Comment by Keyframe
8 days ago
Wouldn't it be wonderful to have a tool where you interact with AI interactively through the codebase via IDE / vim / emacs tree? Say, you open your codebase and start with prompts and AI+tool navigates to a function or a place where it needs to and modifies stuff while chatting to you about it? Or you jump to somewhere, highlight where you are to scope down the focus of it (while it still retains all of the code in history/memory). Sort of like pair programming. It sounds so obvious that I'm almost sure I've missed that already existing somewhere. I think I tried google's thing (forgot the name) but it sucked / wasn't that.
I think you’re describing Aider.chat. There are 2 Emacs packages for it, one official and a very recent fork. Aider is a cli so it works great with vim as well.
In Emacs I’ve had good experience with gptel as well but I prefer aider for the coding workflow
Yep, I've particularly been enjoying the recent "watchfiles" feature where a comment can be added to the source file, and ending it with "ai?" or "ai!" triggers use of said comment as a prompt to ask about or change that section upon save.
I'll check it out, thanks!
Apologies if I'm missing something, but aren't you describing Cursor/Copilot/Windsurf?
you're not. looks like that's kind of it, but would the thing have the context of the whole project when I'm in a file/class/function? With copilot, in my case, it was so far mostly like a fancy autocomplete that has immediate vicinity in its memory where it would be vastly more useful if it had the context of the whole project / all files.
Cursor indexes the entire code use with embeddings. It works well in small single app projects
1 reply →
the vscode extension cline also does this
This doesn't sound good to me, you end up with a large codebase that no human has actually laid eyes on. When you get a bug weird enough that you can't reason the LLM through it, then what? What if a bug is because of interactions between two systems, and you don't own one of them? What if there's an issue due to convoluted business process failures, that just end in a bug report like "my data is missing!"? I honestly think in the latter case, the LLM will just fix a 'bug' and miss the forest for the trees.
I prefer the idea of the other comment reply where you use AI as a tool to explore a codebase and assist you, not something you instruct to do the work. It can accelerate you building that experience and intuition at a level we've never been able to do before.
An llm itself is a large codebase that no human has laid eyes on, instead you validate it through testing.
Regarding testing, I’ve had an interaction with windsurf where I told it there was a bug in the application it generated. It replied “I’ve added some log statements, can you run it and tell me what you see, then I’ll know what to fix”… The llm was instructing me…
Nothing like that at all. For example I have a few codebases kind of large (for certain quantity of large) where I know the code since either I wrote it or participated heavily in. Talking snippets at a time loses a ton of context which would yield better offered solutions if you had, well.. the whole context.
I tried various solutions but I still haven’t found a chat tool that allows me to navigate a large monorepo. I’d like to be able to say "open the file where there is the function to do <xyz>", but current tools don’t understand that.
This works fine in Cursor. As far as I know, you can't say "open the file..." but you can say "where is the function to do <xyz>" and it'll include a link to the file in it's response and then you can click to open it.