/start — Hand a task to the team

Usage:

/start "<plain English task description>"
/start --resume                     # restart the daemon after a crash

What this does

/start "<task>" is a thin wrapper around the repo-root team.sh:

./team.sh start "<task description>"

That command:

1. Confirms the daemon isn't already running (or cleans up a stale lock).
2. Runs the planner agent in foreground. The planner splits your task into one or more sub-tasks and writes .state/active.json plus a .state/tasks/<id>/meta.json per sub-task.
3. Spawns the daemon (pipeline-daemon.py) in the background. The daemon then runs the rest of the pipeline end-to-end — analyst → architect → developer → (3 reviewers + tester + security in parallel) → qa → documenter → retrospective.
4. Returns control to you. Watch progress with /status or ./team.sh logs --daemon.

When to use it

• You have a feature, bugfix, or improvement and you want the team to ship it.
• The planner detects the shape from your phrasing (bug-shaped, feature-shaped, improvement-shaped) and tags flow_type accordingly — no separate slash command needed. Round 4 folded /feature, /bugfix, /improve, /local-loop, /process-issues into this single entry point.
• For batch work, queue tasks in .state/backlog.jsonl with team.sh backlog add "<task>". This is a manual queue — the daemon does not consume it automatically. Pop the next entry yourself with team.sh backlog peek (or team.sh backlog list to see all pending) and hand it to /start "<task>".

--resume

If the daemon crashed mid-task, /start --resume (or ./team.sh resume-daemon) brings it back. Idempotency in every agent (role_done.<role> flags in meta.json) means agents that already finished will exit immediately on re-spawn; the daemon picks up exactly where it stopped.

Backlog queue

There is no daemon-side autopilot and no --backlog flag on /start or team.sh start. .state/backlog.jsonl is a plain manual queue managed via team.sh backlog (add / list / peek / purge-consumed / clear). To work through it, team.sh backlog peek the next entry yourself and pass it to /start "<task>" — there is no automatic dequeue.

Safety limits (configurable)

LSD_POLL_INTERVAL          = 3 sec        # how often the daemon polls .state/
LSD_MAX_PARALLEL_TASKS     = 3            # how many tasks run in parallel
LSD_MAX_PARALLEL_SUB       = 4            # subagents per task (review fan-out)
LSD_AGENT_TIMEOUT_SEC      = 900          # hard kill any agent past this

Override via env vars at daemon start time.

Notes

This is the same pipeline we run in production daily. The cookbook ships the exact daemon and agent definitions we use, anonymized. /start is the single command that kicks the team off.