swarm（swarm 搜索）

Swarm

Fan out a list of tasks to subagents with bounded concurrency, collect results, and return a compact summary.

Loading

**The REPL's eval tool supports ES module imports from @/skills/*. Use await import("@/skills/swarm") to load this skill — the REPL installs a custom module loader that resolves those paths against the skills backend. Do not** inline index.ts into your eval body. Importing is the supported, tested path; copying the source is an anti-pattern that duplicates logic and drifts on skill updates.

When to use

You have many independent tasks (e.g. "summarize each of these 20 files", "classify each of these 50 tickets", "research these 15 topics") and want them to run concurrently rather than sequentially.

Usage

runSwarm(...) returns a summary object, not an array. Destructure .results for the per-task output — the summary itself is not iterable.

const { runSwarm } = await import("@/skills/swarm");

const { results, completed, failed } = await runSwarm({
  tasks: [
    { description: "Summarize /notes/alpha.md" },
    { description: "Summarize /notes/beta.md" },
    { description: "Summarize /notes/gamma.md" },
  ],
  concurrency: 3,           // optional, defaults to 5, capped at 10
  subagentType: "general-purpose",  // optional; per-task override wins
});

console.log(`${completed} ok, ${failed} failed`);
for (const r of results) {
  console.log(r.id, r.status, r.output ?? r.error);
}

Contract

The runSwarm(opts) function accepts:

• tasks (required): an array of {description: string, subagentType?: string}.
• concurrency (optional, default 5, capped at 10): max parallel subagent invocations.
• subagentType (optional, default "general-purpose"): the default subagent to dispatch each task to. A task's own subagentType takes precedence.

Returns a summary object:

{
  total: number;          // tasks.length
  completed: number;      // subagents that returned a result
  failed: number;         // subagents that threw
  results: {              // one entry per task, in input order
    id: number;           // 0-indexed position in `tasks`
    status: "completed" | "failed";
    output?: string;      // on success — subagent's final message
    error?: string;       // on failure — error message
  }[];
}

Design notes

• Dispatch goes through tools.task, which the REPL's PTC layer exposes for us. The skill does not register any subagent itself — it's a fan-out pattern on top of what the agent already has.
• Failures are caught per-task: one failed subagent does not abort the swarm. Check the failed count and the per-task status.
• Concurrency is bounded with a semaphore-style pool rather than Promise.all on everything. For 100 tasks with concurrency=5, this keeps memory and tool-call rate predictable.