Hi Yan,
I'd recommend you to try to start out with some messaging pattern instead of reaching for special mailboxes.
Did you try just catching the exception and retrying your logic rather than relying on supervision for this?
A second option would be to have the sender retry sending the message after a timeout. A third option would
be to provide a parent actor that outside actors interact with who contains the retry logic (which messages are in
flight and what child is processing what message, how many retries for a specific message) and lets the children
just fail without caring about the retry logic.