Vercel deployment

Compatibility: Pactmark 0.1.x. Local adapter evidence exists — build, route, security and accessibility contracts, plus streamed events. No live Pactmark Vercel deployment has been verified. Live deployment remains unsupported until the readiness record contains URL, log, database-resume, rollback and teardown evidence.

The adapter is deliberately thin

app/api/agent/[...pactmark]/route.ts

import { nextVercelHandler } from "@/src/host";

export const runtime = "nodejs";
export const dynamic = "force-dynamic";
export const revalidate = 0;
export const fetchCache = "force-no-store";

export const GET = nextVercelHandler;
export const POST = nextVercelHandler;
export const OPTIONS = nextVercelHandler;

src/host.ts

import { createVercelRouteHandler } from "@pactmark/vercel";

export const nextVercelHandler = createVercelRouteHandler({
  basePath: "/api/agent",
  runtime: pactmarkRuntime,
  readEnvironment: () => ({
    PACTMARK_PROFILE: process.env.PACTMARK_PROFILE,
  }),
  authenticate: async (request) => resolvePrincipalAndTenant(request),
});

createVercelRouteHandler wraps createAgentFetchHandler and supplies the request-time environment, the abort signal, an optional waitUntil, and the runtime’s capabilities. That is the entire adapter — everything else is the portable kernel.

Use the Node.js runtime. Fluid Compute gives you full Node APIs, streaming and longer durations, so there is no reason to reach for an edge runtime and lose executor and store options.

Authentication comes first

async function resolvePrincipalAndTenant(request: Request) {
  const session = await verifySession(request); // your IdP, verified
  if (!session) return undefined;               // → 401 before any runtime access
  return {
    principal: { type: "user" as const, id: session.userId },
    tenant: { id: session.tenantId },
    authenticationStrength: session.strength,
    decisionRoles: session.roles,
  };
}

The tenant and principal are resolved before runtime access. A client cannot select them, and an unauthenticated request never reaches a store or a model.

Durability lives outside the invocation

This is the part that most deployments get wrong.

What the function does

Accept an authenticated command, append events, stream progress, return. It is a transport, not a lifecycle owner.

What the function must not own

Run state, continuation, background progress, or the assumption that it will still be alive when a human approves something.

For a production profile you need:

A hostname-verified TLS Postgres store

sslmode=verify-full, least-privilege role, ordered migrations applied out of band.

A separately operated worker

Vercel functions are not a place to run a durable wake-up loop. Run the worker where a long-lived process belongs — a container, a VM, a scheduled job with proper leasing.

Real background scheduling

waitUntil is for non-critical work such as flushing configured telemetry. It is not a durable wake-up mechanism and must not be used as one.

A long function timeout is not a durability guarantee. Memory state, background work inferred from an open response, or an unauthenticated preview cannot satisfy the production gate — and readiness will say so.

Streaming on Vercel

SSE works on the Node.js runtime with no special configuration. Set dynamic = "force-dynamic" and fetchCache = "force-no-store" so nothing is cached, and make sure your client reconnects with afterSequence — see Streaming and clients.

Preview deployments

A release preview must install the exact attested tarballs from a vendored frozen lockfile. It must not resolve Pactmark packages from the public registry before publication, and a preview with anonymous access is not evidence of anything.

The reference fixture supports an explicit preview profile with allowAnonymousDevelopment and a fixed anonymous authority — clearly separated from the production path, which requires a real authentication hook. Keep that separation in your own host; do not let a preview flag be readable in production configuration.

Environment variables

vercel env add PACTMARK_PROFILE production
vercel env add DATABASE_URL production
vercel env pull .env.local

Read environment variables at request time inside the adapter via readEnvironment. Do not capture them at module scope where they can be serialized into a build artefact.

What is actually evidenced today

Claim	Status
The route adapter builds and serves the command surface	Local contract evidence
Security and accessibility contracts pass	Local contract evidence
Streamed events work through the adapter	Local contract evidence
A live Pactmark Vercel deployment behaves durably	No current deployment

Historical Vercel evidence was test-only Preview work and its resources were removed.

Deployment overview

How this target compares with Node and Cloudflare.

Readiness checklist

What has to be true before this becomes a production system.

Getting started

Core concepts

Guides

Deployment

Production

Security

Vercel deployment

The adapter is deliberately thin

Authentication comes first

Durability lives outside the invocation

What the function does

What the function must not own

Streaming on Vercel

Preview deployments

Environment variables

What is actually evidenced today

Deployment overview

Readiness checklist

​The adapter is deliberately thin

​Authentication comes first

​Durability lives outside the invocation

What the function does

What the function must not own

​Streaming on Vercel

​Preview deployments

​Environment variables

​What is actually evidenced today

Deployment overview

Readiness checklist

The adapter is deliberately thin

Authentication comes first

Durability lives outside the invocation

Streaming on Vercel

Preview deployments

Environment variables

What is actually evidenced today