Skip to content
Early bird + Black Friday · 73% off

Get ZeroDirs for $179 $49

Claim $49 deal

Changelog

Notable changes to the ZeroDirs Starter. The format follows Keep a Changelog, and versions follow Semantic Versioning as they apply to a codebase you own:

  • Major — a change that will not merge cleanly into a customised copy, or that changes the shape of site.config.ts.
  • Minor — new features and new config fields, with defaults that keep existing sites building.
  • Patch — fixes and documentation.

Quote the version below when you ask for support, and paste the full output of the failing command. The version is also in package.json. Support: hi@zerodirs.com.

Three paths are yours and no release edits them: site.config.ts, src/content/ and public/. Everything else is ours to change, so Changed and Removed entries are where a merge conflict would come from — zerodirs.com/docs/about/updates/ covers how to apply one.

  • pnpm cf:secrets is one command, no prompts (design.md N74): it reads .env.prod, checks the shape of every value, sends all runtime secrets to the Worker in one request and reads back what the Worker has. --dry-run shows the plan without sending, --list shows the names, what your site.config.ts needs and what is set, and bare names (pnpm cf:secrets STRIPE_SECRET_KEY) send only those — from .env.prod, no longer from a prompt. --all still works and means nothing; --from-env and the fallback to .env are gone (test-mode keys must never reach production by accident). A key in .env.prod that .env.example does not know stops the run. To upgrade: nothing in your files changes; create .env.prod from .env if you have not, and drop --all from any note you kept.
  • The header, redrawn (design.md N68): the Submit call-to-action is a pill with a paper-plane icon and the word “Submit” (its accessible name is still “Submit a {noun}”); the search field is gone from the header — the home page carries the big one — and the theme toggle and a new account icon sit to the right of the pill. The account icon links to /dashboard/ and replaces the “Account” item in the nav; for a signed-out visitor it opens the sign-in dialog (next entry), and for a signed-in one it is their avatar, which opens a menu — name, address, Dashboard, Bookmarks, Log out (design.md N78). Content pages are static and identical for everyone, so the header learns who is signed in from a second cookie, zd_user, that holds a name, an address and an avatar URL and nothing a thief could use; the session itself stays in the HttpOnly zd_session. To upgrade: add 'send', 'user', 'layout-dashboard', 'bookmark', 'log-out' and 'chevron-down' to theme.icons in site.config.ts; without them the pill shows only its label and the menu only its words, and nothing fails.
  • Signing in happens in a dialog (design.md N76): pressing the account icon opens the sign-in card on the page the visitor is already on — email and password, “Continue with Google” when both Google secrets are set, “Forgot password?” and “Don’t have an account? Sign up” — instead of sending them to /login/. A wrong password comes back into the dialog, with the address kept and the password field focused; a right one sends the visitor on to the page they were after, or to /dashboard/. The card is asked for when the dialog opens (/login/dialog/, rendered by the Worker), because a static page cannot know who is signed in or whether the Worker holds the Google secrets. /login/ stays — it is where /submit/ and /dashboard/ send a signed-out visitor, and what a browser without JavaScript gets — and is redrawn as the same card. It costs one more inline script on every page while features.accounts is on, about 2.6 KB; still no island and no <script src>, and the e2e smoke test counts it. To upgrade: take the merge. If you had edited src/templates/account/LoginPage.astro, its markup now lives in src/components/account/AuthCard.astro; the three form-validation colours moved from src/components/forms/Field.astro into src/styles/global.css.
  • Submitters sign in with a password, not an emailed link (design.md N77). An account is made on /signup/ with an email and a password (eight characters or more), and signed in on /login/ or in the dialog with the same; “Continue with Google” is unchanged. Passwords are stored as PBKDF2-HMAC-SHA-256 at 100,000 iterations with a per-password salt, in a form that names its own parameters, so the count can be lowered later (the free plan’s CPU budget is tight for it) without invalidating anything. The one emailed link left is “Forgot password?”: /login/forgot/ queues a reset link that works once and for twenty minutes, /auth/reset/ is where it lands, and the password-reset email template replaces login-link. Ten wrong passwords in a quarter of an hour lock the address for the rest of it; the form says “wrong email or password” and nothing more, whichever it was. Nothing checks the mailbox at sign-up, so an address is only proven by a Google sign-in or a completed reset — and the first time it is proven, a password somebody else registered for it is dropped, with its sessions: the mailbox’s owner gets the account, and sets a password of their own through the reset link. To upgrade: pnpm db:migrate:local, then pnpm db:migrate:remote once (0002_passwords.sql adds two columns to users and marks every account that exists as proven). Accounts made under 0.2.0 have no password: Google keeps working for them, and “Forgot password?” sets one. requestLoginLink / consumeLoginLink are gone from src/server/auth/; /auth/callback/ with it.
