From 6ae64adfd57cab0da73ee6b9377e4078aa87afed Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 30 Oct 2026 03:53:49 -0700 Subject: [PATCH] fix(anthropic): map tool search only beside a deferred tool, guard non-string tool types, and honor additional_drop_params tools --- .../messages/responses_adapters/handler.py | 3 +- .../responses_adapters/tool_translation.py | 16 +++++- .../responses_adapters/transformation.py | 4 +- .../responses/count_tokens/token_counter.py | 3 -- .../responses/count_tokens/transformation.py | 19 +------ .../llms/openai/responses/transformation.py | 10 ++-- .../test_responses_adapters_handler.py | 50 +++++++++++++++++++ .../count_tokens/test_token_counter.py | 13 +++-- 8 files changed, 84 insertions(+), 34 deletions(-) diff --git a/litellm/llms/anthropic/messages/responses_adapters/handler.py b/litellm/llms/anthropic/messages/responses_adapters/handler.py index 1df439bab4e..ef023e99ec6 100644 --- a/litellm/llms/anthropic/messages/responses_adapters/handler.py +++ b/litellm/llms/anthropic/messages/responses_adapters/handler.py @@ -24,6 +24,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import ( prompt_cache_breakpoint_for_wire, with_prompt_cache_breakpoint, ) +from litellm.litellm_core_utils.tool_fields import drop_policy from litellm.responses.utils import auto_reasoning_summary from litellm.types.integrations.custom_logger import CONVERTED_STREAM_KEYS from litellm.types.llms.anthropic import ( @@ -316,7 +317,7 @@ def _build_responses_kwargs( request_data["thinking"] = thinking if tool_choice: request_data["tool_choice"] = tool_choice - if tools: + if tools and "tools" not in drop_policy(forwarded_kwargs).additional_drop_params: request_data["tools"] = tools if top_p is not None: request_data["top_p"] = top_p diff --git a/litellm/llms/anthropic/messages/responses_adapters/tool_translation.py b/litellm/llms/anthropic/messages/responses_adapters/tool_translation.py index ecd95b60794..13a41f93d47 100644 --- a/litellm/llms/anthropic/messages/responses_adapters/tool_translation.py +++ b/litellm/llms/anthropic/messages/responses_adapters/tool_translation.py @@ -232,7 +232,21 @@ def _custom_function(name: str, tool: Mapping[str, object]) -> ResponsesTool: def is_tool_search(tool: Mapping[str, object]) -> bool: - return tool.get("type") in ANTHROPIC_TOOL_SEARCH_TOOL_TYPES + tool_type: Final = tool.get("type") + return isinstance(tool_type, str) and tool_type in ANTHROPIC_TOOL_SEARCH_TOOL_TYPES + + +def _defers_a_function(tool: Mapping[str, object]) -> bool: + translated: Final = _translated_tool(tool, False) if tool.get("defer_loading") is True else None + return isinstance(translated, ResponsesTool) and translated.payload.get("type") == "function" + + +def keeps_tool_search(tools: Sequence[Mapping[str, object]], *, maps_tool_search: bool) -> bool: + return ( + maps_tool_search + and any(is_tool_search(tool) for tool in tools) + and any(_defers_a_function(tool) for tool in tools) + ) def _with_deferral(translated: TranslatedTool, defer_loading: object) -> TranslatedTool: diff --git a/litellm/llms/anthropic/messages/responses_adapters/transformation.py b/litellm/llms/anthropic/messages/responses_adapters/transformation.py index 34c3d530fb5..c0765f18670 100644 --- a/litellm/llms/anthropic/messages/responses_adapters/transformation.py +++ b/litellm/llms/anthropic/messages/responses_adapters/transformation.py @@ -30,7 +30,7 @@ from litellm.llms.anthropic.messages.responses_adapters.tool_translation import ResponsesTool, TranslatedTool, UntranslatableTool, - is_tool_search, + keeps_tool_search, log_dropped_tools, responses_tool, unforceable_tool_choice_error, @@ -472,7 +472,7 @@ class LiteLLMAnthropicToResponsesAPIAdapter: drop_params: bool = True, maps_tool_search: bool = False, ) -> tuple[TranslatedTool, ...]: - runs_tool_search: Final = maps_tool_search and any(is_tool_search(tool) for tool in tools) + runs_tool_search: Final = keeps_tool_search(tools, maps_tool_search=maps_tool_search) translated: Final = tuple(responses_tool(tool, runs_tool_search=runs_tool_search) for tool in tools) untranslatable: Final = tuple(tool for tool in translated if isinstance(tool, UntranslatableTool)) if untranslatable and not drop_params: diff --git a/litellm/llms/openai/responses/count_tokens/token_counter.py b/litellm/llms/openai/responses/count_tokens/token_counter.py index 02e6c98ca07..75fe22ee928 100644 --- a/litellm/llms/openai/responses/count_tokens/token_counter.py +++ b/litellm/llms/openai/responses/count_tokens/token_counter.py @@ -148,7 +148,6 @@ class OpenAITokenCounter(BaseTokenCounter): messages_format: TokenCountMessagesFormat, output_format: JsonValue, api_base: str | None, - drop_params: bool, ) -> ResponsesCountRequestParts: match messages_format: case "anthropic": @@ -159,7 +158,6 @@ class OpenAITokenCounter(BaseTokenCounter): tools=tools, output_format=output_format, maps_tool_search=openai_runs_tool_search(model_to_use, api_base), - drop_params=drop_params, ) case "openai": input_items, instructions = OpenAICountTokensConfig.messages_to_responses_input( @@ -239,7 +237,6 @@ class OpenAITokenCounter(BaseTokenCounter): messages_format=messages_format, output_format=output_format, api_base=resolved.api_base, - drop_params=deployment_drops_params(deployment), ) # If no input items were produced (e.g., system-only messages), fall back to local counting diff --git a/litellm/llms/openai/responses/count_tokens/transformation.py b/litellm/llms/openai/responses/count_tokens/transformation.py index 8aad129da40..f723245a096 100644 --- a/litellm/llms/openai/responses/count_tokens/transformation.py +++ b/litellm/llms/openai/responses/count_tokens/transformation.py @@ -20,11 +20,7 @@ from litellm.llms.anthropic.messages.native.utils import normalize_native_messag from litellm.llms.anthropic.messages.responses_adapters.transformation import ( LiteLLMAnthropicToResponsesAPIAdapter, ) -from litellm.llms.openai.responses.transformation import keeps_anthropic_tool_search -from litellm.responses.litellm_completion_transformation.tool_search import ( - tools_with_anthropic_tool_search_mapped, - tools_without_anthropic_tool_search, -) +from litellm.responses.litellm_completion_transformation.tool_search import tools_with_anthropic_tool_search_mapped from litellm.responses.utils import ResponsesAPIRequestUtils from litellm.types.llms.anthropic import AnthropicMessagesRequest, AnthropicOutputSchema from litellm.types.llms.openai import ResponsesAPIRequestParams, replayable_responses_input @@ -165,14 +161,6 @@ def chat_content_blocks_to_responses_content( return "\n".join(part["text"] for part in parts if part["type"] == "input_text") -def _tools_inference_sends( - model: str, tools: Sequence[Mapping[str, object]], *, drop_params: bool -) -> Sequence[Mapping[str, object]] | None: - if keeps_anthropic_tool_search(model, tools, drop_params=drop_params): - return tools_with_anthropic_tool_search_mapped(tools) - return tools_without_anthropic_tool_search(tools) - - class OpenAICountTokensConfig: """ Configuration and transformation logic for OpenAI Responses API token counting. @@ -357,7 +345,6 @@ class OpenAICountTokensConfig: tools: Sequence[Mapping[str, object]] | None, output_format: JsonValue = None, maps_tool_search: bool = False, - drop_params: bool = False, ) -> ResponsesCountRequestParts: """Convert with the adapter /v1/messages inference uses, so the count sees the same input it bills.""" anthropic_request: Final = AnthropicMessagesRequest( @@ -379,9 +366,7 @@ class OpenAICountTokensConfig: translated.get("text"), ) counted_tools: Final = ( - _tools_inference_sends(model, translated_tools, drop_params=drop_params) - if maps_tool_search and translated_tools is not None - else translated_tools + tools_with_anthropic_tool_search_mapped(translated_tools) if maps_tool_search else translated_tools ) return input_items, instructions, None if counted_tools is None else tuple(counted_tools), text diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py index 3e2b58f7fbf..9b8515a9ecd 100644 --- a/litellm/llms/openai/responses/transformation.py +++ b/litellm/llms/openai/responses/transformation.py @@ -158,17 +158,12 @@ def _tool_deferral(tool: object) -> _ToolDeferral: return _ToolDeferral() -def _runs_tool_search(model: str, tools: Sequence[object]) -> bool: +def _runs_tool_search(model: str, tools: Sequence[ALL_RESPONSES_API_TOOL_PARAMS]) -> bool: return any(_tool_deferral(tool).defers for tool in tools) and litellm.supports_tool_search( model=model, custom_llm_provider=LlmProviders.OPENAI.value ) -def keeps_anthropic_tool_search(model: str, tools: Sequence[object], *, drop_params: bool) -> bool: - callers_own_search: Final = any(is_openai_tool_search_tool(tool) for tool in tools) - return callers_own_search or not drop_params or _runs_tool_search(model, tools) - - def _without_defer_loading(tool: ALL_RESPONSES_API_TOOL_PARAMS) -> ALL_RESPONSES_API_TOOL_PARAMS: deferral: Final = _tool_deferral(tool) if not deferral.defers: @@ -733,7 +728,8 