Merge 34353d15d4 into aeede53112

litellm/llms/anthropic/messages/adapters/transformation.py

@@ -1116,12 +1116,7 @@ class LiteLLMAnthropicMessagesAdapter:
         unknown backend, because that provider declares this param and forwards it to a proxy
         that resolves the real target itself, where a derived cache key has no such guarantee.
         """
-        if not model or not custom_llm_provider:
-            return False
-        supported_params: Final = litellm.get_supported_openai_params(
-            model=model, custom_llm_provider=custom_llm_provider
-        )
-        return "reasoning_effort" in (supported_params or ())
+        return LiteLLMAnthropicMessagesAdapter._target_declares("reasoning_effort", model, custom_llm_provider)
 
     def _translate_metadata_to_openai(
         self,
@@ -1150,16 +1145,43 @@ class LiteLLMAnthropicMessagesAdapter:
         self,
         anthropic_message_request: AnthropicMessagesRequest,
         new_kwargs: ChatCompletionRequest,
+        custom_llm_provider: str | None = None,
     ) -> None:
-        """Translate Anthropic tool_choice to OpenAI format."""
+        """Translate Anthropic tool_choice to OpenAI format.
+
+        ``disable_parallel_tool_use`` becomes ``parallel_tool_calls=False`` only for a target that
+        declares the param, so a provider without it keeps answering instead of rejecting the
+        request the way it already does an OpenAI-shaped ``parallel_tool_calls``.
+        """
         if "tool_choice" not in anthropic_message_request:
             return
         tool_choice: Final = anthropic_message_request["tool_choice"]
         if not tool_choice:
             return
-        new_kwargs["tool_choice"] = self.translate_anthropic_tool_choice_to_openai(
-            tool_choice=cast(AnthropicMessagesToolChoice, tool_choice)
+        anthropic_tool_choice: Final = cast(AnthropicMessagesToolChoice, tool_choice)
+        new_kwargs["tool_choice"] = self.translate_anthropic_tool_choice_to_openai(tool_choice=anthropic_tool_choice)
+        if anthropic_tool_choice.get("disable_parallel_tool_use") is True and self._target_declares_parallel_tool_calls(
+            anthropic_message_request["model"], custom_llm_provider
+        ):
+            new_kwargs["parallel_tool_calls"] = False  # rebind-ok: out-param store like tool_choice above
+
+    @staticmethod
+    def _target_declares_parallel_tool_calls(model: str | None, custom_llm_provider: str | None) -> bool:
+        """Whether the target declares ``parallel_tool_calls`` among its supported params.
+
+        Without a resolved provider the flag stays behind, which is what this bridge sent before
+        it carried one at all, for the same reasons as ``_target_declares_reasoning_effort``.
+        """
+        return LiteLLMAnthropicMessagesAdapter._target_declares("parallel_tool_calls", model, custom_llm_provider)
+
+    @staticmethod
+    def _target_declares(param: str, model: str | None, custom_llm_provider: str | None) -> bool:
+        if not model or not custom_llm_provider:
+            return False
+        supported_params: Final = litellm.get_supported_openai_params(
+            model=model, custom_llm_provider=custom_llm_provider
         )
+        return param in (supported_params or ())
 
     def _translate_stop_sequences_to_openai(
         self,
@@ -1384,6 +1406,7 @@ class LiteLLMAnthropicMessagesAdapter:
         self._translate_tool_choice_to_openai(
             anthropic_message_request=anthropic_message_request,
             new_kwargs=new_kwargs,
+            custom_llm_provider=custom_llm_provider,
         )
         ## CONVERT TOOLS
         tool_name_mapping = self._translate_tools_to_openai(

litellm/llms/anthropic/messages/responses_adapters/handler.py

@@ -173,12 +173,20 @@ def _forwarded_kwargs(extra_kwargs: Mapping[str, object] | None) -> Mapping[str,
 
 
 def _provider_returns_encrypted_reasoning(model: str, custom_llm_provider: object) -> bool:
+    return _responses_config_declares("include", model, custom_llm_provider)
+
+
+def _provider_declares_parallel_tool_calls(model: str, custom_llm_provider: object) -> bool:
+    return _responses_config_declares("parallel_tool_calls", model, custom_llm_provider)
+
+
+def _responses_config_declares(param: str, model: str, custom_llm_provider: object) -> bool:
     provider: Final = (
         custom_llm_provider if isinstance(custom_llm_provider, str) else litellm.get_llm_provider(model=model)[1]
     )
     provider_model: Final = local_model_name(model, provider)
     responses_config: Final = ProviderConfigManager.get_provider_responses_api_config(provider, provider_model)
-    return responses_config is not None and "include" in responses_config.get_supported_openai_params(provider_model)
+    return responses_config is not None and param in responses_config.get_supported_openai_params(provider_model)
 
