My workflow is very similar. But I just ask agents to write design in html instead of markdown, due to its richer layout and better interactivity. When the design is about UI or anything related to graphics, this approach is extremely efficient.

I also found that having the design reviewed by multiple agents has very little marginal value. The review agent will always find something to improve, but mostly it’s just nit and not anything super important.

I used to let Claude just upload the html design doc to Claude artifacts for me to review. Recently I switched to codex and started to use my own tool https://github.com/hyperlogue/r3 to complete this workflow.

This looks really cool, especially on the remote control part. I've experienced exactly the same problem, but ended up building my own tool through an opposite architectural approach than Saggar.

The main difference is that I want to keep my existing workflow in Kitty and tmux/zellij. Since all of them provide solid remote control protocol, I wrote captain-miao (https://github.com/hyperlogue/captain-miao) which is a standalone TUI app that runs inside these terminals, observing the agent sessions, managing tabs and windows, and allow me to quickly jump to the exact window/tab that needs my attention. The sessions are still native windows/panes inside the terminal/multiplexer that I configured and polished for years.

One difficulty I found when trying to expanding the agent support beyond claude code and codex, many agents don't provide a good way to inject hooks to listen to the session status change. wondering how Saggar solves this problem. do it infer state from the screen buffer?

Looks great, well done! Kitty is great too, makes sense.

I have purposefully kept things quite tight for my own sanity, but I think it makes sense market-share wise at the moment.

At the moment I'm also assuming that if someone is using Cursor AI tooling/Grok, they're probably using Cursor IDE and terminals in there.

I am planning on looking into integrating Pi.dev, which I love the philsophy of (though don't use personally just because API usage costs so much).

Just an anecdata: I really dislike the Cursor IDE's session/worktree handling, but my company only provide Cursor licences. So I'm using herdr, cursor cli ("agent"), a custom workflow skill that reads a manifest of the repos I work with + the setup commands they need.

Ah interesting! Thanks for sharing.

The remote control appears to use a third party relay, and I don't see encryption mentioned.

I really want to use this, but I definitely don't trust a vibe coded relay to my shell.

Fair concerns! The third party is Cloudflare, and it is encrypted