I'm also curious about this choice. I found it surprising when I first examined the code generated by this library, since it doesn't seem like idiomatic Python to me, so I'm wondering what the reasoning behind it was.

A particular consideration for my use case: it's not certain that consumers of our generated code will be effectively enforcing type-checking (some will be, some won't be is my guess). In the case where they aren't, then using a union of [ok, error] rather than an exception that's raised on error increases the chances that they'll get some inscrutable downstream error because they thought they had a success response but actually had a failure one, and forgot to check which it was.