finally make token counting work in apps

Author: RayBytes

chatmock/routes_openai.py

@@ -87,7 +87,6 @@ def chat_completions() -> Response:
         tool_choice=tool_choice,
         parallel_tool_calls=parallel_tool_calls,
         reasoning_param=reasoning_param,
-        include_usage=(not is_stream) or include_usage,
     )
     if error_resp is not None:
         return error_resp
@@ -256,7 +255,6 @@ def completions() -> Response:
         input_items,
         instructions=BASE_INSTRUCTIONS,
         reasoning_param=reasoning_param,
-        include_usage=(not stream_req) or include_usage,
     )
     if error_resp is not None:
         return error_resp

chatmock/upstream.py

@@ -40,7 +40,6 @@ def start_upstream_request(
     tool_choice: Any | None = None,
     parallel_tool_calls: bool = False,
     reasoning_param: Dict[str, Any] | None = None,
-    include_usage: bool | None = None,
 ):
     access_token, account_id = get_effective_chatgpt_auth()
     if not access_token or not account_id: