Hacker News Reader

Comment by edgyquant | Hacker News Reader

What are people’s workflows for using eMacs with generative coding? I’m trying to get a workflow where I can generate code and approve it like a PR (maybe leave comments for a next round of generation) but I’m new to eMacs and the ai coding package isn’t working with the latest eMacs

View on HN

Agent Shell connected via ACP to local Claude Code.

https://github.com/xenodium/agent-shell

Can't say enough good things about agent-shell.el! Xenodium has done amazing work there.

I'll have to try it again. It definitely feels like the right way to do things, but when I tried it it had a couple of show-stopper limitations and a show-stopper bug, so I've been using claude-code-ide.el instead.

I'm a huge fan of agent-shell, too. It keeps getting better and better, and it's infinitely customizable, like any good Emacs package.

Wonderful to hear. Thank you sir!

Seconded, agent-shell is my favorite LLM interface bar none

Have you tried https://eca.dev ?

I haven’t! It seems akin to ACP, is that accurate? Do you recommend it?

No, ECA doesn't work like ACP, it's provider-independent and operates similar to an lsp-server. It is in fact, written by the maintainer of lsp-mode and author of clojure-lsp.

I've tried many different tools for AI tasks in Emacs, eca and gptel remain the bread-and-butter of my workflows.

I use windows 11, wsl2, WezTerm.

emacs -nw directly on the shell.

its like working on a linux computer, but from windows.

I run vterm inside emacs, and claude code inside vterm. I have special commands for extracting file routes, and flow between dired and magit. Everything runs like magic.

I also have copilot-chat and copilot-mode installed. From time to time I use it to review what claude has done, or to chat a little bit.

I’m trying to use a similar workflow, but when I attempt to use Claude code in vterm, I have issues.

My use case is resuming a session and not being able to scroll back to the past history. Otherwise if I start a new session, it works fine.

It’s probably related to me being on an old version of vterm, but I’m just wondering if you ever have come across this.

Definitely try ghostel, it's really solid.

Thanks, I would, and I’ve used it on my personal machines (it works great), but I’m trying to stick to the apt package repos on my $WORK machine.

Ubuntu has a fair number of elpa packages in its repos, so I’m trying to make it work if possible. Not sure if I’m missing anything config wise within the confines of vterm

I found my answer, I needed to be using page up and page down keys to “scroll” through the claude history - which is apparently required due to Claude capturing my C-v and M-v’s in its “text-box”.

I’ve been using gptel, I’ll have it generate org babel blocks I can tangle (write to files) or have it use tools to write and edit files in the project I’m working in, and then review with magit.

I use cLaude cOde in my eMacs (through ghostel, which is unbelievably great!) to generate code :-)

I then use mAgit to review it.

The iPhone autocorrects it to eMacs... it's so annoying.

https://en.wikipedia.org/wiki/EMac

Setting up a keyboard replacement is the way to go here. Solves the issue for all devices.

IMO, generative coding on Emacs works nicely because magit is awesome for code review. A good agent integration package is going to have a command you can use to send the current cursor line & file to the agent so you can reference and comment.

Which AI coding package? Agent-shell? Gptel-agent?

It’s just called ai-code I’m new to eMacs so this popped up but had a bunch of issues I was trying to debug https://github.com/tninja/ai-code-interface.el

Never heard of that one. The most popular ways for agent stuff seems to be https://github.com/xenodium/agent-shell , or just opening a terminal in emacs and using claude code. If you want to run it through an emacs terminal, consider using https://github.com/dakra/ghostel#ghostel (or vterm, or eat)

Thanks for posting the ghostel link, I hadn’t heard of it. Excited to try it out and see how it compares to vterm

Reply on news.ycombinator.com