multi-agent failure modes: Swarm Failure Modes

# detect/split_brain.py
def assert_single_plan(store, ticket_id: str, run_id: str):
    plans = store.list_plans(ticket_id)
    active = [p for p in plans if p.status == "active"]
    if len(active) > 1:
        raise RuntimeError(f"split_brain:{ticket_id}:{[p.hash for p in active]}")
    if active and active[0].run_id != run_id:
        raise RuntimeError("stale_run_holds_plan")

// detect/dupPr.ts
export function prFingerprint(title: string, files: string[]): string {
  const norm = title.toLowerCase().replace(/\s+/g, " ").trim();
  const f = [...files].sort().join("|");
  return `${norm}::${f}`;
}

export function findDupes(open: { title: string; files: string[]; url: string }[]) {
  const map = new Map<string, string[]>();
  for (const p of open) {
    const k = prFingerprint(p.title, p.files);
    map.set(k, [...(map.get(k) || []), p.url]);
  }
  return [...map.values()].filter((v) => v.length > 1);
}

# lease.py
def acquire_implementer(redis, run_id: str, ttl=600) -> bool:
    # ✅ SET NX EX — only one implementer mutates the branch
    return bool(redis.set(f"lease:impl:{run_id}", "1", nx=True, ex=ttl))