This is related to the way Azure DataBricks mixes magic commands and python code. When notebook (from Azure DataBricks UI) is split into separate parts, one containing only magic commands %sh pwd and others only python code, committed file is not messed up.