_tool_call(self): + from litellm.responses.litellm_completion_transformation.custom_tools import extract_custom_tool_names + + iterator = self._make_iterator() + iterator.responses_api_request = { + "tools": [ + { + "type": "namespace", + "name": "functions", + "tools": [ + { + "type": "custom", + "name": "exec", + "format": {"type": "grammar", "syntax": "lark", "definition": "start: /.+/"}, + } + ], + } + ] + } + iterator._custom_tool_names = extract_custom_tool_names(iterator.responses_api_request.get("tools")) + iterator._namespace_tool_names = LiteLLMCompletionResponsesConfig.namespace_tool_name_map( + iterator.responses_api_request.get("tools") + ) + + iterator._queue_tool_call_delta_events( + [{"index": 0, "id": "call_exec", "function": {"name": "exec", "arguments": '{"content":"ls"}'}}] + ) + iterator._queue_final_tool_call_done_events( + ModelResponse( + id="chatcmpl-exec", + created=1, + model="us.openai.gpt-5.6", + object="chat.completion", + choices=[ + Choices( + finish_reason="tool_calls", + index=0, + message=Message( + content=None, + role="assistant", + tool_calls=[ + ChatCompletionMessageToolCall( + id="call_exec", + type="function", + function=Function(name="exec", arguments='{"content":"ls"}'), + ) + ], + ), + ) + ], + ) + ) + + added = iterator._pending_tool_events[0] + assert added.item.type == "custom_tool_call" + assert added.item.name == "exec" + done = iterator._pending_tool_events[-1] + assert done.item.type == "custom_tool_call" + assert done.item.input == "ls" + def test_streaming_unqualified_namespace_tool_calls_restore_namespace(self): """A unique nested tool name without the namespace still maps back.""" iterator = self._make_iterator() diff --git a/tests/test_litellm/responses/test_additional_tools.py b/tests/test_litellm/responses/test_additional_tools.py new file mode 100644 index 00000000000..bef3b27eacd --- /dev/null +++ b/tests/test_litellm/responses/test_additional_tools.py @@ -0,0 +1,48 @@ +from litellm.responses.additional_tools import hoist_additional_tools + +_EXEC_TOOL = {"type": "custom", "name": "exec", "format": {"type": "grammar", "syntax": "lark", "definition": "start: /.+/"}} +_WAIT_TOOL = {"type": "function", "name": "wait", "parameters": {"type": "object", "properties": {}}} +_TOP_LEVEL_TOOL = {"type": "function", "name": "top_level", "parameters": {"type": "object", "properties": {}}} +_USER_MESSAGE = {"type": "message", "role": "user", "content": "Run ls"} + + +def test_string_input_passes_through_with_existing_tools(): + hoisted = hoist_additional_tools("hello", [_TOP_LEVEL_TOOL]) + + assert hoisted.input == "hello" + assert hoisted.tools == (_TOP_LEVEL_TOOL,) + assert hoisted.hoisted == () + + +def test_input_without_additional_tools_items_is_returned_untouched(): + request_input = [_USER_MESSAGE] + + hoisted = hoist_additional_tools(request_input, None) + + assert hoisted.input is request_input + assert hoisted.tools == () + assert hoisted.hoisted == () + + +def test_additional_tools_items_are_stripped_and_appended_after_top_level_tools_in_item_order(): + request_input = [ + {"type": "additional_tools", "id": "at_1", "role": "developer", "tools": [_EXEC_TOOL]}, + _USER_MESSAGE, + {"type": "additional_tools", "id": "at_2", "role": "developer", "tools": [_WAIT_TOOL]}, + ] + + hoisted = hoist_additional_tools(request_input, [_TOP_LEVEL_TOOL]) + + assert hoisted.input == [_USER_MESSAGE] + assert hoisted.tools == (_TOP_LEVEL_TOOL, _EXEC_TOOL, _WAIT_TOOL) + assert hoisted.hoisted == (_EXEC_TOOL, _WAIT_TOOL) + + +def test_additional_tools_item_without_a_tools_list_is_stripped_and_contributes_nothing(): + request_input = [{"type": "additional_tools", "id": "at_1", "role": "developer", "tools": "exec"}, _USER_MESSAGE] + + hoisted = hoist_additional_tools(request_input, None) + + assert hoisted.input == [_USER_MESSAGE] + assert hoisted.tools == () + assert hoisted.hoisted == () diff --git a/tests/test_litellm/responses/test_custom_tool_call.py b/tests/test_litellm/responses/test_custom_tool_call.py index e80301c3b2f..2ed71ee3ecf 100644 --- a/tests/test_litellm/responses/test_custom_tool_call.py +++ b/tests/test_litellm/responses/test_custom_tool_call.py @@ -55,6 +55,24 @@ class TestCustomToolUtilities: names = extract_custom_tool_names(tools) assert names == set() + def test_extract_custom_tool_names_walks_namespace_tools(self): + tools = [ + {"type": "function", "name": "regular_tool"}, + { + "type": "namespace", + "name": "functions", + "tools": [ + {"type": "custom", "name": "exec"}, + {"type": "function", "name": "wait"}, + "ignored", + ], + }, + {"type": "namespace", "name": "empty", "tools": "not-a-list"}, + ] + + names = extract_custom_tool_names(tools) + assert names == {"exec"} + def test_extract_custom_tool_names_none(self): """Test extraction with None input.""" names = extract_custom_tool_names(None)