Profiled the H2 client and removed the main client-side hot spots:

• completeRequest: fast-path in-order completions by advancing kRunningIdx (like H1) instead of an O(n) indexOf+splice; out-of-order completions keep the splice path
• use on instead of once for self-removing stream handlers, and skip listener removal on already-closed streams (handlers null-guard on the stream state)
• single pre-shaped stream state object (stable hidden class), finalizeRequest/openStream hoisted out of per-request closures
• cache the session ref state to skip redundant ref()/unref() calls (session ref proxies to the socket)
• memoize getProtocolFromUrlString

Benchmarks (single connection, local H2 server, interleaved A/B vs main): +5-6% req/s at 100 parallel requests, +27% at 1000, +51% at 2000; +15% with a 10-connection Pool. Remaining cost is dominated by node core session.request() internals.

🤖 Generated with Claude Code