Coordinator and worker

Use a coordinator to persist intent and dispatch work, and a Node or Bun worker to transfer files. The queue, database, authentication and artifact store are supplied by your application. Convex and Vercel Fluid can fill these roles, but neither is required.

Send references, not bytesCopy Anchor Link

An application-owned job could look like this:

export interface DeployJob {
  publishId: string
  appId: string
  artifactId: string
  manifestKey: string
  operationId: string
}

The worker loads the manifest from authorized storage and resolves its opaque source references. Keep tokens and signed URLs out of the queue message. Validate the job against the stored publish intent before choosing credentials or a destination.

Coordinate through the control planeCopy Anchor Link

import { makeVercelClient, makeVercelControl } from "@deploykit/vercel/control"

export const createControl = (token: string, teamId: string) =>
  makeVercelControl(makeVercelClient({ token, teamId }))

This exposes Effect operations for project/status/access/activation work without the transfer implementation. The control entry point is checked as a browser-compatible bundle; keep it on the trusted server because it uses provider credentials.

Save the publish intent and schedule delivery using your database/queue's transaction or outbox mechanism. A database insert followed by an unrelated queue send is not an atomic operation. The workflow guide explains which checkpoints belong to which process.

Run the transferCopy Anchor Link

Use the complete stored artifact example in the worker. For sustained workloads, move its client construction outside the function and reuse that client. This shares memory, staging and upload-throttling budgets across jobs. Close it during process shutdown.

Save the deployment receipt before polling or checking. For Vercel, request production plus activation: "deferred" when verification must happen before activation. Pages has different activation semantics; do not assume changing the adapter preserves that behavior.

Decide how to acknowledge a jobCopy Anchor Link

A lease can prevent concurrent workers from processing the same intent, but lease expiry does not prove that a remote create failed. Operation IDs support discovery, not provider deduplication. See recovery for the checked reconciliation example.

Transport failures between processesCopy Anchor Link

Provider.encodeFailure and Provider.decodeFailure preserve supported tagged SDK failures in versioned JSON. They do not serialize every possible application error, Effect defect or interruption. Use your own envelope to distinguish those outcomes.

Keep detailed failure records in access-controlled storage. Forward allowlisted structured fields to telemetry, and keep queue acknowledgement independent of telemetry delivery.