vercel-deploy

---
name: vercel-deploy
description: 'Ship a Next.js 16 App Router project to production on Vercel end-to-end: preview → smoke test → a *staged* production deployment (`vercel --prod --skip-domain`) that serves no traffic yet → `vercel promote` → domains and DNS, plus a rollback runbook. Requires `stack.framework` in {`next`, `monorepo`} and `phase >= feature_complete`; sets `phase = "deployed"` on success. Idempotent. Triggers on: "deploy", "ship it", "go live", "push to production", "manda in produzione", "attacca il dominio", "rollback the deploy". Not for: the Vercel project config — `vercel.json`, region, env-var matrix, monorepo root directory are `module-add deploy` (run it first); the pre-deploy gates (`vercel-doctor`, `compliance-audit`, `shadscan`, `launch-audit`, `app-security`); CI and git hooks (`module-add ci`); an Expo/RN app (`rn-eas-deploy`); an eve agent (`eve deploy`).'
---

# vercel-deploy — ship the web app to production on Vercel

The web counterpart of `rn-eas-deploy`. It is the **only** skill that moves a web project to `phase = "deployed"`.

> **Versions checked 2026-10-22**: Vercel CLI `65.0.0` (published 2026-10-20). Docs: <https://vercel.com/docs/cli/deploy>, <https://vercel.com/docs/cli/deploying-from-cli>, <https://vercel.com/docs/cli/promote>, <https://vercel.com/docs/cli/rollback>, <https://vercel.com/docs/deployments/promoting-a-deployment>, <https://vercel.com/docs/instant-rollback>, <https://vercel.com/docs/git>, <https://vercel.com/docs/domains/working-with-domains/add-a-domain>.

> **What the re-check on 2026-10-22 found.** Seven major versions and 72 releases passed between
> `58.5.1` and `65.0.0`, so this was read from the published package rather than the docs. The
> mechanism this skill rests on is **unchanged and not deprecated**: `--skip-domain` still reads
> *"Disable the automatic promotion (aliasing) of the relevant domains to a new production
> deployment. You can use `vc promote` to complete the domain-assignment process."* `promote`,
> `rollback` (with `rollback status` and a 3m default timeout), `alias`, `domains` and `ls` are all
> still there.
>
> **One thing is genuinely new and belongs in this story: `vercel rolling-release` (`rr`)** —
> *"Rolling releases gradually shift traffic to a new deployment in stages, allowing you to monitor
> for errors before serving all traffic"*, with `enable`, `configure`, `approve`, `abort`,
> `complete` and `fetch`. It did not exist at `58.5.1`. It is a **stronger** version of what the
> staged deploy below does by hand: `--skip-domain` + `promote` is all-or-nothing at the moment of
> promotion, while a rolling release moves traffic in stages and can be aborted mid-way.
> **This skill still ships the staged flow**, because it needs no project-level configuration and
> works on every plan; where a project has rolling releases enabled, say so and use them instead of
> promoting in one step. See <https://vercel.com/docs/rolling-releases>.

## Scope — read this first

This skill **ships**. It does not configure.

| Concern | Owner |
|---|---|
| `vercel.json`, function region, env-var matrix per environment, monorepo Root Directory, project linking | **`module-add deploy`** — run it first |
| husky, lint-staged, GitHub Actions | **`module-add ci`** |
| cost/perf gate before shipping | **`vercel-doctor`** |
| legal/privacy gate before shipping | **`compliance-audit`** |
| UI-quality / accessibility gate before shipping | **`shadscan`** |
| launch-readiness gate before shipping | **`launch-audit`** |
| code-security gate before shipping (dependencies, secrets, authorization) | **`app-security`** |
| preview → smoke → staged production → promote, domains + DNS, rollback, `phase = "deployed"` | **this skill** |

If `vercel.json` or `.vercel/project.json` is missing, do not improvise them here — say so and route to `module-add deploy`. Two skills writing the same config is how the region silently changes between runs.

