Merge 3734837163 into a392a39627

litellm/llms/anthropic/chat/transformation.py

@@ -417,6 +417,10 @@ DROP_SERVICE_TIER_PROVIDER_WARNING: Final = (
     "Dropping `service_tier` for model=%s on provider=%s. Only the Anthropic API routes accept it."
 )
 
+DROP_THREAD_PROVIDER_WARNING: Final = (
+    "Dropping `thread` for model=%s on provider=%s. Only the Anthropic API keeps Message Threads."
+)
+
 DROP_UNSUPPORTED_TOOL_SEARCH_WARNING: Final = (
     "Dropping the OpenAI `tool_search` tool and the `defer_loading` of the other tools for model=%s on %s "
     "(drop_params=True) because this route cannot run that tool search."
@@ -698,6 +702,17 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
         litellm.verbose_logger.warning(DROP_SERVICE_TIER_PROVIDER_WARNING, model, custom_llm_provider)
         optional_params.pop("service_tier", None)
 
+    @staticmethod
+    def maybe_drop_thread_param(
+        model: str,
+        optional_params: dict[str, object],  # mutable-ok: mirrors override contract
+        custom_llm_provider: str | None = None,
+    ) -> None:
+        if "thread" not in optional_params or custom_llm_provider in (None, "anthropic"):
+            return
+        litellm.verbose_logger.warning(DROP_THREAD_PROVIDER_WARNING, model, custom_llm_provider)
+        optional_params.pop("thread", None)
+
     @staticmethod
     def _raise_invalid_reasoning_effort(model: str, value: object, llm_provider: str) -> NoReturn:
         """Raise a ``BadRequestError`` for an unrecognised ``reasoning_effort``.

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

@@ -288,7 +288,7 @@ class AnthropicAdapter:
         #########################################################
         # Created Typed Request Body
         #########################################################
-        request_body: Final = AnthropicMessagesRequest(model=model, messages=messages, **kwargs)
+        request_body: Final = AnthropicMessagesRequest(model=model, messages=messages, **kwargs)  # pyright: ignore[reportUnknownArgumentType]  # the request params arrive as a bare dict
 
         (
             translated_body,

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

@@ -793,7 +793,7 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
             ),
             max_tokens=max_tokens,
             model=model,
-            **anthropic_messages_optional_request_params,
+            **anthropic_messages_optional_request_params,  # pyright: ignore[reportUnknownArgumentType]  # the request params arrive as a bare dict
         )
         request: Final = (
             {**dict(anthropic_messages_request), "thinking": gated.thinking}

litellm/llms/anthropic/messages/native/utils.py

@@ -301,6 +301,11 @@ class AnthropicMessagesRequestUtils:
                 optional_params=filtered_params,
                 custom_llm_provider=custom_llm_provider,
             )
+            AnthropicConfig.maybe_drop_thread_param(
+                model=model,
+                optional_params=filtered_params,
+                custom_llm_provider=custom_llm_provider,
+            )
             for param in ("temperature", "top_p", "top_k"):
                 if param in filtered_params:
                     AnthropicModelInfo._apply_sampling_param(  # pyright: ignore[reportPrivateUsage]  # same gating the /chat/completions path applies; forking it would drift

litellm/types/llms/anthropic.py

@@ -440,6 +440,7 @@ class AnthropicMessagesRequestOptionalParams(TypedDict, total=False):
     reasoning_effort: str | None
     safeguards: ReadOnly[list[dict[str, object]] | None]
     service_tier: ReadOnly[str | None]
+    thread: ReadOnly[Mapping[str, object] | None]
 
 
 class AnthropicMessagesRequest(AnthropicMessagesRequestOptionalParams, total=False):

tests/unit/llms/anthropic/messages/native/test_anthropic_messages_handler.py

@@ -1633,6 +1633,66 @@ async def test_anthropic_messages_forwards_safeguards_and_unknown_beta_to_anthro
     assert response["safeguard_results"] == safeguard_results
 
 
+@pytest.mark.asyncio
+async def test_anthropic_messages_forwards_message_threads_continue_to_anthropic():
+    """Shape is what Claude Code 2.1.312 sends with the message-threads-2026-08-12 beta, captured 2026-10-29."""
+    from litellm.llms.anthropic.messages.native import handler
+
+    thread = {"type": "continue", "previous_message_id": "msg_01ABC"}
+    tool_result_only = [
+        {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "toolu_01", "content": "README contents"}]}
+    ]
+    captured: dict[str, object] = {}
+
+    def upstream_records_the_request(request: httpx.Request) -> httpx.Response:
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_02DEF",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-haiku-4-5",
+                "content": [{"type": "text", "text": "ok"}],
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+            request=request,
+        )
+
+    upstream = AsyncHTTPHandler()
+    upstream.client = httpx.AsyncClient(transport=httpx.MockTransport(upstream_records_the_request))
+
+    await handler.anthropic_messages(
+        max_tokens=16,
+        messages=tool_result_only,
+        model="anthropic/claude-haiku-4-5",
+        custom_llm_provider="anthropic",
+        api_key="sk-test",
+        client=upstream,
+        thread=thread,
+        extra_headers={"anthropic-beta": "message-threads-2026-08-12"},
+    )
+
+    assert captured["body"]["thread"] == thread
+    assert captured["body"]["messages"] == tool_result_only
+
+
+@pytest.mark.parametrize("custom_llm_provider", ["vertex_ai", "bedrock", "azure_ai", "databricks"])
+def test_anthropic_messages_drops_thread_where_message_threads_are_not_served(custom_llm_provider: str) -> None:
+    from litellm.llms.anthropic.messages.native.utils import AnthropicMessagesRequestUtils
+
+    optional_params = AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
+        params={"max_tokens": 64, "thread": {"type": "create"}},
+        model="claude-haiku-4-5",
+        custom_llm_provider=custom_llm_provider,
+    )
+
+    assert "thread" not in optional_params
+    assert dict(optional_params).get("max_tokens") == 64
+
+
 @pytest.mark.asyncio
 async def test_anthropic_messages_streaming_forwards_safeguards_and_keeps_safeguard_results():
     """Shapes are what Claude Code 2.1.278 sends and api.anthropic.com returns, captured 2026-09-21."""