Considering upcoming Claude changes

[melonamin]

I have a need for a wrapper like that. I tested it with Ralphex, and it seems to work just fine.

Total slop at this point, but I'm planning to clean it up https://github.com/melonamin/agentrun

@umputun this is just FYI; you're probably already planning something.

[umputun]

Yeah, I started to think about exactly the same approach - a shadow tmux session activated on ralphex start and killed on completion. Probably, for extra-paranoid detection prevention, some randomization patterns on input.

If you want to try to make it happen, it's all yours. I see it as an optional mode (well, at least optional for now) and some sort of intercept inside the ralphex flow (i.e., a proxy of some kind) at the execution level. This is just a preliminary design; I have not finalized anything. We can discuss if there are any questions, or you can just PR whatever version makes sense for you, and, as usual, I will grill it afterward ;)

[melonamin]

I see two possible shapes:

• ralphex stays as is, in a "safe zone." agentrun is optionally suggested as a drop-in replacement for the claude command (override already supported in ralphex).
• optional mode in ralphex itself, more narrow than agentrun - only needed features and nothing else.

You prefer the second one?

[umputun]

Both approaches seem fine to me. The first one is good because of the separation of concerns, and I like it. Probably, agentrun should support some startup/shutdown mechanism, so ralphex could control the life cycle, but this is a nice-to-have territory. The only potential problem with the first approach is that we need to deploy agentrun somehow, or at least provide instructions on how to install and set it up. The second approach keeps everything together, with no distribution concerns and clear flow control. However, this is an orthogonal concern to the core ralphex functionality.

The bottom line: I don't have a strong opinion on which way to go; either will do. If we can figure out how to address development/distribution concerns, I would pick agentrun. On the other hand, if the "only needed features" are simple enough, it can live inside.

[umputun]

In the last hour, I have seen a lot of similar solutions for -p emulations on both Twitter and Reddit. I guess this is a good signal to just allow users to bring whatever they want to act as a clade instead of putting our own implementation inside. And having yours as an "officially endorsed hack" in the documentation/FAQ and whatever other place you can think of.

[umputun]

Another idea that may be worth exploring for your wrapper: with a lot of workarounds published recently, almost all of them use the same tmux-behind idea. This is somewhat concerning, as the next logical step could be "we don't allow interactive sessions in tmux anymore." I know it sounds over-paranoid, but who could predict the recent -p changes? Who knows what they will do next.

An alternative is to use some kind of pty interface to wrap it. Something like https://github.com/creack/pty. Probably at this level, the ability to customize it and make it less prone to detection is better.

[melonamin]

Hey, I just wanted to let you know that it looks like I won’t have time for this over the next couple of weeks, so please don’t wait for me.

the logical step could be "we don't allow interactive sessions in tmux anymore." I know it sounds over-paranoid, but who could predict the recent -p changes? Who knows what they will do next.

This cat-and-mouse game is stupid, but we have to play it. I think the convergence point is somewhere in the future, where subscriptions just go away or they aren't much cheaper than API access. Hopefully, the API becomes cheaper, and not the other way around.

An alternative is to use some kind of pty interface to wrap it. Something like creack/pty. Probably at this level, the ability to customize it and make it less prone to detection is better.

I actually started with this, and in my limited testing, it works just fine for the Ralphex use case. I have another use case in which I basically need a persistent and steerable Claude session; that's why I built AgentRun that way.

Right now, from a user perspective, I can continue using Ralphex with agentrun or any other wrapper. So, as far as I’m concerned, you can limit this to a note in the README. Ralphex can, and probably should, stay the orchestrator and not spend time fighting the flicker company.

[pySilver]

This might be useful for ones who wants to create a wrapper on their own https://rmux.io/