change

backend/app/cad_agent/adapters/structured_llm.py

@@ -46,13 +46,67 @@ class StructuredModelGateway:
 
     async def call_tool(self, *, messages: list[dict[str, Any]], tool: dict[str, Any], provider_id: str, model_id: str, required_tool_name: str) -> dict[str, Any]:
         provider, model = self._provider_model(provider_id, model_id)
-        payload = self._payload(provider, model.id, messages, [tool], required_tool_name)
-        response = await self._request(provider, payload)
+        response, compatibility_mode = await self._request_tool_call(
+            provider, model.id, messages, [tool], required_tool_name,
+        )
         try:
-            return self._normalized_response(provider, response)
+            normalized = self._normalized_response(provider, response)
+            normalized["usage"]["structured_compatibility_mode"] = compatibility_mode
+            return normalized
         except (KeyError, TypeError, ValueError) as error:
             raise StructuredModelError("Provider response cannot be normalized as a structured tool-call result") from error
 
+    async def _request_tool_call(
+        self,
+        provider: ProviderConfig,
+        model_id: str,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        required_tool_name: str,
+    ) -> tuple[dict[str, Any], str]:
+        """Keep forced tool choice unless a provider explicitly rejects thinking mode.
+
+        Some OpenAI-compatible gateways reject an otherwise valid forced
+        ``tool_choice`` whenever reasoning is enabled or implicit. This is a
+        request compatibility issue rather than an author error. First retry
+        without the client-requested reasoning option; only if that exact
+        rejection remains do we allow automatic choice among the single tool
+        already exposed to the model. The workflow's one-call/schema checks
+        still reject free-form or wrong-tool responses.
+        """
+        try:
+            return await self._request(
+                provider,
+                self._payload(provider, model_id, messages, tools, required_tool_name),
+            ), "forced"
+        except StructuredModelError as first_error:
+            if not self._thinking_tool_choice_rejection(first_error):
+                raise
+        try:
+            return await self._request(
+                provider,
+                self._payload(
+                    provider, model_id, messages, tools, required_tool_name,
+                    include_reasoning=False,
+                ),
+            ), "reasoning_disabled"
+        except StructuredModelError as second_error:
+            if not self._thinking_tool_choice_rejection(second_error):
+                raise
+        return await self._request(
+            provider,
+            self._payload(
+                provider, model_id, messages, tools, required_tool_name,
+                include_reasoning=False,
+                force_tool_name=False,
+            ),
+        ), "single_tool_auto"
+
+    @staticmethod
+    def _thinking_tool_choice_rejection(error: StructuredModelError) -> bool:
+        message = str(error).casefold()
+        return "thinking" in message and ("tool_choice" in message or "tool choice" in message)
+
     async def conformance(self, *, provider_id: str, model_id: str, tools: list[dict[str, Any]]) -> dict[str, Any]:
         failures: list[dict[str, str]] = []
         usage = {"prompt_tokens": 0, "completion_tokens": 0}
@@ -190,20 +244,31 @@ class StructuredModelGateway:
             })
         return values
 
-    def _payload(self, provider: ProviderConfig, model_id: str, messages: list[dict[str, Any]], tools: list[dict[str, Any]], required_tool_name: str) -> dict[str, Any]:
+    def _payload(
+        self,
+        provider: ProviderConfig,
+        model_id: str,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        required_tool_name: str,
+        *,
+        include_reasoning: bool = True,
+        force_tool_name: bool = True,
+    ) -> dict[str, Any]:
         if len(tools) != 1 or not required_tool_name:
             raise StructuredModelError("The v3 protocol requires exactly one named tool per provider call.")
+        request_options = provider.request_options if include_reasoning else {}
         if provider.api_style == "responses":
             payload: dict[str, Any] = {
                 "model": model_id, "input": self._responses_input(messages), "tools": self._responses_tools(tools),
-                "tool_choice": {"type": "function", "name": required_tool_name},
-                **provider.request_options,
+                "tool_choice": {"type": "function", "name": required_tool_name} if force_tool_name else "auto",
+                **request_options,
             }
             return payload
         return {
             "model": model_id, "messages": messages, "tools": tools,
-            "tool_choice": {"type": "function", "function": {"name": required_tool_name}},
-            "temperature": 0, **provider.request_options,
+            "tool_choice": {"type": "function", "function": {"name": required_tool_name}} if force_tool_name else "auto",
+            "temperature": 0, **request_options,
         }
 
     @staticmethod