Feat: use modelDetails.shouldSkipTemperature from discovery to gate sampling params

Add _should_skip_sampling_params cached_property to UiPathBaseLLMClient.
It reads shouldSkipTemperature from get_model_info() (authoritative) and
falls back to the is_claude_opus_4_or_above() name heuristic when the
model is not found in discovery.

All five Anthropic clients now use self._should_skip_sampling_params
instead of calling is_claude_opus_4_or_above directly, so the logic
automatically adapts when the backend sets shouldSkipTemperature on
new models without requiring a code change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: cosminacho

packages/uipath_langchain_client/src/uipath_langchain_client/base_client.py

@@ -49,6 +49,7 @@
     get_captured_response_headers,
     set_captured_response_headers,
 )
+from uipath.llm_client.utils.model_family import is_claude_opus_4_or_above
 from uipath_langchain_client.settings import (
     UiPathAPIConfig,
     UiPathBaseSettings,
@@ -190,6 +191,19 @@ def uipath_async_client(self) -> UiPathHttpxAsyncClient:
             logger=self.logger,
         )
 
+    @cached_property
+    def _should_skip_sampling_params(self) -> bool:
+        """True if the model's discovery metadata marks temperature/top_k/top_p as unsupported.
+
+        Reads ``modelDetails.shouldSkipTemperature`` from the model discovery API.
+        Falls back to a name-based heuristic when the model is not found in discovery.
+        """
+        try:
+            info = self.client_settings.get_model_info(model_name=self.model_name)
+            return bool(info.get("modelDetails", {}).get("shouldSkipTemperature", False))
+        except Exception:
+            return is_claude_opus_4_or_above(self.model_name)
+
     def uipath_request(
         self,
         method: Literal["POST", "GET"] = "POST",

packages/uipath_langchain_client/src/uipath_langchain_client/clients/anthropic/chat_models.py

@@ -15,7 +15,6 @@
 )
 from uipath_langchain_client.utils import (
     CLAUDE_OPUS_4_UNSUPPORTED_SAMPLING_PARAMS,
-    is_claude_opus_4_or_above,
 )
 
 try:
@@ -165,7 +164,7 @@ def _get_request_payload(
         **kwargs: Any,
     ) -> dict:
         payload = super()._get_request_payload(input_, stop=stop, **kwargs)
-        if is_claude_opus_4_or_above(self.model):
+        if self._should_skip_sampling_params:
             for param in CLAUDE_OPUS_4_UNSUPPORTED_SAMPLING_PARAMS:
                 payload.pop(param, None)
         return payload

packages/uipath_langchain_client/src/uipath_langchain_client/clients/bedrock/chat_models.py

@@ -15,7 +15,6 @@
 )
 from uipath_langchain_client.utils import (
     CLAUDE_OPUS_4_UNSUPPORTED_SAMPLING_PARAMS,
-    is_claude_opus_4_or_above,
 )
 
 try:
@@ -85,7 +84,7 @@ def setup_uipath_client(self) -> Self:
     @override
     def _converse_params(self, **kwargs: Any) -> dict:
         params = super()._converse_params(**kwargs)
-        if is_claude_opus_4_or_above(self.model_id):
+        if self._should_skip_sampling_params:
             inference = params.get("inferenceConfig")
             if isinstance(inference, dict):
                 inference.pop("temperature", None)
@@ -119,7 +118,7 @@ def setup_model_id(cls, values: Any) -> Any:
     @model_validator(mode="after")
     def setup_uipath_client(self) -> Self:
         self.client = WrappedBotoClient(self.uipath_sync_client)
-        if is_claude_opus_4_or_above(self.model_id):
+        if self._should_skip_sampling_params:
             self.temperature = None
             if self.model_kwargs:
                 self.model_kwargs = {
@@ -181,7 +180,7 @@ def _get_request_payload(
         **kwargs: Any,
     ) -> dict:
         payload = super()._get_request_payload(input_, stop=stop, **kwargs)
-        if is_claude_opus_4_or_above(self.model):
+        if self._should_skip_sampling_params:
             for param in CLAUDE_OPUS_4_UNSUPPORTED_SAMPLING_PARAMS:
                 payload.pop(param, None)
         return payload

packages/uipath_langchain_client/src/uipath_langchain_client/clients/normalized/chat_models.py

@@ -69,7 +69,6 @@
 from uipath_langchain_client.utils import (
     CLAUDE_OPUS_4_UNSUPPORTED_SAMPLING_PARAMS,
     is_anthropic_model_name,
-    is_claude_opus_4_or_above,
 )
 
 _DictOrPydanticClass = Union[dict[str, Any], type[BaseModel], type]
@@ -256,7 +255,7 @@ def _default_params(self) -> dict[str, Any]:
             params.pop("temperature", None)
 
         # Claude Opus 4+ reasoning models reject temperature, top_k and top_p entirely.
-        if is_claude_opus_4_or_above(self.model_name):
+        if self._should_skip_sampling_params:
             for param in CLAUDE_OPUS_4_UNSUPPORTED_SAMPLING_PARAMS:
                 params.pop(param, None)
 

tests/langchain/clients/bedrock/test_unit.py

@@ -73,6 +73,8 @@ def test_is_claude_opus_4_or_above(self) -> None:
     def test_unsupported_params_stripped_from_payload(
         self, opus4_client: UiPathChatAnthropicBedrock
     ) -> None:
+        # Force _should_skip_sampling_params without a real discovery call.
+        opus4_client.__dict__["_should_skip_sampling_params"] = True
         with patch.object(opus4_client, "_client") as mock_client:
             mock_client.messages.create.return_value = MagicMock(
                 content=[MagicMock(type="text", text="hi")],
@@ -94,6 +96,8 @@ def test_sampling_params_kept_for_other_models(
             settings=client_settings,
             temperature=0.5,
         )
+        # Discovery returns False for haiku; ensure params are not stripped.
+        haiku.__dict__["_should_skip_sampling_params"] = False
         with patch.object(haiku, "_client") as mock_client:
             mock_client.messages.create.return_value = MagicMock(
                 content=[MagicMock(type="text", text="hi")],