  • /account/ is now /dashboard/ (design.md N78, N79), an area of its own with a sidebar — Overview, Submissions, Bookmarks under Dashboard, Account under Settings, then Visit site — and a top bar with the Submit button, the theme toggle and the account menu. The sidebar folds to a rail of icons on a wide screen (every icon is still the link, with its name as the tooltip) and stays folded from page to page through a zd_rail cookie; on a phone it is a drawer. Overview: four numbers (submitted, live, in review, clicks in the last 30 days) and a table of every submission that ?q= searches by name or host and ?sort=/?dir= order by name, status, clicks, submission or publication date — plain links and a GET form, no script. Submissions: the same rows under tabs, with what can be done about each — Upgrade a submission waiting in the free queue onto a priced plan (a popover with the plans, each leading to its checkout), Complete payment, View, Submit again, and Withdraw a submission that is still waiting and not paid for: it asks once, then the submission becomes “Withdrawn” (stored as rejected with the reason Withdrawn by the submitter., so the operator sees why it left the queue; no email) and the address is free to be submitted again. Account: the avatar, address and sign-up method with a Log out button, the name, and the password — change it with the current one, or set a first one on an account that has only used Google; either signs out every other device. /account/ answers with a permanent redirect to /dashboard/, and dashboard joins the reserved route segments and the robots.txt disallow list. To upgrade: take the merge; add 'zap', 'arrow-up', 'arrow-down', 'chevrons-up-down', 'panel-left', 'lock' and 'calendar' to theme.icons (each is optional, like the others); a routes.*Base named dashboard has to be renamed, which pnpm check:config will say. The src/templates/account/AccountPage.astro of 0.2.0 is gone; the dashboard’s pages are src/templates/account/Dashboard*.astro on src/layouts/Dashboard.astro.
  • ⌘K / Ctrl K opens a command palette on every page: a search field, the site’s quick links and its featured listings (logo, name, host), filtered as you type; Enter runs the full search on /search/. The ⌘K badge in the home-page search field opens it too, and reads Ctrl K off a Mac. It is a native <dialog> with server-rendered rows and one inline script of about two kilobytes — no island, no <script src>, so rule 17 still holds and the e2e smoke test counts it. theme.icons should list search and x for its two icons (the demo config already does).
  • The first pnpm dev of a fresh clone takes about twenty seconds instead of three and a half minutes, and pnpm content:reset --local a few seconds instead of the same. Every logo used to go into the local bucket through its own wrangler r2 object put --local (and out through its own delete), 1.5 s of wrangler start-up per object; against --local the seed and the reset now open the bucket binding once in-process, through wrangler’s getPlatformProxy(), and put concurrently or delete in one batch. Downloads run sixteen at a time instead of eight, and a host that has not answered in five seconds is a letter tile rather than a twelve-second wait. --remote is unchanged: one wrangler process per object, no S3 credentials needed.
  • /signup/, a create-account page (design.md N76, N77), where the sign-in dialog’s “Sign up” leads: a name (optional, it ends up on the account), an email and a password, or Google when it is configured, with the terms and privacy pages linked when legal has them. signup joins the reserved route segments and the robots.txt disallow list. To upgrade: a routes.*Base named signup has to be renamed, which pnpm check:config will say.
  • Bookmarks (design.md N79): a Bookmark button on every listing page, under “Visit”, for a signed-in visitor — pressed, it saves the listing in place and reads “Bookmarked” from then on; pressed by a signed-out visitor it opens the sign-in dialog and comes back to the page. /dashboard/bookmarks/ lists what was saved, newest first, with Open and Remove. The button is a form that works without JavaScript (the post lands on the bookmarks page), plus one inline script of about 0.8 KB on listing pages only, which asks /api/bookmarks/<id>/ whether the listing is saved — a static page cannot know — and posts to it with fetch. To upgrade: pnpm db:migrate:local, then pnpm db:migrate:remote once (0003_bookmarks.sql adds the bookmarks table). features.accounts off leaves the button out; there is no flag of its own. A seed or pnpm content:use keeps bookmarks: seeded listings keep their ids, and a bookmark whose listing is gone is simply not shown.
