[anotherpaulg]

This looks super interesting, thanks for sharing. I like that you're exploiting the new functions API to give GPT agent-style access to explore a codebase. I have played with that previously with gpt-3.5 and plan to do some more experiments with gpt-4 someday soon.

I am also working on an open source CLI tool in this space [0]. I've taken a different approach, more focused on chatting with GPT to have it edit the code in your local git repo.

But my tool also provides GPT with a semantic map of your repo and the ability to ask to see particular files, etc. I use it to answer questions about unknown codebases all the time, and then start asking it to make changes. I have a chat transcript that illustrates that here [1]. As another example I needed a new feature in the glow tool and was able to make a PR [2] for it, even though I don't know anything about that codebase or even how to write golang.

Also, there's a small discord [3] where a few of us working on "AI coding tools" have been sharing ideas. You might be interested in joining the conversation over there.

[0] https://github.com/paul-gauthier/aider

[1] https://aider.chat/examples/2048-game.html

[2] https://github.com/charmbracelet/glow/pull/502

[3] https://discord.gg/fHcgCRGu

[narenkmano]

Thanks! Yeah, it's been a lot of fun hacking around with these models. I just checked out your open-source project; looks super interesting! It has a lot of things I've been wanting to implement. I particularly love the command completion and concept of a repository map.

The approach you've taken is the one I use the most in my workflow nowadays. I manually provide context to GPT4 and ask it to translate/fix issues in my code (at least provide high-level ideas about resolving issues). Pretty cool to see a CLI version of that.

Having the human in the loop has been crucial in practice. Your approach of having the LLM interact with your code base and collaboratively make edits is interesting.

Joining the discord; thanks for the invite!