@inbrowser/resumable

@inbrowser/resumable is a generic engine for resumable streaming jobs. A producer yields typed events, a JobStore appends those events to an ordered log, and subscribers can tail the log again from a known offset after a network drop, route handoff, or page reload.

The package has no LLM knowledge. LLM inference is one consumer through @inbrowser/relay; the same engine can back any long-running job that reports incremental events.

What It Provides

• createJobEngine, which starts producers, appends their events, exposes snapshots, and returns a terminal marker when work finishes.
• A JobStore<TEvent> contract for durable backends.
• createMemoryJobStore for tests and local development.
• createRtdbJobStore for Firebase Realtime Database persistence through REST writes and RTDB SSE watches.
• Post-terminal TTL support through ttlMs, expiresAt, and sweepExpired.
• @inbrowser/resumable/testing probes for durability and TTL sweep behaviour.

Quick Start

import { createJobEngine } from '@inbrowser/resumable';
import { createMemoryJobStore } from '@inbrowser/resumable/memory';

type ChunkEvent = { kind: 'chunk'; text: string };

const engine = createJobEngine<ChunkEvent>({
  store: createMemoryJobStore<ChunkEvent>(),
});

const { jobId } = await engine.start(async function* () {
  yield { kind: 'chunk', text: 'hello ' };
  yield { kind: 'chunk', text: 'world' };
});

for await (const item of engine.subscribe(jobId)) {
  if (item.kind === 'event') {
    console.log(item.seq, item.value);
  }
  if (item.kind === 'terminal') {
    console.log(item.status);
  }
}

await engine.stop();

Use subscribe(jobId, { from }) to resume from an offset. Events before from are skipped, so a client that has consumed events 0 and 1 reconnects with from: 2.

Durable Store

Use RTDB when the event log must survive process restart or a subscriber reconnecting through another server instance:

import { createJobEngine } from '@inbrowser/resumable';
import {
  createRtdbJobStore,
  serviceAccountTokenProvider,
} from '@inbrowser/resumable/rtdb';

type ChunkEvent = { kind: 'chunk'; text: string };

const store = createRtdbJobStore<ChunkEvent>({
  url: process.env.RTDB_URL!,
  auth: serviceAccountTokenProvider({ keyFile: './service-account.json' }),
  rootPath: 'resumable_jobs',
  defaultTtlMs: 7 * 24 * 60 * 60 * 1000,
});

const engine = createJobEngine<ChunkEvent>({
  store,
  sweep: { intervalMs: 60 * 60 * 1000 },
});

The durable store preserves the event log. It does not automatically restart a producer if the process running that producer is killed.

For efficient RTDB sweeps, add an index on the store root path:

{
  "rules": {
    "resumable_jobs": {
      ".indexOn": ["expiresAt"]
    }
  }
}

Documentation

The documentation follows the Diataxis approach: each page serves one kind of user need.

• Tutorial: build a resumable stream - learn by doing with the memory store.
• How to use RTDB for durable jobs - wire the production store, sweeps, and probes.
• API reference - facts about exports, types, stores, and probes.
• How resumable jobs work - the design rationale and failure model.

Package Exports

• @inbrowser/resumable - createJobEngine plus core types.
• @inbrowser/resumable/memory - in-process JobStore.
• @inbrowser/resumable/rtdb - Firebase RTDB JobStore and token providers.
• @inbrowser/resumable/testing - durability and TTL probe helpers.

Relationship To @inbrowser/relay

@inbrowser/relay uses this package to keep an LLM generation running on the server while browsers disconnect and reconnect. This package intentionally stops below HTTP framework concerns and below any LLM provider protocol.