It is often specified in the various tutorials of MVVM, that the goal of MVVM is not to eliminate the code-behind and that some event-handling may still be necessary in the code-behind.
What are the scenarios in which you need to write events in code-behind rather than using commands in viewmodel?
In general, if your code pertains to UI logic, keep it in the view's XAML or code-behind. The view model is only responsible for bridging and binding data between the view and the model.
An example can be found in one of my questions, How do I make a WPF window movable by dragging the extended window frame? One of the events I use is SourceInitialized, in which I access the Window's window handle to perform some Windows API magic. But all this relates to the window, and has nothing to do with application logic beyond the window, so I restrict it all to the window's code-behind file, leaving the view model completely ignorant of it.