I think the OG github copilot version felt magical in that time. But I've never tried working like that with current frontier models. Considering how they train it now it might never work without changes how model respond.