A fun problem I play with sometimes: eek out a small-but-super-powerful subset of standard shell tools (sh, awk). For example, if you could create a forth in a dozen lines of bash, then you could just recreate it whenever you needed it.

Then you could start refactoring code to it. Easier than a rewrite.

I got quite far along at building a message-driven paradigm in awk. The input is a stream of well-defined messages. As you read it, append new messages to the end of the stream. I found that the awk I was using (gnu I think) was complicated by IO buffering behaviour. Still, an interesting problem space.

Lex may offer strong opportunities too, although is not part of the posix base.