Talk to an agent in a pane

Each pane in the terminal tab is a live Claude Code session running in that agent’s home folder.

Opening one

Click an agent in the terminal tab. If no session is open, the board starts one; otherwise you attach to the open one.

A pane outlives its ticket and the engine that dispatched it; only the board closing ends it.

Typing into it

Two ways in:

• keystrokes, delivered raw and unsubmitted: Esc is a keystroke, not three characters of text
• a paste, delivered as a bracketed paste plus Enter, the only way a multi-line prompt survives a box that submits on newline

The engine pastes a ticket into a live pane; a fresh session gets it as a command line argument.

The dock

The panel docks left, right or bottom. The splitter resizes it between 30 and 60 percent of the window; a toggle takes it to 90 percent.

Reconnecting

The page pings its socket every 15 seconds and treats 40 seconds of silence as a dead link. On a dropped transport it re-attaches to the same terminal, backing off through 400ms, 900ms, 2s, 4s, 8s and then 15s.

Three close codes mean the session is over. None is retried:

Code	Meaning
1000	the board closed the pane deliberately
4404	there is no session for this agent
4500	a pane could not be opened

Every retry is attach-only. Only your click may spawn a session.

Permission mode

Pane sessions run in permission mode auto. The generated wall decides what the agent may touch; you are asked only about what it names neither way.

When the board restarts

There is no reattach across restarts: every pane closes with the board. Click an agent to open a fresh one.

Where it does not work

A pane needs a pseudo-terminal, and the board reports what this machine supports. The terminal is vendored, so nothing is fetched from a CDN at runtime.