See previous work in T98063. I'm not sure what to do here, we should start by having a conversation. My instinct so far is that all of the failure stuff should be handled by a calling function which has the responsibility of catching ResponseProcessingException, finalizing the transaction and so on. For workflows that involve another call to do_transaction, I would like to see that part of the process moved upwards as well, but don't know what that will look like yet.