The Finder-style columns for tool calls, approvals and raw context is the right metaphor - most agent harnesses hide that structure behind a chat log, and being able to actually see thread branches side by side made it much easier to follow where a sub-thread diverged. Single Go binary and no account/cloud dependency is a big plus for something that's going to be running with real tool-call permissions on my machine. The JS plugin model for tools is a nice touch if you want to wire in something custom.
left a review above - the "got annoyed at a tool and built my own" origin story tracks with JUCE and Tracktion, and it shows in the attention to the Finder-column layout here. one honest note from actually using it: the docs for writing a custom JS plugin are thin right now, I ended up reading example plugins in the repo instead of following a guide. a short "write your first tool" walkthrough would go a long way given the whole pitch is that every tool is pluggable.