robopipe.io — website

Sales-oriented marketing site for Robopipe (industrial machine vision by KOALA42).

Stack: Next.js 16 (App Router) + Payload CMS 3 embedded in one app · PostgreSQL · Tailwind v4 · next-intl (cs/en) · Resend · Google Cloud Storage media. Deployed to Coolify as a Docker image.

Local development

docker compose up -d        # Postgres 17 on :5432
cp .env.example .env        # defaults work out of the box
pnpm install
pnpm payload migrate        # apply DB migrations
pnpm seed                   # demo content in cs+en, admin user
pnpm dev

• Site: http://localhost:3000 (/ redirects to the stored language choice, else by browser language, see below)
• Admin: http://localhost:3000/admin — seeded login admin@robopipe.io / admin

Optional env: RESEND_API_KEY (lead notification emails; logged to console when unset), GCS_BUCKET + GCS_CREDENTIALS (media storage; local ./media when unset).

Architecture notes

• Localization — content is field-localized in Payload (cs default, en), URLs are prefixed (/cs/*, /en/*), and slugs are localized (/cs/cenik ↔ /en/pricing). Section segments (/blog, /case-studies) are constant. hreflang alternates are emitted per document; the language switcher reads them from the head to land on the translated slug.
• Blog visibility — posts/case studies appear only in locales where they have a slug (queries run with fallbackLocale: false); marketing pages are expected to be fully translated.
• Pages are block-based (src/blocks/*): hero, feature grid, content, media, stats, logo cloud, testimonials, use-case cards, pricing table, FAQ accordion, case-study grid, blog teaser, CTA banner, contact form.
• Leads — the contact form posts a server action (src/actions/submitLead.ts): zod validation, honeypot + per-IP rate limit, stored in the leads collection (public REST create is closed), notification via Resend to site-settings.leadNotificationEmail.
• Revalidation — any published change triggers coarse site-wide ISR revalidation (src/hooks/revalidate.ts); pages also refresh every 10 min.
• Root language — / goes to the language stored in the NEXT_LOCALE cookie. Only an explicit choice writes it (footer switcher, locale suggestion). Without it, Accept-Language decides: cs and sk go to /cs, everything else to /en (src/proxy.ts). next-intl itself does not read or write the cookie (localeCookie: false).
• SEO and GEO (issue #2, docs/specs/2-seo-geo/) — buildMeta() (src/lib/meta.ts) gives every page a self canonical, hreflang with x-default (the English URL) and Open Graph. JSON-LD comes from one module (src/lib/jsonLd.ts): Organization and WebSite on every page, Product (offers read from the pricing table in Payload) on the home page, BlogPosting, FAQPage from the visible FAQs, BreadcrumbList. The sitemap and /llms.txt read one list of indexable documents (src/lib/indexable.ts). Plus robots, per-locale RSS (/{locale}/blog/rss.xml), the 301 map from the old Webflow and Shopify URLs in next.config.ts and an editor-managed redirects collection.
• Use cases — case studies are shown as "Případy užití" / "Use cases" at the same URLs: listing /{locale}/case-studies, detail /{locale}/case-studies/{slug} with problem, solution, result, technique and FAQ sections (fields on the collection, empty sections are skipped).
• Memberships — site-settings.memberships (e.g. Česká inovační platforma) feeds the footer, logo bands with source: memberships and Organization.memberOf.
• Speed on phones (issue #30, docs/specs/30-mobile-speed/) — only Space Grotesk and Barlow are preloaded. Inter (figures only) loads weight 600 when a page needs it. Media always gives next/image a sizes: the container width by default, exact values from src/lib/imageSizes.ts in the blocks. The hero poster is an image under the video. The hero block has an optional mobileVideo for screens up to 767 px, and on Save-Data no video plays. The Google tag script loads when the browser is idle after load.
• Draft preview — Payload live-preview + /next/preview?secret=… (draft mode), exit via /next/exit-preview.

Schema changes

Dev uses migrations too — after changing collections run:

pnpm payload migrate:create <name>
pnpm payload migrate
pnpm generate:types

Commit the generated migration and src/payload-types.ts.

Tests

pnpm test:int                                   # vitest, writes to the local DATABASE_URL
E2E_BASE_URL=http://localhost:3310 pnpm test:e2e  # playwright against a running, seeded server

Integration tests stop when DATABASE_URL is not localhost (vitest.setup.ts). E2E_BASE_URL defaults to http://localhost:3000. PLAYWRIGHT_CHROMIUM_PATH points Playwright at another Chromium binary if the downloaded one does not start.

