launch-team

Launch a **team** of Claude Code agents — one persistent session per pane — to work a single task together. You are the **conductor**: pick the roles, write the brief, start the team. From there the agents coordinate themselves, each messaging teammates by typing into their panes. `launcher.mjs` does the deterministic setup (panes, role prompts, the shared roster); your judgment is the roles and the brief. This skill only works **inside a supported multiplexer session** (herdr or zellij). The transport is selected automatically (herdr is the default; override via `transport.config.json`). If no multiplexer session is active, stop and tell the user to start herdr (or zellij) first. ## Step 1 — Decide the team From the arguments, settle the roles and how many agents. One agent is the **orchestrator** (decomposes the task, delegates, integrates); the rest are workers. If the user gave only a task, propose a default and confirm: - `orchestrator` — Designer & Orchestrator - `implementer` — Implementer - `reviewer` — Reviewer Use lowercase kebab-case ids (they become pane names). Keep the team to 2–4 agents. Done when: the agent ids, their roles, and which one is the orchestrator are settled. ## Step 2 — Write the brief and config Write a **brief**: a self-contained description of the goal. Reference existing artifacts (PRDs, plans, issues, diffs) by path/URL instead of pasting them; redact secrets. Then write a **config JSON** to the OS temp directory (not the workspace):...