core: fix issue when switching models (mainly between providers) where past reasoning/metadata would be sent to server and cause 400 errors since they came from another account/provider