Organize test infrastructure better. Add more tests. It's at the point that it's probably safe to write them in its own language.
Make using Boehm's GC instead an option.
Avoid trying to throw in the kitchen sink.
Implement more primitives. At least get around R4RS-ish. SLIB compatibility would be wonderful.
I can understand how tailcalls work after (another) 6 months but probably only because I wrote it. It would be nice if everything trampoline related was contained within eval(). At least document that tc() is used for tailcalls, make it external, and put all the specials that need it into their own file.