Skip to content

Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.

1.7 KB
Raw Blame

The ⌘M chat view answers back. You can approve a plan — accept-edits or manual — send one back for revision, and pick AskUserQuestion options, all from the chat view, without a keystroke going into the terminal pane. It rides the existing PermissionRequest hook channel. The same change fixes a silent no-op that was worse than it looked: the node header's ✓ Approve, and the phone's allow, did nothing on a plan, because Claude Code drops a bare allow for tools that require interaction — the answer has to carry the updated input with it.

And it reads like a conversation. The panel gained a claude.ai-style thread — user bubbles, unbubbled assistant text, a turn action row — and a real composer with file attach, dictation, and model · effort labels. A working indicator and live refresh show the agent answering, and a long transcript now loads in pages instead of making you wait for the whole file.

It also got cheaper to leave open. CPU and GPU drop during camera moves, animations, polling and autosave: the viewport raster scale freezes while the camera moves, the minimap projects transform-only, the working glow settles after four pulses, SSH transcript tails back off while a transcript is idle, and scrollback snapshots are paced and skipped when unchanged. A co-attach joiner no longer loses tmux's alternate screen, and links inside rendered markdown can no longer navigate the app window away from your canvas.

One note for existing installs: the managed hook script updates to revision 5, and its PermissionRequest entry gets an explicit timeout: 600 — without it a plan or question cannot be held open long enough for you to answer it.