Just dropping in to say strong agree.
The biggest stumbling block for me has been that the effect handlers are captured as part of the continuation; I was expecting them to need to be manually re-established every time the continuation is resumed. The way it works makes sense when you always `continue` within the body of an effect handler, but it feel a little unfortunate that there's no way to change the handlers for future resumptions.
Other than that, the possibilities for userland coroutines, dynamically-scoped variables, and centralization of state (like `ref` but provided by a scoped allocator) are incredibly exciting. I've wanted a system like this for a long time, and for some use-cases (scheduling of concurrent simulation tasks) I had to abuse threads and thread-locals to get the separation I wanted.