pHaeusler
Direct terminal interaction
Why not let it simply use the terminal directly?
Then you don't have to keep implementing and maintaining fixed actions it has to use, and it can use existing training rather than wasting prompt space to explain your api as it grows.
GPT seems to already have a pretty good grip on cat, echo, git, etc, and if you put it behind a docker container it could even install tools to help it self and create it's own dev environment.
Safety comes to mind of course but that's fairly solvable by having human in the loop for now (which also lets you help it debug problems).
I don't yet have GPT4 api access to try this, but I've been playing with this concept using the turbo model and an app I made here ModelProgrammer and it seems to get stuck in weird circles like modifying it's tests rather than the code under test that I find having a human to guide out of is helpful anyway.
That's an interesting idea - worth experimenting
My intuition is that it would be too generic and be difficult to get working repeatedly
So far it only seems to take a few examples of editing it's replies before it starts to get things like, using echo to reply to you (that's harder for it), and then telling it to do things like commit it's work or write out to file, etc all get picked up very quickly (literally just tell it don't forget to commit your work and it'll go and do that). A hard one is getting it to write commands short enough to fit in it's output limit but, it doesn't take more than a context length to do it, and that's without me compressing the terminal output which currently includes a copy of all of it's input.
It's very helpful to be able to guide it through by editing what it sent though, that's why I just went and made a GUI for this.