The problem with the current walkdir implementation is that without a rich path object you end up making a lot of redundant system calls. Combined with the inherent overhead of nested generators, it just *feels* bad (even in situations where worrying about the speed is truly a case of premature optimisation)

The caching in pathlib should deal with the problem of redundant system calls, so rebasing walkdir on top of that would probably be a good idea (i.e. all walkdir APIs would produce pathlib paths, and implicit convert strings they encounter to paths).