A more definitely safe way to reduce, but not eliminate, the monadic junk in never:

never =  fix $ Step . pure . Step . pure . Step . pure . Step . pure . ... . Effect . return

I'm just a bit worried about building a long chain of pointers to follow, and also about what a CPU's branch predictor will think about this sort of stop-start thing. It might be worth benchmarking.

BTW, I'm not at all sure this is actually a good thing to do. never likely isn't a performance-critical function, and this implementation is utterly wild and unsupported. I mostly wanted to see if it was even possible. But if you want it, here it is.

I can't say I'm particularly in support of this. Seems like clever for clever's sake. If it could at least be verified with a benchmark, I'd be more inclined.

@ocharles I think you're right. But it would be worth thinking about whether the Step constructor should really be lazy anyway. I'm always a bit uncomfortable about strict polymorphic fields. Stream ((->) a) forces the values of functions, which perversely distinguishes _|_ from const _|_ and plays unpredictably with GHC's eta expansion. Now that type (the type of stream sinks) doesn't really seem too important in this package, but I still think it's worth considering. It would always be possible to offer strict operations even with a lazy constructor, although getting it right requires some care and there might be (small?) performance regressions.