v0.1.4

What's changed

Performance work on the websocket receive path, plus two behaviours restored from 0.1.3.

Faster

Measured on CI, base vs head, three paired rounds on the same runner — 9 faster, 0 regressed, 37 unchanged:

• l2book_dispatch_50_coins −40.0%
• eip712_agent_digest_wasm −26.7%, eip712_multisig_digest_wasm −24.5%
• ws_request_round_trip −23.3%
• subscribe_200_coins −18.8%
• prepare_request −17.6%
• reconnect_resubscribe_burst −15.3%
• l1_action_hash_order_100_wasm −10.8%

What changed under those numbers:

• The socket no longer re-boxes every inbound frame. ReconnectingWebSocket was allocating a MessageEvent and paying a full EventTarget.dispatchEvent per frame purely to hand it to the SDK's own consumer, which then ran its own dispatch. That consumer now takes a direct frame hook; the public message event is still dispatched whenever anything outside the package listens. 1.42×, 305 ns/frame.
• fastAssetCtxs decodes natively. Where the runtime exposes node:zlib, a per-frame DecompressionStream + writer + reader + four promises collapses into one inflateRawSync, delivered in the dispatch tick rather than through a promise queue. 7.6× on delta frames, 1.8× on full snapshots. DecompressionStream remains the browser / React Native fallback, and both paths are differentially tested to agree — including multi-chunk payloads and every malformed input.
• Cheaper listener dispatch. A lone listener is stored unboxed and promoted to a copy-on-write array on the second registration, so multi-listener dispatch no longer copies into a snapshot buffer and clears it. 2.6× at 8 listeners, 1.18× at 1.
• One less serialization per subscribe. The dispatcher reuses the subscription id it already built by concatenation as the wire frame.

Large frames move least: JSON.parse dominates them and the SDK cannot touch that.

Fixed

• fastAssetCtxs could deliver the wrong frame. The payload was read from the recycled event shell inside a queued continuation, so once frames arrived faster than decompression completed, the continuation read a later frame's data — delta updates silently dropped and others delivered twice, corrupting downstream price state. The payload is now captured synchronously.
• A listener could receive a frame after unsubscribing. Dispatch skipped the liveness check for a listener unsubscribed by an earlier listener mid-dispatch. 0.1.3 matched EventTarget here; the behaviour is restored.
• createL1ActionHash validates vaultAddress again. It is publicly exported and was taking the address with no runtime check — a 0x${string} type constrains neither charset nor length at runtime, and nothing at all for JavaScript callers. A malformed address silently hashed to different bytes than the caller described, and a 32-byte address hashed identically to its 20-byte truncation. Both now throw, as they did at 0.1.3.

Behavioural change

HyperliquidEventTarget is no longer a subclass of EventTarget. Dispatch is now a hand-rolled listener map, which is where most of the per-frame win above comes from. addEventListener / removeEventListener keep their semantics — including duplicate-registration dedup, once, AbortSignal, and not delivering to a listener removed mid-dispatch — but instanceof EventTarget is now false and dispatchEvent is not available. Reaching for either was never part of the documented surface; if you depended on it, that access needs replacing. Flagged here because it landed in a patch release.

Also

New perf scenarios cover the fastAssetCtxs decode path and the user-account channels, neither of which was measured before.