[racket] REPL input chunking for non-sexp language

I'm working (with sk and jpolitz) on a non-sexp language built on top
of racket.

We have basic support for it in the repl inside DrRacket, but none at
all from the racket commandline repl (which also means no support for
embedding inside other editors) - and the former seems to be using
s-exps to figure out when to send the input to our eval (I think - I
haven't found documentation describing how this works).

So my question is:

Is there a way to specify how input is split before sending to eval,
both so that DrRacket could follow our conventions (which might, for
example, match a "fun" with matching "end"), and so that the
commandline repl knows how to split input at all (as right now it just
keeps waiting for input until EOF - and EOF also causes the repl to
quit!) Is there something that read/read-syntax should signal, or another
handler to use?

Thanks!
Daniel