For instance in a lot of cases what you really want is to convert a grammar to some reasonable abstract syntax tree and I've found often you have to write a lot of code into those frameworks to do that, whereas you ought to be able to write something declarative that generates the AST objects and the translation code.

Another problem is a lack of composability. Particularly in DSL creation it would be desirable to "cut and paste" grammars into each other; it's very possible to embed SQL into a language like C, Java, or such -- you ought to be able to embed one parser in another but generally it is not easy.

Also there is the problem of error message generation, which is bad enough when you do are doing things that are "easy with yacc" but when you are doing hardcore transpilation of some kind, the error messages are often completely incomprehensible. This is a big factor in any kind of "rules engine" built around an existing language; error handling is disastrously bad in Drools, for instance -- you might think homoiconicity would improve matters with Clara in Clojure but it really doesn't.