I (well, Alex was a bit quicker, actually) propose a binding form,
similar to Shiver's let-optionals, that gives you everything needed to
provide keyword-style interfaces to a procedure while not limiting them
to one particular behaviour:

[syntax] (let-keywords* <ARGLIST> ((<VARIABLE> <KEYWORD> [<DEFAULT>])
             ...) BODY ...)

The semantics should be obvious, for example:

(define (message-box title . args)
  (let-keywords* args ((title 'title "<untitled>") (buttons 'buttons '(ok)))
    ...))

Default values are optional and default to #f (or something else). One
could also imagine an optional trailing "rest" argument after the last
binding to keep remaining arguments, but I think that would have to
large an impact of performance, so I personally tend to leave
much unspecified.

Notes:
- the <KEYWORD> expressions are evaluated and don't necessarily
  have to be keyword objects
- the semantics are not as confusing as with DSSSL lambda-lists
  (#!key before or after #!rest ?, in what context are which variables
  visible?, etc.)
- You have more control over the actual argument processing
  (how do you implement
  (message-box "Hello!" button: "ok" button: "cancel")
  with DSSSL keyword args? Seperating out the binding form gives
  the user the chance to implement custom behaviour)
- This can be implemented on every R5RS Scheme
- You don't muck with lambda-list syntax, which I think is the
  major problem with this SRFI (and writing definition functions for
  something like generic functions, or implementing alternative
  macro systems are just a few examples where you really
  appreciate this)
- You can still provide keyword-based interfaces to the outside,
  so you get the functionality of DSSSL keywords without the
  warts (funny lambda list markers, fixed and complicated behaviour,
  unnecessary coupling to keyword objects).

cheers,
felix