A batshit-style automatic batch scheduler for CrateStack's generated TypeScript RPC client (transport rpc schemas), shipped as an RpcLink (issue #182) rather than a fetch override — so it composes with @cratestack/link-logger, a retry link, or an auth-refresh link instead of clobbering them.

React Native compatibility evidence is tracked in docs/react-native-compatibility.md.

Multiple unary calls issued in the same tick collapse into a single POST /rpc/batch request instead of firing one POST /rpc/{op_id} each.

Usage

import { createBatchLink } from "@cratestack/link-batch";
import { CratestackRpcRuntime } from "./generated/runtime"; // your project's generated client

const runtime = new CratestackRpcRuntime("https://api.example.com", {
  links: [createBatchLink()],
});
const client = new MyGeneratedClient(runtime);

// These three calls, if issued in the same tick, become ONE /rpc/batch request:
const [a, b, c] = await Promise.all([
  client.widgets.get(1),
  client.widgets.get(2),
  client.widgets.get(3),
]);

Batching semantics

Window: defaults to queueMicrotask — calls fired synchronously in the same tick (e.g. inside Promise.all) collapse. Pass windowMs to widen the window across ticks.
Partitioning (fixed in #273; versions before this carried a bug here — see "Known limitations" below): each flush is split into partitions by transport envelope — headers (excluding Idempotency-Key, which is carried per-frame, not per-request), fetch reference, codec reference, and the resolved batch URL — and every partition is sent as its own POST /rpc/batch. Calls that share an envelope, the overwhelmingly common case, still collapse into exactly one request; calls that don't (e.g. two different Authorization headers) each keep their own, in a separate request, rather than one silently overwriting the other's.
Aggregate headers: pass headers to createBatchLink({ headers }) to declare headers that every synthesized /rpc/batch request carries, merged over each partition's own — a same-named per-call header is overridden by this value, not the other way around. Use this for service-level headers (e.g. an API key) rather than per-tenant auth, which should come from per-call headers that drive partitioning.
maxBatchSize is enforced per partition, not globally across the flush — a single oversized partition chunks into several concurrent requests; it never borrows headroom from a different partition.
Dedup: only calls sharing an explicit idempotencyKey are collapsed into one request frame by default — that's already the caller's own signal that the call is a safe repeat. Calls with no idempotency key are never auto-collapsed, since the server does no dedup of its own and silently merging two textually-identical but unmarked mutations would be unsafe. Pass a custom dedupe for full value-based collapsing, or () => null to disable dedup and only batch. Dedup runs within a partition — it never collapses calls across two different envelopes.
Correlation: results are fanned back out by matching each response frame's id, not array position — the server's /rpc/batch contract already guarantees order (see docs/design/rpc-transport.md §3.2), but id-based matching is strictly more robust and mirrors the repo's own Rust client-side batch debouncer (examples/rpc-batch-debounce).
Failure isolation: a partition that fails (network error, non-OK response) only rejects the callers queued in that partition — it never affects a concurrently in-flight partition.

Known limitations

Aborting an individual call's AbortSignal only cancels it if its batch hasn't been sent yet — it does not cancel an in-flight /rpc/batch request.
(Fixed in #273, kept here for anyone reading an older version's docs) Before #273, the synthesized /rpc/batch request reused the first queued call's headers/fetchFn/codec for the whole flush — per-call custom headers on later calls in the same window were silently dropped from the aggregate request instead of applied. This is why partitioning exists now.