I know it’s quite generic, but check out Divergent transitions - a primer for some general strategies.
I think the part most immediately relevant is to simplify the model to find a) the most complex model you can reliably fit (preferably without altering tree depth or adapt_delta) and b) the least complex model that has issues. Similarly if the issues manifest also on a reduced dataset it is easier to debug with it. This can simplify debugging a lot.