I wrote out a desktop-app concept around Pi's small coding-agent runtime.1 The motivation was the friction around switching workspaces and getting back up to speed, rather than adding another chat pane with a different skin.
The main screen I wanted was closer to a workbench: which repositories and worktrees are active, what changed, what is stale or blocked, and which branch owns the conversation.
Recover the state of work
The initial bets were a workspace overview, meaningful diff previews, a context recovery card and visible repository/task state. Provider configuration also needed to be easier to inspect than a hidden collection of settings.
Performance was part of the question, not a result. I wanted to test whether long conversations, mounted terminals, file trees and large diffs could stay responsive without every background update disturbing the whole interface.
This was scope-setting research. The next useful step was a small prototype around those workspace and recovery interactions, before expanding into broader multi-runtime ambitions. Writing down the product boundary gave that prototype something more specific to test than whether the app looked different.