Judah Jacobson wrote:

I've already demonstrated how a library writer can solve that problem in: http://www.haskell.org/pipermail/libraries/2008-January/009034.html

Yes, that is indeed very nice. But I don't think it is a practical solution right now, for the following reasons: It requires that users be in a special monad; it doesn't work for regular MonadIO. So it isn't really a practical solution until it gets into MTL and people start porting their code. Providing an instance for MonadIO1 (or whatever we would call it) is an extra step for monad writers that may often be skipped. Also, there are obviously many other more complex cases beyond a simple bracket. How general is this technique? As you point out, even for bracket it does not work unless the "after" action is IO. What other restrictions are there in the general case? Perhaps we could fix those problems by also providing liftIOk for all kinds k that occur in the signatures of various functions in Control.Exception. But that would make the new monad even more unwieldy. As I suggested before, I think the best solution would be to define the functions of Control.Exception in terms of lower-level primitives that could then be used to define the functions directly for other monads. Perhaps something like this would be general enough to cover everything: -- Delay any exception and put in buffer startDelayingExceptions :: IO () -- Stop delaying exceptions stopDelayingExceptions :: IO () -- Retrieve delayed exception and clear buffer getDelayedException :: IO (Maybe Exception) There would be one exception buffer per thread. New exceptions would replace old ones, so users of this mechanism would be required to catch as catch can. Based on Simon Marlow's observation in our previous discussion on this topic, we would still want to leave around the current primitives as a special case optimization for IO. Regards, Yitz