[cyanydeez]

i dont understand. ssh allows a plethora of options for a text based IO,so why are we having to touch anything on a remote system.

i get what vscode is doing, but your agent just deals in text.

just build an ssh protocol and pipe it into the shell of the users choice. run parallel ssh and share connections.

just seems bizarre how we are happy to install complex servers for something that is litterally just piping text.

linux is here from it, why change the known secured interface .

[frenchie4111]

While I agree it is predominantly text getting sent over the wire, it would be pretty complex to implement harness on just a text stream without the websocket abstraction (I thought about it for a while). If you want to just SSH and run an agent you can, but the point of harness is that it has better ergonomics than a bare terminal can provide

[cyanydeez]

I understand; from a pratical standpoint, installing it on all servers I want to manage is just out of question.

Having a central AI with all it's trimmings is more reliable then custom servers everywhere.

[frenchie4111]

Agreed - my intention is definitely not "harness on every server" and more "make it easy for the backend half of harness to run on your personal dev box". This just comes from my experience at BigTechCorp where all of the engineers (myself included) used beefy EC2 instances for development (so tests ran fast, lol) and VSCode remote to actually edit files

[cyanydeez]

I've just installed a raspbery pi zero w/limited memory, and having to constantly correct the local model because it can't ssh correctly is annoying; but I can't install any kind of harness on it.

Being able to just keep a single instance and only worry about transport protocols would be a bonus.