/**
 * Prompt-cache warming micro-benchmark — isolates the cache effect.
 *
 * The existing api-benchmark grows the conversation each turn and uses a tiny
 * (~250 token) system prompt, which is BELOW Anthropic's minimum cacheable
 * prefix (1024 tok Sonnet/Opus, 2048 Haiku) — so caching never engages and the
 * A/B is dominated by turn-to-turn throughput variance.
 *
 * This test uses a large, FIXED prefix (well above the min) and a tiny variable
 * suffix, then measures, per arm:
 *   COLD  — fresh cache key, first hit (cache WRITE, no read).
 *   WARM  — prewarm the prefix, then measure the next call's cache_read + TTFT.
 *
 * It reports cache_read tokens and TTFT directly, so the cache effect is visible
 * regardless of throughput noise.
 *
 * Usage: npx tsx src/core/cache-warm-benchmark.ts
 * Env:   GG_CW_PROVIDER (default anthropic), GG_CW_MODEL (default claude-haiku-4-5-20251001),
 *        GG_CW_REPS (default 4)
 */
export {};
//# sourceMappingURL=cache-warm-benchmark.d.ts.map