Cookie policy script

pnpm seed:cookie-policy puts the cookie paragraphs (issues #21, #22, #26) on the published privacy page in an existing database, cs and en. It replaces the old "Analytics" paragraph and the sentence that names Vercel. The texts are in src/seed/cookiePolicy.ts, and pnpm seed uses the same texts. The script is idempotent and publishes the page. It takes --dry-run, --production and --remote=<URL> like seed:seo-geo. It exits with code 2 when it cannot find a paragraph to replace, so you edit that part in the admin.

SEED_REMOTE_EMAIL=… SEED_REMOTE_PASSWORD=… pnpm seed:cookie-policy --remote=https://robopipe.io --dry-run

Google Ads offline conversions

GET /api/google-ads/company-leads.csv is the source of the scheduled upload in Google Ads Data Manager (issue #26). It lists leads from the last 90 days with marketing consent and a company email, without test leads. One function decides this: isGoogleAdsLead in src/lib/googleAds.ts. The list of public mail domains is in src/lib/leadEmail.ts. The email goes out only as a SHA-256 hash. The conversion name comes from Site settings → Measurement → Lead conversion name.

The route needs HTTP Basic. The username and the password are in the admin under Site → Google Ads export (issue #28). The database keeps only a salted scrypt hash of the password, and the API never returns it. Only logged-in users can read or change the global. Without a username and a password, the route answers 401 to everybody. Setup steps in Google Ads are in docs/specs/26-offline-conversions/implementation.md.

curl -u "google-ads:<password>" http://localhost:3000/api/google-ads/company-leads.csv

SEO and GEO content script

The landing, task and about pages, the case study sections, FAQs, home page blocks, navigation, memberships, solutions, company facts, pricing FAQs, the vision guide changes and the auto annotation blog draft live as files in seed-assets/copy/seo-geo/. pnpm seed applies them to a fresh local database. pnpm seed:seo-geo applies them to an existing one:

pnpm seed:seo-geo --dry-run             # list planned changes only
pnpm seed:seo-geo                       # drafts only, published content and globals unchanged
pnpm seed:seo-geo --status=published    # publish pages, case studies, posts and globals

The script is idempotent: it finds existing documents and never overwrites a field that has a value. The blog post always stays a draft. It refuses a DATABASE_URL outside localhost unless you pass --production. Production steps are in docs/specs/2-seo-geo/implementation.md and docs/specs/6-seo-geo-stage2/implementation.md.

File	Content
landing-pages.json, landing-<key>.cs.md, landing-<key>.md	Solution, task and about pages. An existing page gets only the blocks it does not have yet (by block id). obsolete lists stage 1 slugs the script reports but never deletes. cardImage sets seo.image (the home page card and the share preview) only while it is empty.
faqs.json	All FAQs. Every key is used by a page, a case study or the pricing page.
settings.json	site-settings.solutions (page keys in display order) and site-settings.company. Written only while empty.
existing-pages.json	Additions to existing content: pricing FAQs, and the first H2, seo.title and one paragraph of the vision guide (changed only from the exact current value).
home.json, navigation.json	Home page blocks, header label, footer about link, the stage 1 footer column to remove. removeBlocks lists home blocks (by blockName) the script removes on every run, also after a restore from the version history. featureLinks adds a link at the end of a featureGrid card text, only in a locale where the link label is empty. When the card has a label in another locale, the script keeps its target page and writes only the label. A link label left empty in English hides the link on /en (no Czech fallback). The script sets the home page order (hero, customer logos, feature grid, process steps, industries, testimonial, solutions, memberships, blog and FAQ, CTA) only from a known state: production after #6, after #9 round 1, or either without the memberships strip. Any other order was changed in the admin and stays. A missing memberships strip then goes after the first testimonialBar.

Copy markers. A document whose text still contains [doplnit Filip] or [ověřit] is never published. It stays a draft and the report says draft (markers): <document>.

Through the REST API. --remote=<URL> runs the same script against a running site through the Payload REST API, without database access. The account comes from SEED_REMOTE_EMAIL and SEED_REMOTE_PASSWORD (an editor role is enough). --dry-run is read-only: the client refuses every write before it is sent. An HTTP error stops the run with the method, path, status and error text.

SEED_REMOTE_EMAIL=… SEED_REMOTE_PASSWORD=… pnpm seed:seo-geo --remote=http://localhost:3000 --dry-run

Search engine verification

The Bing (msvalidate.01) and Seznam (seznam-wmt) verification tags come from Site settings → Verification. An empty value renders no tag. A change through the admin or POST /api/globals/site-settings (Payload uses POST for globals) shows on every page without a deploy. The content script never writes these values.

AI-generated blog posts

Editors queue article ideas in Blog → Blog topics (short description + optional per-article instructions). A nightly Coolify scheduled task (03:00 UTC) hits /api/cron/run-jobs (authorized via CRON_SECRET), which runs the generate-blog-post workflow (src/jobs/):

1. Claude (claude-fable-5, web search enabled, server-side refusal fallback to Opus 4.8) writes the English article as structured JSON, picking a category.
2. A second call adapts it into Czech (not a literal translation).
3. A hero photo is fetched from Pexels and uploaded to Media with localized alt text.
4. A bilingual draft post is assembled (markdown → Lexical via convertMarkdownToLexical); the topic's target date prefills publishedAt; a "draft ready" email goes out via Resend.

Publishing is always manual — review the draft and hit Publish. Regenerate by setting the topic's status back to Queued (content is overwritten; old text stays in version history), or skip the nightly wait entirely with the Generate now button in the topic sidebar (POST /api/blog-topics/:id/generate — responds immediately and runs the job via after(); the button polls until the draft is ready). The base writing prompt lives in the Blog AI global, editable without a deploy. Steps are resumable inline tasks, so a failed/timed-out run resumes where it stopped on the next cron hit.

Required env vars: ANTHROPIC_API_KEY (the Anthropic org must allow ≥30-day data retention — Fable requirement), PEXELS_API_KEY, CRON_SECRET. To run the queue locally: create a topic, then curl -H "Authorization: Bearer $CRON_SECRET" "http://localhost:3000/api/payload-jobs/run?limit=10".

Deployment (Coolify)

The app ships as a Docker image (Dockerfile, Next.js standalone output) deployed with Coolify's Dockerfile build pack. The build runs pnpm payload migrate && pnpm build against the production database, so each deployment migrates the DB before next build prerenders against it.

1. Database — a PostgreSQL resource from Coolify's template on the same server/destination as the app. Use its internal URL for DATABASE_URL: Coolify builds with --add-host entries for the destination network's containers, so the internal hostname also resolves during the image build.

2. Media — a Google Cloud Storage bucket with public read (allUsers → Storage Object Viewer) and a service account with Storage Object Admin on it. Media URLs point straight at https://storage.googleapis.com/<bucket>/….

3. Application — new resource → Git repository → build pack Dockerfile, port 3000, domains https://robopipe.io and https://shop.robopipe.io (the old Shopify host is redirected by next.config.ts). Under Advanced → Build, prefer Build secrets → Docker BuildKit secrets so build-time values don't end up in image metadata; the Dockerfile accepts both secrets and plain build args.

4. Environment variables

   Variable	Build	Runtime	Notes
   DATABASE_URL	✓	✓	internal URL of the Postgres resource
   NEXT_PUBLIC_SERVER_URL	✓	✓	https://robopipe.io, inlined into the client bundle
   GCS_BUCKET	✓	✓	media URLs are baked into prerendered pages
   PAYLOAD_SECRET		✓	the build uses a placeholder when it's not available
   GCS_CREDENTIALS		✓	service-account key JSON, on one line
   PREVIEW_SECRET, CRON_SECRET, RESEND_API_KEY, ANTHROPIC_API_KEY, PEXELS_API_KEY		✓
   INDEXNOW_KEY		✓	Optional fallback. The IndexNow key belongs in the admin: Site settings → Verification → IndexNow key (verification.indexNowKey, 8–128 letters, digits, dashes). It is served at /indexnow-key.txt and a change applies on the next request. Publishing a page, post or case study then notifies Bing and Seznam. This variable is used only while the field is empty. Neither set: nothing is sent

   Changing a build-time value needs a redeploy; runtime-only values just a restart.

5. Scheduled task (nightly blog generation) — on the application, Scheduled Tasks → add: frequency 0 3 * * *, timeout 900 s (one article takes ~3–5 min), command:

   wget -qO- --header="Authorization: Bearer $CRON_SECRET" http://localhost:3000/api/cron/run-jobs

• Build memory — pnpm build raises the Node heap to 8 GB; the Coolify server needs that much free RAM during builds (or lower --max-old-space-size in package.json).
• Single instance — the lead/newsletter rate limits are in-memory and ISR caches on the container filesystem; run one replica.
• Analytics — Google Analytics 4 and the Google Ads tag with a cookie banner (consent mode v2). The IDs are in the admin, Site Settings → Measurement, not in env vars. Empty fields = no Google tag, no banner. Fill them only after the privacy page describes the cookies (pnpm seed:cookie-policy, see below).
• Geo — the locale-suggestion feature reads Vercel's x-vercel-ip-country header in src/proxy.ts, so it stays dormant on Coolify unless that header is provided (e.g. mapped from Cloudflare's cf-ipcountry).
• Seeding prod (one-time): run pnpm seed locally with the production DATABASE_URL (expose the Postgres resource publicly for the duration) and the GCS variables exported, then redeploy so prerendered pages pick up the content.