Basically, the server "knows" (as in can know) the references that exist on the client and vice versa. When a watched reference changes value on the client an HTTP request/response is sent over the wire. The request carries the old and new value, the response is basically redundant but since it's HTTP it needs to be there. Hence, the response might carry new forms to be executed on the client.

The client watching references on the server is obviously more tricky. node.js already supports keep-alive by default. There probably needs to be some kind of chuncked reply mechanism that allows the server to send changes without the client having to send polling requests. Obviously, if the reference is not going to change frequently it might be cheaper to just poll every once in a while.

on-server reminds me of Hop. Take a look at http://hop.inria.fr/, maybe http://hop.inria.fr/usr/local/lib/hop/2.1.1/weblets/home/art... in particular.