Coordinator (this skill)
├─> Spawns Gamemaster (background, long-running)
├─> Spawns Player-1 (background, long-running)
├─> Spawns Player-2 (background, long-running)
└─> Spawns Player-3 (background, long-running)

All agents run in parallel, coordinate via files + hooks:
- Gamemaster writes turn-signal.json → Stop hook guides to wait
- Players use inotifywait (blocking) for turn-signal.json
- Players write player-actions/*.json → Stop hook guides to wait
- Gamemaster uses inotifywait (blocking) for player actions
- Zero polling overhead - all coordination via event-driven hooks

const gameName = "$0"; // e.g., "markovs-chains"
const numPlayersOverride = "$1" ? parseInt("$1") : null;

// Read game rules
const rulesPath = "games/" + gameName + "/RULES.md";
const rulesContent = await Read(rulesPath);

// Parse YAML frontmatter for config
// Extract: name, version, players, starting_cards, max_turns, win_condition

rm -f games/${gameName}/state/game-state.json
rm -f games/${gameName}/state/turn-signal.json
rm -f games/${gameName}/state/player-actions/*.json

mkdir -p games/${gameName}/state/player-actions
mkdir -p games/${gameName}/logs
mkdir -p games/${gameName}/traces

// 1. Spawn gamemaster (Sonnet for reasoning)
await Task({
  subagent_type: "general-purpose",
  model: "sonnet",
  description: "Gamemaster for " + gameName,
  prompt: gamemasterPrompt,
  run_in_background: true
});

// 2-4. Spawn players (Haiku for speed)
for (let i = 1; i <= numPlayers; i++) {
  await Task({
    subagent_type: "general-purpose",
    model: "haiku",
    description: "player-" + i + " for " + gameName,
    prompt: playerPrompts[i-1],
    run_in_background: true
  });
}