fix: don't send thinking params to official OpenAI API (#115)

The OpenAI-compatible provider was sending `thinking`, `enable_thinking`,
and `chat_template_kwargs` parameters to all endpoints including the
official OpenAI API, which rejects unrecognized arguments with a 400 error.
These params are now only sent to local/compatible servers (LM Studio,
llama.cpp, vLLM, etc.).

Closes #115

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: hydropix

src/core/llm/providers/openai.py

@@ -34,6 +34,10 @@ def __init__(self, api_endpoint: str, model: str, api_key: Optional[str] = None,
         self._detected_context_size: Optional[int] = None
         self._context_detector = ContextDetector()
 
+    def _is_official_openai_endpoint(self) -> bool:
+        """Check if the endpoint is the official OpenAI API."""
+        return "api.openai.com" in self.api_endpoint
+
     @staticmethod
     def _normalize_endpoint(endpoint: str) -> str:
         """
@@ -94,16 +98,14 @@ async def generate(self, prompt: str, timeout: int = REQUEST_TIMEOUT,
             "model": self.model,
             "messages": messages,
             "stream": False,
-            # Disable thinking/reasoning mode for local servers and compatible APIs
-            # This prevents models from outputting <think>...</think> blocks
-            "thinking": False,
-            "enable_thinking": False,
-            # Some servers use chat_template_kwargs
-            "chat_template_kwargs": {
-                "enable_thinking": False
-            }
         }
 
+        # Only add thinking-disable params for local/compatible servers, not official OpenAI API
+        if not self._is_official_openai_endpoint():
+            payload["thinking"] = False
+            payload["enable_thinking"] = False
+            payload["chat_template_kwargs"] = {"enable_thinking": False}
+
         client = await self._get_client()
         for attempt in range(MAX_TRANSLATION_ATTEMPTS):
             try: