Agent recipe: delayed & scheduled work

These are operational instructions for AI agents asked to run work later — deferred, retried or recurring — in a codebase that already uses @imqueue. A human-oriented walkthrough of the same material is at /blog/scheduled-work-without-a-job-system/.

When to apply this recipe

Apply it when the user asks for work that runs later over machinery they already have:

• "send / notify / charge / clean up in N minutes or hours";
• "retry this failed call later", with backoff;
• "run this every N minutes" as a sweep or reconciliation;
• "do we need BullMQ / a job queue / cron for this?" — check this recipe's decision table before adding a dependency.

Do not apply it when the requirement is declarative scheduling (cron strings, timezones, calendar dates), operator control of the queue (priorities, attempt caps, dead-lettering, pause/resume, a dashboard), or a multi-step resumable workflow. Those need a real scheduler, a dedicated job system, or an orchestration engine — say so instead of approximating them.

Choose the mechanism first

Facts these recipes rely on

Verified against @imqueue/core, @imqueue/rpc and @imqueue/job sources and by measurement against a live Redis.

The call contract. Generated clients take two optional trailing parameters, in this order: imqMetadata?: IMQMetadata, then imqDelay?: IMQDelay. The delay is always last. IMQDelay(timer, unit) accepts a unit of 'ms' | 's' | 'm' | 'h' | 'd', defaulting to 'ms'.

Those two parameters are stripped from the request by identity (instanceof), not by position. All of the following are measured; do not guess between them:

• method(data, undefined, delay) — emit this form on @imqueue/rpc >= 3.4.0. A trailing undefined on a delayed call is a placeholder and is never delivered, whether or not metadata is also passed, and however many trailing placeholders there are. So method(a, undefined, undefined, delay) sends [a], and a skipped optional declared param falls back to its default.
• method(data, new IMQMetadata({ ... }), delay) compiles and runs on every version. Emit this form when the installed version is <= 3.3.0 or unknown, because there the placeholder survives into the request as a real argument (serialized null) and a method whose params are all required rejects the call with IMQ_RPC_INVALID_ARGS_COUNT.
• On 3.3.1 only, the rule is narrower: one placeholder is dropped, and only when no metadata is passed. Treat 3.3.1 as "prefer the bag" too.
• method(data, delay) runs, but fails type-check with TS2345 (IMQDelay is not assignable to IMQMetadata) on every version. Do not emit it, and do not reach for as any to silence it — on 3.3.1 that cast changes what a skipped optional param delivers.
• Nothing is dropped when there is no delay, on any version: method(a, undefined) delivers null, so a default does not fire.
• The service-side arity check is declared === received when every declared parameter is required, and declared >= received when at least one is optional. So a method with an optional parameter accepts a call that omits trailing arguments, and a surviving placeholder that lands in an optional slot passes the check instead of being rejected. A count above the declared total is rejected either way. A passing call therefore proves nothing about a method with a different signature — do not generalise from it.

Caller-side. callTimeout has no default, and unset means wait forever; a delay extends its budget rather than firing early. The pending promise's resolver lives only in the caller's memory, so never await a long delayed call inside a request handler — a restart loses the resolver while the reply still arrives.

Enqueue is not a durability confirmation. send(toQueue, message, delay?, errorHandler?) takes milliseconds and resolves with a locally generated UUID before the write is confirmed — pass errorHandler to learn about failures. push(job, { delay }) returns synchronously and takes no error handler, so a failed enqueue only reaches the queue's logger.

@imqueue/job handler return contract — get this exactly right:

Catch handler errors explicitly and return a delay; do not rely on a throw. There is no declarative attempts policy and no dead-letter destination — carry the attempt counter in the job payload and write your own park-for-review table.

Delivery mode. safeDelivery defaults to false in @imqueue/core and @imqueue/rpc, and to true through @imqueue/job (whose safeLockTtl maps to safeDeliveryTtl). The lease covers the hand-off only, so a process killed mid-handler still loses that attempt. At-least-once is the guarantee either way: make deferred handlers re-runnable. See /blog/guaranteed-message-delivery-cost/.