  • A submission that was refused — or, now, withdrawn — could still be paid for: its status page kept its “Complete payment” button, and /api/checkout/ took the money for a listing that would never be reviewed. Both now apply to a pending submission only.
  • The Google sign-in failed with “Sign-in did not complete” on every attempt: the OAuth state cookie was URL-encoded twice on the way out and never matched on the way back. The reason a Google round trip declines is now logged (auth: Google sign-in did not complete — …), so a wrong client secret or an unregistered redirect URI can be told apart from a refused address. Under pnpm dev the redirect URI is the local origin (http://localhost:4321/auth/google/callback/), so “Continue with Google” can be tried locally with that URI registered as a second one.
  • Deploys no longer fail with “required secrets have not been set” (design.md N75). Since 0.2.0, wrangler.jsonc listed every runtime secret under secrets.required, and wrangler deploy refuses to deploy while any of them is missing on the Worker — so pnpm cf:setup (which has to deploy once before it can put a secret), Workers Builds and the GitHub Actions workflow all failed for a site that leaves an optional provider off. The secrets block is gone from wrangler.jsonc altogether: an empty list would unblock the deploy, but any secrets block makes wrangler hand the local Worker only the names it lists, and pnpm preview would run without the secrets in your .env. What the list was there for — a worker-configuration.d.ts that does not depend on your .env — is now pnpm cf:typegen reading the names from .env.example (wrangler types --env-file .env.example); CI checks with pnpm cf:typegen --check. To upgrade: take the merge; if you had added a secrets block or names of your own, remove them, and use pnpm cf:typegen rather than a bare wrangler types. Env now lists every key of .env.example, the four build-time ones included. One thing the list did by accident is now done by name: astro preview reads its secrets from the build’s snapshot of .env (dist/server/.dev.vars) or from process.env, never from the .env file itself, so a build made before .env existed served a Worker with no TOKEN_SECRET; integrations/preview-secrets.ts sets wrangler’s CLOUDFLARE_INCLUDE_PROCESS_ENV for preview, and pnpm test:e2e rebuilds when .env is newer than dist/.
  • The first pnpm dev of a fresh clone exited with The file does not exist at "node_modules/.vite/deps_ssr/handler-….js"; the second one worked. Vite discovered the preact renderer’s server entry on the first request, re-optimized the SSR dependencies and reloaded, and the Cloudflare plugin’s worker was still holding the chunk that had just been replaced. The entry is now pre-bundled at startup, and CI starts the dev server on a cold cache and fails when it does not answer.
  • One environment file. Build-time variables and runtime secrets now share .env, with .env.example as the committed key set and .env.prod (gitignored) as what production gets. .dev.vars is gone: wrangler reads it in preference to .env, so keeping both meant a stale file could silently shadow every secret you edited. To upgrade: cat .dev.vars >> .env && rm .dev.vars, then pnpm cf:typegen. pnpm dev refuses to start while a .dev.vars is still there, and prints that same line. wrangler.jsonc now declares secrets.required, so wrangler types writes the same Env on every machine and CI no longer seeds a secrets file before checking the committed types.
  • Listings always come from D1. Every build reads the database: with the three D1 credentials set (CLOUDFLARE_ACCOUNT_ID, D1_DATABASE_ID, D1_READ_TOKEN — a production builder) the real one over its REST API, exactly as before; with none of them set (your machine, CI) the local Miniflare database under .wrangler/, through wrangler d1 execute --local. A partial set fails the build and names the rest, and a builder that has lost its variables has no local database either, so it fails instead of shipping the seed listings as a success. pnpm dev seeds the local database with the demo dataset — rows and logos — the first time it finds it empty, and never touches a database that holds anything. A local build links every logo through the /media/ route, which the Worker now serves in every environment (production links keep using media.baseUrl), and the OG images of a local build carry no logo. Outbound click counting therefore works in every build: every listing has a row id. To upgrade: listings you kept in src/content/listings/ go into the database once — pnpm db:seed:local --from src/content/listings, then the same with --remote — and the directory, public/logos/, LISTINGS_SOURCE and ZERODIRS_DATASET can go; a Workers Builds or Actions setup that built from Markdown now needs the three credentials. pnpm cf:doctor says which database a build here reads.
