team

Launch Agent Team: $ARGUMENTS

MANDATORY: Worktree-isolated executors (no team_name)

For any spawn that creates N ≥ 2 concurrent executors, EACH spawn MUST use:

• subagent_type: "executor" (.claude/agents/executor.md)
• isolation: "worktree" — the Claude Code harness primitive
• run_in_background: true — so all executors run concurrently
• NO team_name / NO TeamCreate. Combining team_name (the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS path) with isolation: "worktree" is silently downgraded by the harness: the executor is routed to the team coordination context and runs on the shared main working copy instead of an isolated worktree. This was observed first-hand (parallel executors clobbering src/app.py, a tree-wide git stash swallowing a sibling's uncommitted files). team_name is prohibited for parallel execution.

Coordinate via the agentId returned by each spawn (SendMessage accepts the agentId as to) and the harness's automatic <task-notification> on completion. Do NOT use TeamCreate, TaskCreate, or SendMessage-by-name.

Solo spawns (N = 1) MAY run on the main working copy and skip isolation: "worktree".

Protocol

Step 0: Resolve the team-lead HEAD SHA (binding)

Before any spawn, from the main repo root:

git rev-parse HEAD

Capture the full 40-char SHA. The harness allocates each worktree's branch from the default branch tip, not from the team-lead's HEAD — so an executor's worktree can start on a stale base. Thread this SHA literally into every executor prompt; the executor resets to it on entry (see Step 2).

Step 1: Provision ONE shared Postgres for the wave

Worktree-isolated executors each have their own working copy but agent-db.sh derives its container name from .claude/skills (constant across worktrees), so per-executor agent-db instances collide. Instead the team-lead brings up ONE shared Postgres and threads its DATABASE_URL into every executor prompt. The integration_db fixture creates a unique database per test, so a single server is safe under concurrency.