Agent teams

A single agent is a copilot. Novalis is built for crews, and it grew that capability in three phases, all shipped.

Observe (Phase A)

Any Claude Code agent team on your machine lights up in Novalis automatically:

• A team tree in the sidebar (⌘⌥T); leads, teammates, live status.
• A Kanban detail pane for the task board, and a read-only transcript pane per teammate.
• A status chip in the bottom bar so a background team’s progress is one glance away.

No setup: Novalis watches the standard team directories and renders what it finds.

Spawn into panes (Phase B)

When a Claude Code lead spawns teammates, a PATH shim translates its multiplexer calls into native Novalis pane operations over a UID-scoped Unix socket. Teammates land as real panes: split, tab, or workspace, your choice. The bridge is opt-in while it hardens; one key turns it on:

[teams]
enabled = true           # the Phase B bridge is opt-in (default off)
spawn_layout = "split"   # or "tab" | "workspace"

You watch every teammate work in a real terminal pane; scroll it, search it, interrupt it.

Native teams (Phase C)

Novalis can run the whole crew itself. Teams: New Native Team in the palette:

• Pick a lead: an LLM lead running its own conversation, or manual mode: you are the lead, from your own pane.
• Teammates run as independent agent conversations, claim tasks through a lock-protected task store, and exchange typed protocol messages through mailboxes.
• The same sidebar, Kanban, and transcript chrome lights up whether the team is external or native.

Approval policies apply to every teammate individually. A crew is still directed by humans.