## Contract

See `references/contracts.md`. Key facts:
- Reads `<project-root>/.workflow/meta.json#stack.framework` — must be `"next"` or `"monorepo"`.
- Requires `meta.json#phase >= "feature_complete"`.
- Sets `meta.json#stack.deploy = "vercel"` and `phase = "deployed"` **only after** the production domain actually serves the deployment.
- Records `stack_config.production_url` + `stack_config.vercel_project` and appends `history`.
- Idempotent: re-running on an already-deployed project becomes a *release* run (deploy → promote), and skips domain setup that is already verified.

## When this skill applies

- Phase is `feature_complete` (or `deployed`, for subsequent releases).
- The user says: "deploy", "ship it", "go live", "manda in produzione".
- The orchestrator routes here from `dev-flow`.

## Monorepo awareness

If `meta.json#stack.framework === "monorepo"`, the web app lives at `apps/web/` and the Vercel project's **Root Directory** is already set to `apps/web` by `module-add deploy`. Consequence: **run the CLI from the repository root**, not from `apps/web/` — with Root Directory configured the CLI is already scoped. `vercel.json` stays in `apps/web/`.

The mobile side ships separately via `rn-eas-deploy`, the agent side via `eve deploy`. All three can be at different points; `phase = "deployed"` for the project means every side that exists has shipped — say which sides remain.

---

## Workflow

### Step 1 — Preconditions

Read `.workflow/meta.json`. Abort with a specific message if:

- `stack.framework === "expo-rn"` → this is `rn-eas-deploy`'s job.
- `stack.framework` is neither `next` nor `monorepo` → refuse; alternative deploy targets are a stack decision, not a flag (see `module-add/references/module-deploy.md` § *Alternative deploy targets*).
- `phase < "feature_complete"` → the build isn't done; say which phase it is and what closes the gap.
- `stack.deploy !== "vercel"` or `<project-root>/.vercel/project.json` is missing → run `module-add deploy` first.

Then confirm the CLI can see the project:

```bash
pnpm dlx vercel@latest whoami            # who the CLI is logged in as
pnpm dlx vercel@latest project inspect   # details of the *linked* project
```

Never `vercel link` from this skill. Re-linking silently repoints the directory at a different project — that is how staging code reaches production.

### Step 2 — Confirm the five gates ran (propose, never block)

Read `meta.json#compliance` (legal), `meta.json#vercel_doctor` (cost/perf), `meta.json#shadscan` (UI quality + a11y), `meta.json#launch_audit` (commercial readiness: metadata, legal and identity pages, a reachable CTA) and `meta.json#security` (can the code be attacked: dependencies, secrets, authorization).

- Missing → tell the user which gate has never run and offer to run it now.
- Present but older than the last material change (compare against `history`) → offer a re-run.
- Present and current → state the findings summary in one line and move on.

No gate blocks a deploy. Report and let the user decide — that is the existing policy for all five, and this skill does not tighten it.

#### Also check observability (propose, never block)

Read `meta.json#stack.observability`. Empty: say that nobody will know when production breaks, and offer `module-add observability`, errors first. Present: run `python3 <skills>/module-add/scripts/observability_check.py <root>` and state its findings in one line (a `found` finding is named, not hidden). Either way the deploy goes on: the user decides.

#### Also check the Next.js security floor

The five gates read the *code*. `app-security`'s dependency scan sees `next` among the packages, but it reports advisories, not which support line you are on or the floor below — that check is here, and `app-security` points to it. Since 2026-07 that is its own risk: Next.js ships **pre-announced security releases roughly monthly**, with two maintained lines whose labels **move**: as of 2026-10-03 they are still **16.3.x Active LTS** and **15.5.x Maintenance LTS**, and both took the 2026-09-30 patches.

```bash
pnpm ls next --depth 0        # or: node -p "require('next/package.json').version"
```

Compare against the floor in `references/contracts.md` § `nextjs_version` (**`16.3.8`** Active LTS or **`15.5.27`** Maintenance LTS, as of 2026-10-03). Below it, say so in one line naming the CVEs that apply to *this* project. The sets, newest first: **2026-09-30** is eight CVEs, of which the one to read first is the **high SSRF in Image Optimization** (conditional on `images.remotePatterns` existing at all) and the one most likely to be ours is the medium **`'use cache'` root-param cache leak** — with `cacheComponents: true` it can serve a page cached for one `[locale]` under another, has **no workaround**, and sits exactly on golden rule ② plus `next/root-params`. **2026-09-22** is a **critical RCE in `next/og` ImageResponse** (`>= 16.2.0 < 16.3.6`). The August set is two critical unauthenticated RCEs (attacker-supplied **AVIF**, which the patch disables outright; **Windows** hosts mixing Pages and App Router), and the July set includes a DoS through **any** Server Action and an SSRF through request-derived `rewrites()`, so most projects we scaffold are in scope of something.

Two things not to get wrong:

- **Move to the current patch of a line that is still patched — and check which line that is.** In July, `16.2.x` was Active LTS and telling someone to jump to 16.3 was wrong. In August the label moved: 16.2 stopped at `16.2.12`, the criticals were fixed in `16.3.3`, and a project still on 16.2 is now unpatched. Read the line labels in the latest release post; never carry them from memory.
- **Hosting on Vercel is not a patch.** Vercel has deployed global WAF rules for past Next.js CVEs, and said in the same breath that they "should not be considered a complete substitute for upgrading". Report the floor for Vercel-hosted and self-hosted projects alike.

Then re-check the floor itself before quoting a number — it moves monthly. **Checked 2026-10-03: the
floor moved.** `16.3.8` is both the newest `16.3.x` and the newest 16.x stable (it is `latest`), and
`15.5.27` is the newest `15.5.x`. The previous record, `16.3.3`/`15.5.24`, had stood since 2026-08-25
through five releases and nine CVEs — **which is the argument for reading the advisory database and not
the version number**: `16.3.4` → `16.3.8` looks like ordinary patch drift and is not. The command for
that is `gh api repos/vercel/next.js/security-advisories`, and it answers from this session where
`nextjs.org` does not.

⚠️ **Several of the September advisories give their patched version as a literal `16.3.?`.** Do not
parse it and do not treat an unparseable range as "not affected" — take the floor from the newest
patch on the line instead.

**`npm view next versions --json` settles this in one command**, and settles it better than the blog:
the blog tells you what was announced, the registry tells you what you can actually install today.
Read the blog for *which line is still patched* — that label moves and the registry cannot tell you —
then take the patch number from npm. `[VERIFY]` against <https://nextjs.org/blog>. Same rule as every
gate here — report, don't block.

### Step 3 — Reproduce the platform build locally

```bash
pnpm build                                   # the same build Vercel will run
pnpm dlx vercel@latest env ls                # every expected var, in every expected environment
```

Two things to actually check in the `env ls` output, not just print:

- Every var the app reads exists in **Preview**, not only Production. A missing preview var is the single most common cause of "works locally and in prod, PR preview 500s".
- Vars that must differ per environment actually differ — `DATABASE_URL` above all. See `references/deploy-checklist.md`.

If `pnpm build` fails, stop. Do not deploy a build you could not reproduce.

### Step 4 — First deploy or subsequent release?

```bash
pnpm dlx vercel@latest ls --prod
```

This decides the path, and the difference is documented behaviour, not a preference:

> *"The first deployment of a new project is always a production deployment, even when you run `vercel` without `--prod`."*

- **No production deployment yet → path A (first deploy).** There is no way to make the first deployment a preview. Tell the user plainly: the first deploy goes to production, so smoke-testing happens *after* it exists and *before* the custom domain points at it. Attach domains in Step 9, after the deployment is verified — a `*.vercel.app` URL nobody knows is a safe place to be wrong.
- **A production deployment exists → path B (release).** Steps 5–8 in full.

