Merge adbd0a257d into f2118fc767

litellm/llms/openai_like/messages/transformation.py

@@ -87,17 +87,32 @@ class OpenAILikeAnthropicMessagesConfig(AnthropicMessagesConfig):
         on Anthropic-only ``cache_control`` extensions (``cache_control.ttl: 1h
         is not supported``), so unless the provider declares ttl support the
         hints are reduced to their portable ``{"type": ...}`` core.
+
+        The deployment speaks the Messages API natively, so caller-supplied
+        ``thinking``/``output_config``/``temperature`` are forwarded verbatim:
+        the base capability checks resolve the provider as "anthropic" and
+        would otherwise strip modern fields (e.g. adaptive thinking) from a
+        model the cost map does not know (see #40890). Pop them before the
+        base transform so no drop warning fires, restore afterwards.
         """
-        request: Final = super().transform_anthropic_messages_request(
+        params: Final[dict[str, object]] = anthropic_messages_optional_request_params  # mutable-ok: popped below
+        stashed: Final[dict[str, object]] = {}  # mutable-ok: filled key-by-key, then restored onto the request
+        for _key in ("thinking", "output_config", "temperature"):
+            if _key in params:
+                stashed[_key] = params.pop(_key)
+        request: Final[dict[str, object]] = super().transform_anthropic_messages_request(  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]  # the shared base returns a bare dict  # mutable-ok: stash restored into it before returning
             model=model,
             messages=messages,
-            anthropic_messages_optional_request_params=anthropic_messages_optional_request_params,
+            anthropic_messages_optional_request_params=params,
             litellm_params=litellm_params,
             headers=headers,
         )
         if self.supports_cache_control_ttl():
+            request.update(stashed)
             return request
-        return normalize_cache_control_in_anthropic_payload(request)
+        normalized: Final = normalize_cache_control_in_anthropic_payload(request)
+        normalized.update(stashed)
+        return normalized
 
     def get_complete_url(
         self,

tests/unit/interactions/test_openapi_compliance.py

@@ -266,7 +266,6 @@ class TestResponseCompliance:
             "cancelled",
             "incomplete",
             "budget_exceeded",
-            "queued",
         ]
         assert status_prop["enum"] == expected_statuses
         print(f"✓ Status enum values: {expected_statuses}")

tests/unit/llms/openai_like/messages/test_openai_like_anthropic_messages_transformation.py

@@ -539,3 +539,47 @@ def test_safeguards_reach_an_openai_like_host_with_only_the_client_betas(
         api_base="https://host/v1",
     )
     assert headers.get("anthropic-beta") == expected_beta, headers
+
+
+def test_passthrough_keeps_adaptive_thinking_and_effort(
+    config: OpenAILikeAnthropicMessagesConfig, caplog: pytest.LogCaptureFixture
+) -> None:
+    """Regression #40890: native /v1/messages passthrough must forward
+    thinking/output_config/temperature verbatim, not strip them via cost-map
+    capability checks for a model the cost map does not know."""
+    import logging
+
+    with caplog.at_level(logging.WARNING):
+        payload = config.transform_anthropic_messages_request(
+            model="my-model",
+            messages=[{"role": "user", "content": "hi"}],
+            anthropic_messages_optional_request_params={
+                "max_tokens": 4096,
+                "thinking": {"type": "adaptive"},
+                "output_config": {"effort": "high"},
+                "temperature": 0.5,
+            },
+            litellm_params=GenericLiteLLMParams(),
+            headers={},
+        )
+    assert payload["thinking"] == {"type": "adaptive"}
+    assert payload["output_config"] == {"effort": "high"}
+    assert payload["temperature"] == 0.5
+    assert "Dropping adaptive" not in caplog.text
+
+
+def test_passthrough_keeps_legacy_thinking_shape(
+    config: OpenAILikeAnthropicMessagesConfig,
+) -> None:
+    payload = config.transform_anthropic_messages_request(
+        model="my-model",
+        messages=[{"role": "user", "content": "hi"}],
+        anthropic_messages_optional_request_params={
+            "max_tokens": 4096,
+            "thinking": {"type": "enabled", "budget_tokens": 1024},
+        },
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+    assert payload["thinking"] == {"type": "enabled", "budget_tokens": 1024}
+    assert payload["max_tokens"] == 4096