  • pnpm content:reset and pnpm content:use act on a database (--local by default, --remote): reset deletes the seeded rows, their logos in the bucket and the demo blog posts and leaves the database empty; use <dataset> seeds it with a dataset and copies the dataset’s blog posts. pnpm listings:import writes to a database (--local or --remote, one is required unless --dry-run) and no longer to Markdown files; pnpm check:config counts the approved rows in the local database for its estimate.
  • pnpm cf:setup finishes the email step instead of only reporting it: with providers.email: 'cloudflare' and no sending domain it offers wrangler email sending enable <domain>, keeps the EMAIL binding when that works, and puts the binding back on a later run once the domain exists — previously the binding was dropped and had to be restored by hand.
  • The default look, redrawn (design.md N51): a search-first home page with the category chips under the field, cards led by a logo tile with the name over its category, a browse strip of category chips on the listing index and category pages, a sticky facts card with the outbound button on the detail page, a stepper and option cards on the submission form, a four-tile queue summary on the admin dashboard, and a sticky translucent header with a Submit button on every page. Still zero JavaScript. The tokens are the same fifteen; default’s radius moves from 0.5rem to 0.75rem, and cards, buttons and chips are now fixed multiples of it (rounded-brand-lg, rounded-brand-pill), so mono stays square and warm rounder with no new knob. Three component classes join .btn and .card: .chip, .input, .hero-glow. A copy that overrides radius gets the same proportions at its own size; one that edited Card.astro, Detail.astro, Header.astro, Footer.astro or HomePage.astro will merge by hand.
  • Paid plans ship on: features.payments: true, providers.payment: 'stripe' and both priced tiers enabled in site.config.ts, the way the demo runs. The two Stripe secrets are yours to set (pnpm cf:secrets STRIPE_SECRET_KEY STRIPE_WEBHOOK_SECRET); until they are on the Worker a paid submission’s checkout answers 503 “Payment is temporarily unavailable”, /admin/ reports both keys, and the free plan keeps working.
  • Email ships on, through Cloudflare Email Sending: providers.email: 'cloudflare' in site.config.ts and the EMAIL (send_email) binding declared in wrangler.jsonc. It needs the Workers Paid plan and a sending domain enabled on your zone; until both are true every message still renders into email_outbox and /admin/ says why. On the Free plan, switch to Resend — the skill walks through it.
  • The README troubleshooting section names the npm install crash of npm 11.4 (Node 24.3) on vitest 4’s peer dependencies and the three ways out; pnpm is unaffected.
  • pnpm cf:secrets is a conversation now. It starts by listing which secrets the deployed Worker already has, shows each key’s help text line by line (every key in .env.example now says where its value comes from), takes names as arguments (pnpm cf:secrets STRIPE_WEBHOOK_SECRET), checks the shape of a value before uploading it — a whsec_ pasted into STRIPE_SECRET_KEY, the placeholder TOKEN_SECRET, a publishable key — and asks before going ahead, and ends cleanly on Ctrl+C with what was set and what is still missing. Every prompt also says whether the current site.config.ts needs that secret at all (a GitHub token is idle on a site that publishes through a deploy hook). Wrangler’s banner no longer interleaves with the prompts.
  • The demo’s top banner (ads.topBanner) is a house ad for ZeroDirs: on your own site, give it your sponsor’s text and url, or set enabled: false.
  • pSEO, the OG images and the related-listings ranking moved to src/features/pseo/, src/features/og/ and src/features/related.ts, and src/features/index.ts is gone. Templates and components import those modules directly, and each entry function checks its own flag (pseoPages() reads features.pseo, relatedListings() reads features.relatedLinks) and takes the config as an argument. A copy that edited one of these files merges it as a rename; one that imported from src/features/index.ts changes that import. The hard rules in AGENTS.md are renumbered.
  • pnpm cf:media: the one step pnpm cf:setup could not finish. It makes the R2 bucket readable from the web — a custom domain when you have the zone id, the r2.dev URL otherwise — and writes the result into media.baseUrl, restoring site.config.ts untouched if the edited file does not validate.
  • pnpm cf:doctor: one read-only screen for “is this actually wired up?”. It checks the tooling, site.config.ts, wrangler.jsonc, .env, which secrets the Worker holds, the R2 bucket and sending domain, and whether the live site was built with the config you have now — each line either ✔ or a problem with the command that fixes it. It writes nothing, and a check it cannot make (no login, --offline) says so instead of guessing. --json for an agent, --strict to fail on warnings, --manual for the steps no CLI can do.
