Switchboard

Every machine runs a small runner that dials out to the hub, like a handset lifting off the cradle. Nothing ever connects inbound, so there are no open ports, no VPN, and no firewall tickets. To add a machine, mint an invite code; one paste on the new machine installs, configures, and joins it in a single motion. The hub never logs into anything. Machines call it.

One machine is the master; the rest lend their agent as workers. Delegate a task to a single worker, or group several and fan one task out to the whole group so each agent takes its turn in parallel. Steer any line mid-flight, or watch every keystroke with the wheel in reach. A roster lists each machine by stable node id, capability, and baton holder: exactly one hand drives a worker at a time. A dropped line suspends, never kills, its session, so the work is still there when it dials back in.

Workers build in their own sandboxes: no git install, no GitHub account, no credentials on any worker. Bind a repo once on the master and every delegated task replies with a branch and a diffstat; the branch lands on your machine automatically as a sealed pack (AES-256-GCM) the hub stores but cannot read. Review the diff locally, then integrate. Conflicts are never auto-resolved. Optionally mirror to GitHub: the master alone pushes, with the token in its local keychain, never on the hub.

With more than one worker on the board, the master drives the mesh the way a frontier agent drives sub-agents. Cut a goal into independent subtasks, route each to the cheapest capable machine, and fan one wave out to a group with a single /sb-delegate @group. Every agent takes its turn in parallel. Collect the replies and reduce them in one place. The master is the scheduler; the hub only routes. Have a different machine verify the result, then integrate. /sb-orchestrate loads the full playbook into the session.