### Step 5 — Preview deploy + smoke test (path B only)

```bash
pnpm dlx vercel@latest deploy
```

`stdout` is always the deployment URL — capture it. Then smoke-test the actual flows the release touches (auth, the changed route, one write path), not just "the homepage 200s".

If the project has **Deployment Protection** enabled — Standard Protection is the Hobby default and protects preview URLs — the URL asks for a Vercel login before it renders. That is expected; log in rather than "fixing" it. Note that with Standard Protection the generated production URL is restricted too, so anything reading `VERCEL_URL` / `NEXT_PUBLIC_VERCEL_URL` for a fetch must target the requested domain instead.

Do not proceed until the user confirms the preview is good.

### Step 6 — Staged production deployment

```bash
pnpm dlx vercel@latest deploy --prod --skip-domain
```

This creates a **production** deployment — real production env vars, real production build — that is **not** assigned to the custom domains. Production traffic keeps going to the current deployment.

`--skip-domain` must be used with `--prod`, and it overrides the project's **Auto-assign Custom Production Domains** setting for this deployment.

This is the safe shape and the one the docs point at: `vercel alias` explicitly is *not* the recommended way to put a deployment on a production domain — `--prod --skip-domain` → `promote` → `rollback` is.

### Step 7 — Smoke-test the staged deployment

The staged URL runs with production environment variables against production data. Re-check the same flows as Step 5, plus anything that behaves differently in production: payment webhooks against live keys, transactional email, anything branching on `VERCEL_ENV === "production"`.

This is the last cheap moment. After Step 8 the fix is a rollback.

### Step 8 — Promote to Current

```bash
pnpm dlx vercel@latest promote <staged-deployment-url>
```

The promotion assigns the production domains **without a rebuild**. Default timeout is `3m`; `--timeout=0` returns immediately and lets it finish server-side.

Verify it took:

```bash
pnpm dlx vercel@latest promote status
pnpm dlx vercel@latest ls --prod
```

A deployment that has been promoted once **cannot be promoted again** — to go back to it later you roll back, not promote.

### Step 9 — Custom domains (idempotent)

Skip entirely if `vercel domains verify` already reports the project's domains as configured.

Otherwise follow `references/domains-dns.md`. The short version:

```bash
pnpm dlx vercel@latest domains add <domain> <project>
pnpm dlx vercel@latest domains verify <domain> --project <project> --format=json
```

The DNS records themselves are added at the registrar, by the user — this skill reads the required records out of `domains verify` / the dashboard and hands them over. It never invents record values: the apex is an **A** record and each subdomain has a **project-specific CNAME target**, so a hardcoded `*.vercel-dns-*.com` value is always wrong for someone.

Vercel recommends `www` as the primary domain with a redirect from the apex, because the DNS spec forbids CNAME on an apex and a CNAME gives the CDN more control over incoming traffic. Say so, then do what the user decides — an apex-primary setup with an A record is explicitly supported.

Once a domain is configured it is **automatically applied to the latest production deployment**.

### Step 10 — Post-deploy verification + rollback runbook

```bash
curl -sS -o /dev/null -w '%{http_code}\n' https://<production-domain>/
pnpm dlx vercel@latest inspect <deployment-url>
```

Then hand the user the rollback runbook from `references/rollback-runbook.md` — in the message, not just as a file path. The single fact that must be said out loud:

> After a rollback, Vercel turns **off** auto-assignment of production domains. Pushes to the production branch stop going live until you `vercel promote` a deployment.

Someone who rolls back at 2am without knowing that spends the next hour wondering why their fix isn't deploying.

### Step 11 — Update meta.json + commit

```json
{
  "stack": { "deploy": "vercel" },
  "stack_config": {
    "vercel_project": "<project-name>",
    "production_url": "https://<production-domain>",
    "production_region": "<fra1|iad1|…>"
  },
  "phase": "deployed",
  "history": [
    { "skill": "vercel-deploy", "ran_at": "<iso>", "deployment": "<url>", "domains": ["…"] }
  ]
}
```

Set `phase = "deployed"` **only after Step 10 confirms the production domain serves the new deployment**. A promote that timed out client-side is not a deploy.

