llm-coalesce · step 1 demochecking provider…

One call, many widgets

Every widget below wants the same answer at the same moment. The Node server behind this page is deliberately dumb — it has no idea what coalescing is; every request it receives triggers a real provider call. The coalescing happens entirely in this page's JavaScript, imported from the real llm-coalesce npm package (not a copy of it): on the left, each widget calls this demo's own createSSEStream() helper on its own; on the right, widgets go through the package's coalescer.stream(), and only the first one actually reaches the server. llm-coalesce itself never parses SSE or talks to any provider — it only ever sees a plain async stream, supplied by the caller. Watch the provider-call counters — they count actual stream responses received by this page.

WIDGETS

Stagger mount times (join mid-stream)

Without coalescing

each widget calls fetch() on its own

stream responses received

tokens generated

Press "Run demo" to start.

With llm-coalesce

widgets share one client-side registry entry

stream responses received

tokens generated

Press "Run demo" to start.

Turn on stagger to mount widgets 0ms / 600ms / 1300ms apart — on the right, a widget joining mid-stream replays the buffered tokens instantly from the browser's own buffer, then continues live, without a second request ever reaching the server. The "mount one more widget after the stream finishes" button is left in deliberately: llm-coalesce's registry only shares requests that overlap in time, so a widget arriving after completion makes a fresh call on both sides — that's a scoped-out feature (a post-completion TTL cache), not a bug; see the package's own "Non-goals" and ADR 0003. Set ANTHROPIC_API_KEY in the server's environment to swap the mock for a real streaming completion — see the README. That choice is independent of everything above: coalescing decides how many requests get sent, not what answers them.

Everything above runs in one script, one module graph — the coalescing was never in question across a bundle boundary. See the microfrontends demo → for the harder case: three genuinely separate module graphs, coordinated only through windowAdapter().