Excellent share, thank you. My question is with your setup, how strictly does Claude Code adhere to using this mode to traverse the codebase over grep? I have found this is to be a huge issue when implementing similar solutions... it loves to just grep.

> If I had to guess 60 million is just enough to build the POC out.

You're kidding.

This looks great. Could use built in safety from destructive actions by LLM?

That’s exactly how it’s designed. The flow is:

- you type a command that starts with yo

- the clanker thinks

- the clanker comes back with a shell command that fits your yo command and fills it in as if you had retrieved it from your shell history by pressing the up arrow

- you have to press enter to actually execute the command. Or you could edit the command just like you can edit commands retrieved from your shell history.

I personally find this approval flow to spark more joy than what the other agent TUIs and CLIs do - they usually pop a modal menu dialog with yes/no/something else. And that’s jarring, because modality is a jarring UX. What yosh does feels groovy because it so so much like just retrieving something from history, or like a speedrun of opening a browser, asking Google or a clanker, and copy pasting.

Link: https://github.com/BrowserBox/BrowserBox

What about context loaded in from tool calls?

You can load context from wherever you want (including inside other contexts and/or external tools), just uc.get() your way into the context. You can use metadata to filter your tool calls. You pick whatever structure works best for you. UltraContext is just the building blocks for your app! What are you building?

Orbstack VM.

How do you know without a doubt you're tracking the same person?

Let us all generate our own custom cursors.

May as well just use Claude Code then.

Well, I do use Claude Code myself, but I'd thought the point of OpenCode was that it could combine the responses of multiple LLMs.

What specifically is slop therein?