Commit: `release: deploy <version> to production on Vercel`.

### Step 12 — Print the next-steps summary

- Production URL, and the deployment id now serving it.
- Which gates ran, and any finding the user chose to ship with.
- The rollback command, verbatim.
- For a monorepo: which sides have shipped and which have not (`rn-eas-deploy` for mobile, `eve deploy` for the agent).
- The maintenance loop: re-run `compliance-audit` + `vercel-doctor` + `shadscan` + `launch-audit` + `app-security` after material changes.

---

## Common anti-patterns (NEVER do)

- ❌ `vercel --prod` straight at a project that already serves traffic. Stage it, smoke it, promote it.
- ❌ `vercel alias` to put a deployment on the production domain. The docs name `promote`/`rollback` as the preferred production commands.
- ❌ `--prebuilt` for a Next.js app whose build reads System Environment Variables — they are missing at build time with that flag.
- ❌ `vercel link` from this skill. Linking belongs to `module-add deploy`.
- ❌ `phase = "deployed"` before the production domain actually serves the deployment.
- ❌ Assuming a rollback restores environment variables. It does not rebuild: env vars and cron jobs revert to the state of the rolled-back deployment.
- ❌ Leaving a project in a rolled-back state without telling the user that pushes to the production branch no longer go live.
- ❌ Hardcoding a `*.vercel-dns-*.com` CNAME target. It is per project.
- ❌ Deploying an app whose only route is the scaffold's default page. A green checkmark on an empty app proves nothing.

## References

- `references/deploy-checklist.md` — the pre-flight list: gates, build, env-var matrix, first-deploy vs release.
- `references/domains-dns.md` — apex vs `www`, A/CNAME records, verification, redirects, when `alias` is still right.
- `references/rollback-runbook.md` — instant rollback, plan limits, the auto-assign trap, undoing a rollback.
- `module-add/references/module-deploy.md` — the configuration this skill assumes exists.

## Updating meta.json (recommended pattern)

When this skill modifies state (artifact written, phase advanced, history appended), use the canonical script when available:

```bash
# Wherever dev-flow is installed (e.g. ~/.claude/skills/dev-flow/), invoke:
python3 .../dev-flow/scripts/update_meta.py <project-root> set-phase deployed
python3 .../dev-flow/scripts/update_meta.py <project-root> append-history \
    --skill 'vercel-deploy' --inputs '{...}' --outputs '{...}' --phase-after deployed
```

The script enforces phase monotonicity, normalizes legacy kebab-case aliases, and writes the canonical sha256 + timestamp into `meta.json#artifacts`. **Fall back to direct JSON editing only if the script is not on PATH** (and warn the user).

## Sources

- Vercel CLI: <https://vercel.com/docs/cli/deploy>, <https://vercel.com/docs/cli/deploying-from-cli>, <https://vercel.com/docs/cli/promote>, <https://vercel.com/docs/cli/rollback>, <https://vercel.com/docs/cli/alias>, <https://vercel.com/docs/cli/domains>, <https://vercel.com/docs/cli/list>, <https://vercel.com/docs/cli/inspect>
- Platform: <https://vercel.com/docs/deployments/promoting-a-deployment>, <https://vercel.com/docs/instant-rollback>, <https://vercel.com/docs/git>, <https://vercel.com/docs/deployment-protection>
- Domains: <https://vercel.com/docs/domains/working-with-domains/add-a-domain>, <https://vercel.com/docs/domains/working-with-domains/deploying-and-redirecting>

• SKILL.md 20 KB
• references/contracts.md 85 KB
• references/deploy-checklist.md 8 KB
• references/domains-dns.md 5 KB
• references/rollback-runbook.md 6 KB