2.5 flash lite support and bug fixes

app/api_helpers.py

@@ -130,7 +130,7 @@ def create_generation_config(request: OpenAIRequest) -> Dict[str, Any]:
             types.SafetySetting(category="HARM_CATEGORY_HARASSMENT", threshold="OFF"),
             types.SafetySetting(category="HARM_CATEGORY_CIVIC_INTEGRITY", threshold="OFF")
     ]
-    config["thinking_config"] = {"include_thoughts": True}
+    # config["thinking_config"] = {"include_thoughts": True}
 
     # 1. Add tools (function declarations)
     function_declarations = []

app/routes/chat_api.py

@@ -97,6 +97,11 @@ async def chat_completions(fastapi_request: Request, request: OpenAIRequest, api
         # This will now be a dictionary
         gen_config_dict = create_generation_config(request)
 
+        if "gemini-2.5-flash" in base_model_name or "gemini-2.5-pro" in base_model_name:
+            if "thinking_config" not in gen_config_dict:
+                gen_config_dict["thinking_config"] = {}
+            gen_config_dict["thinking_config"]["include_thoughts"] = True
+
         if "gemini-2.5-flash-lite" in base_model_name:
             gen_config_dict["thinking_config"]["include_thoughts"] = False