 
 def _reads_prompt_cache_breakpoint(model: str, forwarded_kwargs: Mapping[str, object]) -> bool:
@@ -333,6 +341,9 @@ def _build_responses_kwargs(
         ),
         auto_summary_level=auto_reasoning_summary(model, forwarded_kwargs.get("custom_llm_provider")),
         drop_params=drops_unsupported_params(forwarded_kwargs.get("drop_params")),
+        declares_parallel_tool_calls=_provider_declares_parallel_tool_calls(
+            model, forwarded_kwargs.get("custom_llm_provider")
+        ),
     )
 
     # Normalize reasoning effort based on model capabilities

litellm/llms/anthropic/messages/responses_adapters/transformation.py

@@ -597,11 +597,15 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
         auto_summary_level: str = "detailed",
         *,
         drop_params: bool = True,
+        declares_parallel_tool_calls: bool = True,
     ) -> dict[str, Any]:
         """
         Translate a full Anthropic /v1/messages request dict to
         litellm.responses() / litellm.aresponses() kwargs.
 
+        ``declares_parallel_tool_calls`` says whether the provider's Responses API accepts
+        ``parallel_tool_calls``; ``tool_choice.disable_parallel_tool_use`` maps onto it only then.
+
         ``include_encrypted_reasoning`` asks the provider for ``reasoning.encrypted_content``
         on every call, so a reasoning model's items can be replayed intact next turn even
         when the client sent no ``thinking`` block; pass False for a provider whose
@@ -666,9 +670,10 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
         # tool_choice
         tool_choice: Final = anthropic_request.get("tool_choice")
         if tool_choice:
-            responses_kwargs["tool_choice"] = self.translate_tool_choice_to_responses_api(
-                cast(AnthropicMessagesToolChoice, tool_choice)
-            )
+            anthropic_tool_choice: Final = cast(AnthropicMessagesToolChoice, tool_choice)
+            responses_kwargs["tool_choice"] = self.translate_tool_choice_to_responses_api(anthropic_tool_choice)
+            if anthropic_tool_choice.get("disable_parallel_tool_use") is True and declares_parallel_tool_calls:
+                responses_kwargs["parallel_tool_calls"] = False
 
         # thinking -> reasoning
         thinking: Final = anthropic_request.get("thinking")

tests/integration/translation/messages/params/test_messages_params_tool_choice.py

@@ -0,0 +1,97 @@
+from dataclasses import replace
+from typing import Final
+
+import pytest
+from integration._support.client import Gateway
+from integration._support.provider import SharedProvider
+from integration.translation.case import TranslationTestCase
+from integration.translation.messages.bases.azure import GPT_6_LUNA_TEST_CASE as AZURE_GPT_6_LUNA_TEST_CASE
+from integration.translation.messages.bases.bedrock_converse import CLAUDE_SONNET_4_6_TEST_CASE
+from integration.translation.messages.bases.openai import GPT_5_4_TEST_CASE as OPENAI_GPT_5_4_TEST_CASE
+from integration.translation.runner import assert_translation
+from pydantic import JsonValue
+
+LOOKUP_SCHEMA: Final[dict[str, JsonValue]] = {
+    "type": "object",
+    "properties": {"city": {"type": "string"}},
+    "required": ["city"],
+}
+ANTHROPIC_TOOLS: Final[list[JsonValue]] = [
+    {"name": "lookup_weather", "description": "Look up the weather for a city.", "input_schema": LOOKUP_SCHEMA}
+]
+RESPONSES_TOOLS: Final[list[JsonValue]] = [
+    {
+        "type": "function",
+        "name": "lookup_weather",
+        "strict": False,
+        "description": "Look up the weather for a city.",
+        "parameters": LOOKUP_SCHEMA,
+    }
+]
+BEDROCK_TOOL_CONFIG: Final[dict[str, JsonValue]] = {
+    "tools": [
+        {
+            "toolSpec": {
+                "name": "lookup_weather",
+                "description": "Look up the weather for a city.",
+                "inputSchema": {"json": LOOKUP_SCHEMA},
+            }
+        }
+    ],
+    "toolChoice": {"auto": {}},
+}
+DISABLE_PARALLEL_TOOL_USE: Final[dict[str, JsonValue]] = {"type": "auto", "disable_parallel_tool_use": True}
+SCENARIO: Final = "tool_choice_disable_parallel_tool_use"
+
+OPENAI_GPT_5_4_DISABLE_PARALLEL_TEST_CASE: Final = replace(
+    OPENAI_GPT_5_4_TEST_CASE,
+    scenario=SCENARIO,
+    litellm_request={
+        **OPENAI_GPT_5_4_TEST_CASE.litellm_request,
+        "tools": ANTHROPIC_TOOLS,
+        "tool_choice": DISABLE_PARALLEL_TOOL_USE,
+    },
+    expected_provider_request={
+        **OPENAI_GPT_5_4_TEST_CASE.expected_provider_request,
+        "tools": RESPONSES_TOOLS,
+        "tool_choice": "auto",
+        "parallel_tool_calls": False,
+    },
+)
+
+AZURE_GPT_6_LUNA_DISABLE_PARALLEL_TEST_CASE: Final = replace(
+    AZURE_GPT_6_LUNA_TEST_CASE,
+    scenario=SCENARIO,
+    litellm_request={**AZURE_GPT_6_LUNA_TEST_CASE.litellm_request, "tool_choice": DISABLE_PARALLEL_TOOL_USE},
+    expected_provider_request={
+        **AZURE_GPT_6_LUNA_TEST_CASE.expected_provider_request,
+        "tool_choice": "auto",
+        "parallel_tool_calls": False,
+    },
+)
+
+CLAUDE_SONNET_4_6_DISABLE_PARALLEL_TEST_CASE: Final = replace(
+    CLAUDE_SONNET_4_6_TEST_CASE,
+    scenario=SCENARIO,
+    litellm_request={
+        **CLAUDE_SONNET_4_6_TEST_CASE.litellm_request,
+        "tools": ANTHROPIC_TOOLS,
+        "tool_choice": DISABLE_PARALLEL_TOOL_USE,
+    },
+    expected_provider_request={
+        **CLAUDE_SONNET_4_6_TEST_CASE.expected_provider_request,
+        "toolConfig": BEDROCK_TOOL_CONFIG,
+        "additionalModelRequestFields": {"tool_choice": {"disable_parallel_tool_use": True}},
+    },
+)
+
+CASES: Final = (
+    OPENAI_GPT_5_4_DISABLE_PARALLEL_TEST_CASE,
+    AZURE_GPT_6_LUNA_DISABLE_PARALLEL_TEST_CASE,
+    CLAUDE_SONNET_4_6_DISABLE_PARALLEL_TEST_CASE,
+)
+
+
+@pytest.mark.parametrize("case", CASES, ids=lambda case: case.id)
+def test_messages_params_tool_choice(case: TranslationTestCase, gateway: Gateway, provider: SharedProvider) -> None:
+    assert_translation(case, gateway, provider)

tests/unit/llms/anthropic/messages/adapters/test_anthropic_messages_adapters_transformation.py

@@ -5578,3 +5578,47 @@ async def test_chat_bridge_tool_result_text_block_without_text_reaches_the_model
             ],
         }
     ]
+
+
+@pytest.mark.parametrize(
+    ("tool_choice", "model", "custom_llm_provider", "expected_parallel_tool_calls"),
+    [
+        ({"type": "auto", "disable_parallel_tool_use": True}, "gpt-5.4", "openai", False),
+        ({"type": "any", "disable_parallel_tool_use": True}, "gpt-5.4", "openai", False),
+        ({"type": "auto", "disable_parallel_tool_use": True}, "openai/gpt-oss-120b", "groq", False),
+        ({"type": "auto", "disable_parallel_tool_use": False}, "gpt-5.4", "openai", None),
+        ({"type": "auto"}, "gpt-5.4", "openai", None),
+        ({"type": "auto", "disable_parallel_tool_use": True}, "command-a-03-2025", "cohere_chat", None),
+        ({"type": "auto", "disable_parallel_tool_use": True}, "gpt-5.4", None, None),
+    ],
+    ids=[
+        "openai-auto-disabled",
+        "openai-any-disabled",
+        "groq-auto-disabled",
+        "openai-auto-enabled",
+        "openai-auto-omitted",
+        "cohere-declares-no-parallel-tool-calls-stays-behind",
+        "no-provider-stays-behind",
+    ],
+)
+def test_translate_anthropic_to_openai_disable_parallel_tool_use_sets_parallel_tool_calls(
+    tool_choice: dict[str, object],
+    model: str,
+    custom_llm_provider: str | None,
+    expected_parallel_tool_calls: bool | None,
+) -> None:
+    from litellm.types.llms.anthropic import AnthropicMessagesRequest
+
+    request: Final = AnthropicMessagesRequest(
+        model=model,
+        max_tokens=64,
+        messages=[{"role": "user", "content": "What is the weather in Paris?"}],
+        tools=[{"name": "lookup_weather", "input_schema": {"type": "object", "properties": {}}}],
+        tool_choice=tool_choice,
+    )
+
+    translated, _ = LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai(
+        request, custom_llm_provider=custom_llm_provider
+    )
+
+    assert translated.get("parallel_tool_calls") == expected_parallel_tool_calls

tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_handler.py

@@ -52,6 +52,30 @@ def test_build_responses_kwargs_never_asks_the_nested_call_to_replay_a_stream():
     assert responses_kwargs["custom_llm_provider"] == "azure"
 
 
+@pytest.mark.parametrize(
+    ("model", "custom_llm_provider", "expected_parallel_tool_calls"),
+    [
+        ("openai/gpt-5.6-luna", "openai", False),
+        ("azure/gpt-5.6-luna", "azure", False),
+        ("perplexity/sonar", "perplexity", None),
+    ],
+    ids=["openai-declares", "azure-declares", "perplexity-declares-no-parallel-tool-calls-stays-behind"],
+)
+def test_build_responses_kwargs_maps_disable_parallel_tool_use_only_where_the_provider_declares_it(
+    model: str, custom_llm_provider: str, expected_parallel_tool_calls: bool | None
+):
+    responses_kwargs = _build_responses_kwargs(
+        max_tokens=1024,
+        messages=MESSAGES,
+        model=model,
+        tools=[{"name": "do_thing", "input_schema": {"type": "object", "properties": {}}}],
+        tool_choice={"type": "auto", "disable_parallel_tool_use": True},
+        extra_kwargs={"custom_llm_provider": custom_llm_provider},
+    )
+    assert responses_kwargs.get("parallel_tool_calls") == expected_parallel_tool_calls
+    assert responses_kwargs["tool_choice"] == "auto"
+
+
 def test_build_responses_kwargs_derives_prompt_cache_key_from_claude_code_session_id():
     responses_kwargs = _build_responses_kwargs(
         max_tokens=1024,

tests/unit/llms/anthropic/messages/responses_adapters/test_responses_adapters_transformation.py

@@ -1203,6 +1203,30 @@ class TestTranslateRequestBroaderCoverage:
         kwargs = _ADAPTER.translate_request(req)
         assert kwargs["tool_choice"] == {"type": "function", "name": "do_thing"}
 
+    @pytest.mark.parametrize(
+        ("tool_choice", "expected_parallel_tool_calls"),
+        [
+            ({"type": "auto", "disable_parallel_tool_use": True}, False),
+            ({"type": "any", "disable_parallel_tool_use": True}, False),
+            ({"type": "auto", "disable_parallel_tool_use": False}, None),
+            ({"type": "auto"}, None),
+        ],
+    )
+    def test_disable_parallel_tool_use_sets_parallel_tool_calls(
+        self, tool_choice: dict[str, object], expected_parallel_tool_calls: bool | None
+    ):
+        req = _make_request(tools=[{"name": "do_thing"}], tool_choice=tool_choice)
+        kwargs = _ADAPTER.translate_request(req)
+        assert kwargs.get("parallel_tool_calls") == expected_parallel_tool_calls
+
+    def test_disable_parallel_tool_use_stays_behind_when_provider_lacks_parallel_tool_calls(self):
+        req = _make_request(
+            tools=[{"name": "do_thing"}], tool_choice={"type": "auto", "disable_parallel_tool_use": True}
+        )
+        kwargs = _ADAPTER.translate_request(req, declares_parallel_tool_calls=False)
+        assert "parallel_tool_calls" not in kwargs
+        assert kwargs["tool_choice"] == "auto"
+
     def test_thinking_translated_to_reasoning(self):
         req = _make_request(thinking={"type": "enabled", "budget_tokens": 12000})
         kwargs = _ADAPTER.translate_request(req)