Promotion path. The producer parks the packed message in the sorted set <prefix>:<queue>:delayed, scored Date.now() + delay, plus an empty companion key <prefix>:<queue>:<id>:ttl set with PX <delay> NX. That key's expiry fires a keyspace notification, and one elected watcher moves everything now due onto the ready list; failing that, workers sweep every watcherCheckDelay (default 5000 ms). Default prefixes: imq for core/rpc, imq-job for jobs. Prompt promotion needs notify-keyspace-events to include Ex.

Accuracy is "no earlier than." Pass whole integer milliseconds: a fractional delay fails to set the alarm key and waits for the next sweep, and an IMQDelay carrying an unrecognised unit string yields no delay at all. The due time comes from the sending process's clock and is compared against the sweeping one.

Absent features. There is no cron, repeat or recurrence primitive anywhere in the ecosystem. There is also no handle for a single scheduled message: nothing cancels, reschedules or inspects one, and the only removal path is wholesale. If an action can be revoked, gate the handler with a state check rather than trying to unschedule it.

Recipe: defer a call the service already exposes

import { IMQDelay, IMQMetadata } from '@imqueue/rpc';

const notifications = new NotificationClient({ callTimeout: 30_000 });

await notifications.start();

// fire-and-forget: do not await a 24h call in a request handler
notifications
    .sendTrialEndingEmail(
        { userId, plan },
        undefined,                                   // metadata slot — skipped
        new IMQDelay(24, 'h'),                       // delay is always last
    )
    .catch(err => logger.error('deferred call failed to enqueue', err));

The service method needs no change: a delayed request carries only from, method, args and optional metadata, so validation, decorators and handlers behave exactly as for an immediate call.

Recipe: retry with backoff you control

import JobQueue from '@imqueue/job';

type Sync = { orderId: string; attempt: number };

new JobQueue<Sync>({ name: 'OrderSync' })
    .onPop(async job => {
        try {
            await warehouse.sync(job.orderId);
        } catch (err) {
            job.attempt += 1;                        // counter lives in the payload

            if (job.attempt > 5) {
                await parkForReview(job);            // your own dead-letter table

                return -1;                           // negative: stop
            }

            return 1000 * 2 ** (job.attempt - 1);    // 1s, 2s, 4s, 8s, 16s
        }
    })
    .start()
    .then(queue => queue.push({ orderId, attempt: 0 }))
    .catch(err => logger.error('OrderSync failed to start', err));

Recipe: recurring sweep

import JobQueue from '@imqueue/job';

const EVERY_MINUTE = 60_000;
const sweeper = new JobQueue<Record<string, never>>({ name: 'ExpireHolds' });

sweeper.onPop(async () => {
    await releaseExpiredHolds();

    return EVERY_MINUTE;    // re-arm; next tick is scheduled after the work
});

await sweeper.start();

Seeding rules — the part that breaks in production:

1. Never seed on process boot unguarded. Every process that pushes a seed starts its own chain, so N replicas rolling through a deploy leave N overlapping sweeps forever.
2. Seed from a one-off task, a migration step, or the scheduler the user already runs.
3. If it must happen at boot, gate it on a SET NX lease whose TTL spans several periods, retried on an interval, so a deploy cannot double-seed and a dead chain re-seeds when the lease lapses.
4. The period drifts: the next fire time is computed after the work finishes, so the real interval is work plus delay. Do not promise wall-clock alignment.
5. A chain is not a schedule. If a re-arm ever fails, the recurrence ends silently — instrument the age of the last completed run and alert on it.

Verify

# prompt promotion requires keyspace expiry events (look for E and x)
redis-cli CONFIG GET notify-keyspace-events

# what is parked right now, and when it is due (scores are due-time in ms)
redis-cli --scan --pattern 'imq:*:delayed'
redis-cli ZRANGE imq:<ServiceName>:delayed 0 -1 WITHSCORES
redis-cli ZCARD imq-job:<JobName>:delayed

# the alarm keys that trigger promotion
redis-cli --scan --pattern 'imq:<ServiceName>:*:ttl'

Then confirm behaviour, not just configuration: schedule a short delay and measure the actual arrival (expect "no earlier than", plus up to watcherCheckDelay without notifications); kill a worker mid-handler and check the work is either re-run or recorded as lost, according to the delivery mode you chose; and run the handler twice to prove it is idempotent.

Failure modes