swarm

Use the swarm workflow: frame a done predicate, partition the task into slices or race arms, then spawn all workers in one message as background agents with worktree isolation. The parent waits, aggregates terminal results, and returns one consolidated report.

Slices partition work so each worker covers a different scope, while races give N workers identical briefs and select a winner. For races you must declare the selection rule up front: first pass, rank all, or best-of.

Use worktree isolation for any worker that checks out a branch, builds, runs the app, or writes files. Workers sharing the parent's checkout race each other, and remote isolation is reserved for workers that must outlive the session.

If a worker drops out, the swarm proceeds with N-1 workers and notes the dropout. Results missing the required SHAs or method are dropped and rerun once; after a second miss the gap is recorded and never counted as a pass.

The parent reads terminal results, validates evidence requirements, applies the declared race rule when used, and keeps a compact result table. The final report contains the table, one-line evidenced issues, and explicit gaps or dropouts, never raw worker dumps.