This is something we’ve seen reported, and I’ve flagged it to our team. Resuming a subagent is meant to keep its original model and reasoning effort, so the effort silently changing (and costing more) is a bug on our side, not anything in your setup.
A workaround that should hold it steady in the meantime: open the main model picker, select the affected model, and pin the reasoning effort you want there (and turn Fast off if you don’t want the fast variant). Resumes fall back to that saved preference, so they should stay on the effort you picked instead of resetting to the model’s default.
On your second point: resume is designed to reuse the subagent’s original model, so once the effort drop is fixed it should carry over automatically without needing an explicit parameter. I’ve passed the override idea along as well.