cactus：用非阻塞决策队列替代 90% 的 Claude Code 使用

sped up 20x, cactus on the left, claude on the right, file finished at the end

mod available! ( working for weeks to release into a new harness capability feels like the future )

I was having serious issues reading Claude prose, like my brain would start fuzzing over on every 'it's not X, it's Y'. I tried the skills for addressing it inline, but it didn't work for me. ( 👋 hands up if you're allergic to AI prose too )

Forcing structured output is the only way to counter Claude's tendency to speak to you as though you are a machine, and frankly, I'd rather machines write for machines anyway.

Cactus is a few things:

1. A cli utility for AI to interact with to generate and iterate upon steering questions.
2. A tui for humans to interact with that collects questions from ALL your agents in your machine, grouped by project. Load context into your brain once, get all your agents for a project moving again.
3. Some hooks to provide information, check that questions are asked, and remind claude about open questions
4. A claude code mod to provide the q/a surface in the sidebar for a given session, enabling it also provides answer injections, so cactus doesn't need to keepalive an awkward background process.

One of the great things about working like this is I don't need to wait for a finished response, nor do I need to speed read to steer the conversation. As you can see in the gif, claude keeps writing questions after I've answered some, and the final response is written only when everything is complete.

Herdr (https://herdr.dev) is also awesome, and enables some tui <> agent interactions in cactus.

https://github.com/eighteyes/cactus

I'd love to know your thoughts. Thanks for clicking.