replay

Overview ¶

Command replay exercises Relay's /v1/messages endpoint against every turn from a captured Claude Code session JSONL. It reconstructs the realistic outbound request body that Claude Code would have sent at each assistant turn (full conversation history up to that point), fires them serially through Relay, then reports overhead p50/p99 harvested from Prometheus.

Serial mode only — p50/p99 reflect sequential latency, not under-load concurrency. Use bench/bench_test.go for load numbers.

Usage:

go run ./bench/replay --session <path> [--relay http://localhost:8080] [--key test-key]