The programmability of a shell just isn't interesting to me. Whenever I start hacking a shell script, it either starts out as just two or three lines or it evolves to the point where I feel like I need a "real" language. At that point, I'll end up hacking it in perl or python.
I would like to see the interactive Read Eval Print Loops of one of these languages finish closing the gap and become good shell replacements. I find I can get close to that sometimes with interactive programming environments, but not quite.
I've been thinking about this also and was planning this week to change one of my login shells to just pythn to trial it. I've now got two small line-based editors running in python (one is an ed clone, the other is simpler and uses list comprehension), and it's simple enough to write wrapper functions around the bits you need to change directory and spawn process. There's no standardisation of effort, but it's completely doable just on your own system. I realise it's probably not at the top of your mind, but can you think what you found missing when you last tried it?
In the meantime, I'm happy enough with bash.