If that's something you're worried about, review the code before running it.
> don't you get anxiety "what if there's an error in tui code and it would mess up my git repo"?
I think you might want to not run untrusted programs in an environment like that, alternatively find a way of start being able to trust the program. Either approaches work, and works best depending on what you're trying to do.
It takes more, not less, time to thoroughly review code you didn't write.
If we're talking receiving random patches where first you have to understand the context, background and so on, then yeah I agree, it'll take longer time probably than what it took for someone to hammer with their fingers. But again, I'm not sure that's how professionals use LLMs right now, vibe-coding is a small hyped world mostly non-programmers seem to engage in.
How can you "check" that which you don't "understand"?
> I'm not sure that's how professionals use LLMs right now
I'm a professional and I can tell you how I use LLMs: I write code with their assistance, they don't write code for me.
The few times I let Claude or Copilot loose, the results were heartbreaking and I spent more time reviewing (and then discarding) the code than what it took me to later write it from scratch.
??? I do understand, since I literally just instructed it, how would I otherwise? I'm not letting the LLM do the design, it's all me still. So the "understand" already exists before the LLM even finished working.
> I'm a professional and I can tell you how I use LLMs: I write code with their assistance, they don't write code for me.
Hey, welcome to the club, me too :) I don't write code, I write English prose, yet nothing is vibe coded, and probably I'll end up being able to spend more time than you thinking about the design and architecture and how it fits in, because actual typing is no longer slowing me down. Yet again, every line is reviewed multiple times.
It's more about the person behind the tools, than the tools themselves I think ultimately. Except for Copilot probably, the times I've tried it I've just not been able to produce code that is even slightly up to my standards. It's a breeze with Codex though (5.2), and kind of hit and miss with Claude Code.
But I do quickly check the output what it does, and especially the commands it runs. Sometimes it throws all code in a single file, so I ask for 'good architecture with abstractions'.
If `gh repo ...` commands get run you can lose everything instantly. You can force push and be left with a single blank commit on both sides. The agent has full control of everything, not just your local data.
Just set up Rclone/restic and get your stuff into a system with some immutability.
Ais are giving you what they get from common patterns, parsing documentation etc. Depending what you're asking this might be an entirely novel combination of commands never run before. And depending on the model/prompt it might solve in a way any human would balk at (push main to origin, delete .git, re-clone from origin. Merged local branches are gone!)
It's like the ai art issues - people struggle with relative proportions and tones and making it look real. Ai has no issues with tones, but will add extra fingers or arms etc that humans rarely struggle with. You have to look for different things, and Ai bugs are definitely more dangerous than (most) human bugs.
(Depends a little, it's pretty easy to tell if a human knows what they're talking about. There's for sure humans who could write super destructive code, but other elements usually make you suspicious and worried about the code before that)