Comment by pama
10 hours ago
My main take is exactly the opposite. Why not build everything with a simple text interface (shell command) so the models learn to use these tools natively in pretraining. Even TUI like codex-cli or claude code are needless abstractions for such use cases and make full automation hard. You could add as many observability or input layers for humans as you want but the core should be simple calls that are saved in historical and documentation logs. [the headless/noninteractive modes come close, as do the session logs]
It would be cool if I didn't have to worry about whether I was "in" or "out" of the AI TUI. Right now, I need at least two terminals open: One running my shell, that I use to issue shell commands, and one running Claude, where I can't. It would be nice if it could just be my shell, and when I wanted to invoke claude, I'd just type:
Right in the shell.
isn't that what Simon Willison's `llm` does?
edit: [link](https://github.com/simonw/llm)
Most AI agents have a 'bash mode' and, you can use Warp terminal which is terminal first, but easy to activate the AI from the terminal. For example, if you mangle a jq command, it will use AI to suggest the right way to do it.
-p
Oh wow, nice. Does it remember context from run to run?
TUI is easy to train on, but hard to use for users. Part of the reason it’s easier to have LLMs use a bunch of Unix tools for us is that their text interface is tedious and hard to remember. If you’re a top 5% expert in those tools it doesn’t matter as much I guess but most people aren’t.
Even a full-featured TUI like Claude Code is highly limited compared to a visual UI. Conversation branching, selectively applying edits, flipping between files, all are things visual UI does fine that are extremely tedious in TUI.
Overall it comes down to the fact that people have to use TUI and that’s more important than it being easy to train, and there’s a reason we use websites and not terminals for rich applications these days.
I use headless mode (-p) and keep my named shell histories as journals (so no curses/TUI or GUI). But session management or branching could improve at the tool level and allow seamless integration with completion tools, which could be a simple fast AI looking at recent sessions or even be visual, say for navigating and extending a particular complex branch of a past session. It is not too hard to work with such shell-based flows within Emacs for me, but it would be nice if there was some standardization effort on the tool fronts and some additional care for future automation. I dont want my AI clicking buttons if it can be precise instead. And I certainly want multithreading. I think of AI more as an OS; it needs a shell more than it needs windows at this point in time.
I agree. I like using Claude or Codex in VM on top of the tmux. Much more flexibility that way. I open a new tmux window for each issue/task big enough to warrant it, issue a prompt to create a worktree and agents and let them go to town. I actually use claude and codex a the same time. I still get observability because of tmux and I can close my laptop and let them cook for a while in yolo mode since the VM is frequently backed up in proxmox pbs. I am a retired hobbyist but this has been a nice force multiplier without devolving a complete viby mess. I hope these new orchestration tool support this like vs code remote development does. Same for cloud. I want them to support my personal "cloud" instead of laggy github mess.