/**
 * Real-API speed benchmark — sends actual requests to a live provider and
 * measures TTFT, output throughput, cache behavior, and wall-clock latency.
 *
 * Unlike the mock benchmark, this hits real endpoints and produces
 * production-representative numbers. Uses gg's own AuthStorage to resolve
 * credentials (same path as a real session).
 *
 * Usage:
 *   npx tsx src/core/api-benchmark.ts
 *
 * Environment overrides:
 *   GG_BENCH_PROVIDER  — provider name (default: "glm")
 *   GG_BENCH_MODEL     — model id (default: "glm-5.3")
 *   GG_BENCH_TURNS     — number of turns (default: 5)
 */
export interface ApiTurnMetrics {
    turn: number;
    prompt: string;
    ttftMs: number;
    outputTokens: number;
    outputDurationMs: number;
    tokensPerSecond: number;
    wallClockMs: number;
    inputTokens: number;
    cacheRead: number;
    cacheWrite: number;
    cacheHit: boolean;
    stopReason: string;
}
export interface ApiBenchmarkResult {
    provider: string;
    model: string;
    turns: ApiTurnMetrics[];
    totalWallClockMs: number;
    totalTtftMs: number;
    totalOutputTokens: number;
    avgTokensPerSecond: number;
    avgTtftMs: number;
    cacheHits: number;
    cacheHitRate: number;
}
export declare function runApiBenchmark(config: {
    provider: string;
    model: string;
    apiKey: string;
    baseUrl?: string;
    accountId?: string;
    cacheRetention?: "short" | "long";
    promptCacheKey?: string;
    turns?: number;
}): Promise<ApiBenchmarkResult>;
export declare function formatApiResult(result: ApiBenchmarkResult): string;
export interface ApiComparisonResult {
    baseline: ApiBenchmarkResult;
    optimized: ApiBenchmarkResult;
    ttftImprovement: number;
    throughputImprovement: number;
    wallClockImprovement: number;
    cacheHitRateDelta: number;
}
/** Format an A/B comparison between two API benchmark runs. */
export declare function formatApiComparison(c: ApiComparisonResult): string;
//# sourceMappingURL=api-benchmark.d.ts.map