/**
 * Agentic swarming — fan out N specialized agents on the same task
 * concurrently and merge their outputs.
 *
 * Inspired by LangGraph's `Send` map-reduce primitive and OpenAI Agents
 * SDK's handoff pattern. The audit ranked these as the two most-portable
 * concepts from the swarming-framework ecosystem; this is the map-reduce
 * half (Send-style fan-out). Sequential handoff stays on the roadmap
 * for /multi-execute.
 *
 * Usage:
 *   /swarm code-architect,silent-failure-hunter,type-design-analyzer  audit the auth flow
 *
 * Each agent receives:
 *   - System prompt: its own ECC agent prompt (the role-specific persona)
 *   - User message: the task verbatim
 *   - Empty tool list: swarm runs are analysis-only; no file edits or
 *     bash commands. If you want a swarm to make changes, /architect (or
 *     other agent slash) first, then act on the synthesis manually.
 *
 * Concurrency: Promise.allSettled — one agent failing doesn't kill the
 * others. Errors surface in the result block.
 *
 * Key pool: every concurrent agent uses streamChat which goes through
 * the key rotation pool from v1.23.0. Users with multiple OpenRouter
 * accounts get true parallel throughput across keys.
 *
 * Cost note: N agents = N model calls. The orchestrator's /help text
 * surfaces this so it isn't surprising.
 */
import type { CrowcoderConfig } from './types.js';
export interface SwarmAgent {
    /** Display name (matches the ECC slug like 'code-architect') */
    name: string;
    /** The system prompt the agent runs with — usually the ECC skill body */
    prompt: string;
}
export interface SwarmResult {
    agent: string;
    text: string;
    durationMs: number;
    /** Set when this agent's call threw — others may still have succeeded */
    error?: string;
}
/**
 * Look up agents by slug and return their full prompts. Throws on the
 * first unknown slug so the caller can correct + retry rather than
 * partially running.
 *
 * Accepts both bare slugs ("code-architect") and the "agent: <slug>"
 * form that listSkills() returns.
 */
export declare function resolveAgents(slugs: string[]): SwarmAgent[];
/**
 * Run all agents concurrently against the same task. Each agent gets a
 * private message history (its prompt + the task); they don't see each
 * other's output. Merging is the caller's job — usually print-with-
 * attribution + an optional synthesis prompt.
 *
 * No tools available to swarm agents — analysis only. If we exposed
 * tools we'd need lock coordination on the filesystem + permission
 * prompts for every parallel write, which defeats the purpose.
 */
export declare function runSwarm(agents: SwarmAgent[], task: string, config: CrowcoderConfig): Promise<SwarmResult[]>;
/**
 * Format swarm results for stdout — attribution headers + clear visual
 * separation between agents. The caller can pipe this into a follow-up
 * synthesis prompt if they want consensus or comparison.
 */
export declare function formatSwarmResults(results: SwarmResult[]): string;