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig): ) -> Sequence[ALL_RESPONSES_API_TOOL_PARAMS] | None: if tools is None or not any(is_anthropic_tool_search_chat_tool(tool) for tool in tools): return tools - if keeps_anthropic_tool_search(model, tools, drop_params=drops_params(litellm_params)): + callers_own_search: Final = any(is_openai_tool_search_tool(tool) for tool in tools) + if callers_own_search or not drops_params(litellm_params) or _runs_tool_search(model, tools): return tools_with_anthropic_tool_search_mapped(tools) verbose_logger.warning( "Dropping tool search under drop_params: OpenAI model '%s' cannot run it or no tool is deferred, so every " diff --git a/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_handler.py b/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_handler.py index 650d412acbf..fb96039e73a 100644 --- a/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_handler.py +++ b/tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_handler.py @@ -889,6 +889,56 @@ async def test_drop_params_drops_tool_search_and_its_deferral_on_an_openai_model ] +UNDEFERRED_WEATHER: Final = {key: value for key, value in DEFERRED_WEATHER.items() if key != "defer_loading"} + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "weather", + [ + pytest.param(UNDEFERRED_WEATHER, id="no-defer-loading"), + pytest.param({**DEFERRED_WEATHER, "defer_loading": "true"}, id="defer-loading-string"), + ], +) +async def test_tool_search_with_no_deferred_tool_answers_400_naming_drop_params_before_any_openai_call( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, weather: Mapping[str, object] +): + with pytest.raises(litellm.UnsupportedParamsError) as error: + await _openai_request_for(respx_mock, monkeypatch, [TOOL_SEARCH, weather], drop_params=False) + + assert error.value.status_code == 400 + assert "tool_search_tool_bm25_20251119" in error.value.message + assert "drop_params" in error.value.message + assert not respx_mock.calls + + +@pytest.mark.asyncio +@pytest.mark.parametrize("tool_type", [pytest.param(["bash"], id="list"), pytest.param({"name": "bash"}, id="object")]) +async def test_a_tool_whose_type_is_not_a_string_reaches_openai_as_a_function_beside_tool_search( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, tool_type: object +): + upstream: Final = await _openai_request_for( + respx_mock, monkeypatch, [TOOL_SEARCH, DEFERRED_WEATHER, {"type": tool_type, "name": "odd_tool"}] + ) + + assert [(tool["type"], tool.get("name")) for tool in _tools_sent(upstream)] == [ + ("tool_search", None), + ("function", "get_weather"), + ("function", "odd_tool"), + ] + + +@pytest.mark.asyncio +async def test_additional_drop_params_tools_sends_no_tools_even_beside_one_openai_cannot_run( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +): + upstream: Final = await _openai_request_for( + respx_mock, monkeypatch, [WEB_FETCH, TYPED_BASH], drop_params=False, additional_drop_params=["tools"] + ) + + assert "tools" not in json.loads(upstream.calls.last.request.content) + + @pytest.mark.asyncio async def test_a_tool_openai_cannot_run_answers_400_before_any_openai_call_without_drop_params( respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch diff --git a/tests/unit/llms/openai/responses/count_tokens/test_token_counter.py b/tests/unit/llms/openai/responses/count_tokens/test_token_counter.py index f6ff8aab381..3de097476c8 100644 --- a/tests/unit/llms/openai/responses/count_tokens/test_token_counter.py +++ b/tests/unit/llms/openai/responses/count_tokens/test_token_counter.py @@ -255,10 +255,17 @@ async def test_anthropic_tool_search_is_counted_as_the_openai_tool_search_infere pytest.param(UNDEFERRED_TOOL_SEARCH_TOOLS, [("function", "get_weather", None)], id="nothing-is-deferred"), ], ) -async def test_drop_params_counts_anthropic_tool_search_only_where_inference_keeps_it( - monkeypatch: pytest.MonkeyPatch, tools: list[dict[str, object]], expected: list[tuple[object, object, object]] +@pytest.mark.parametrize( + "litellm_params", + [pytest.param({}, id="without-drop-params"), pytest.param({"drop_params": True}, id="drop-params")], +) +async def test_anthropic_tool_search_is_counted_only_beside_a_deferred_tool( + monkeypatch: pytest.MonkeyPatch, + tools: list[dict[str, object]], + expected: list[tuple[object, object, object]], + litellm_params: Mapping[str, object], ) -> None: - assert await _counted_tools(monkeypatch, "gpt-5.6-luna", {"drop_params": True}, tools) == expected + assert await _counted_tools(monkeypatch, "gpt-5.6-luna", litellm_params, tools) == expected @pytest.mark.asyncio