  • pnpm cf:secrets --all: every runtime secret in .env.prod (or .env) in one wrangler secret bulk, values on stdin. Empty, placeholder and wrong-shaped values are left out and named, so a truncated paste cannot reach the Worker. --dry-run lists what it would send without sending anything.
  • pnpm cf:ci: the build-time half of the same file pushed to GitHub Actions with gh variable set / gh secret set, and the Build step’s env: block in .github/workflows/deploy.yml rewritten from .env.example between marker comments, so a variable you add is never missing in CI. --enable sets USE_GITHUB_DEPLOY.
  • pnpm cf:setup: from a checkout to a running site in one command — Worker name, custom domain or *.workers.dev, D1 and R2 created and written into wrangler.jsonc, the send_email binding left out until the account can send, production migrations, build, deploy, and TOKEN_SECRET. Idempotent, with --dry-run and --workers-dev.
  • pnpm dev (and pnpm test:e2e) now runs scripts/dev-setup.ts first: .env from its example when missing, the D1 migrations applied to the local database, and the demo dataset seeded into it while it is empty — so a fresh clone reaches /, /submit/ and /admin/ on its first pnpm dev with no Cloudflare account.
  • A go-live skill: the steps cf:setup leaves to you — domain, the real database for the builder, Workers Builds and the Deploy Hook, email, payments — one question at a time.
  • Submitter accounts (features.accounts, on in the demo): sign in with an emailed link that works once for twenty minutes, or with Google once GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET are set; /account/ lists a submitter’s own submissions with their review status; /submit/ requires a session and takes the address from it. Sessions are 30-day cookies whose ids are stored hashed; the hourly cron sweeps what expired. One expand-only migration (0001_accounts.sql) adds users, sessions, login_tokens and listings.user_id. The configure-accounts skill has the Google steps. Ten skills.
  • Ad placements, each one boolean in the new ads block: a banner under the header, a sponsor wall under the home page hero, a card inside every listing grid, and a corner card that can be dismissed without JavaScript. /advertise/ lists the placements and their prices and is built only while ads.page.enabled. Everything is static and marked as an ad (rel="sponsored"); the demo shows all four with placeholder copy.
  • The free review queue says how long it is. A free submission’s status page and receipt email state when it is expected to go live — position in the queue divided into the last 30 days of real approvals, or queue.estimateDays (21) while there are none — and offer the priced plans to skip ahead: one click to Stripe Checkout, and the tier changes only once the payment lands. The operator gets one reminder a day at queue.digestHourUtc (09:00 UTC) while anything is waiting, with the oldest wait and the paid submissions close to their review promise; null switches it off. The admin dashboard shows the same numbers.
  • site.builtWith (default true): a “Built with ZeroDirs” link in the footer, one boolean to remove. Not a licence condition.
  • A configure-email skill: the three email providers side by side, the Cloudflare Email Sending path, the Resend path for the Workers Free plan, and how to verify with the outbox and the /admin/ configuration report.
  • tests/unit/put-secrets.test.ts: the pure parts of pnpm cf:secrets — the help text every key gets, name selection, and each shape check, including that no check ever repeats the value.
  • .dev.vars and .dev.vars.example. See the .env entry under Changed.
  • Markdown as a listings source: src/loaders/files.ts, src/content/listings/ (a copy of the ai-tools dataset), public/logos/ (its 14 tiles, still in seed/datasets/ai-tools/logos/), the LISTINGS_SOURCE and ZERODIRS_DATASET build variables, pnpm content:check, and pnpm listings:import’s --out / --to. See the D1 entry under Changed for the one-time move.
  • docs/. The documentation is the site at zerodirs.com/docs, which is generated from the code and cannot drift the way the copies here did, and nothing in the starter links to the old files any more. Their text is in the git history (git log -- docs/). If your copy edited a file under docs/, the merge reports a modify/delete conflict: keep the file or delete it, nothing reads it.
  • The submit Action built its email provider without the Worker’s bindings, so with providers.email: 'cloudflare' the console provider “sent” every receipt: the outbox rows were marked sent, nothing left the account, and no error was raised anywhere. Both Actions and the cron now hand env.EMAIL over through one helper, and a unit test fails if any of them ever builds a provider without it.
  • features.search: false failed the build: /search/ was still built, island included, but had left the routes manifest, so the zero-JS check no longer exempted it. The page now returns a body-less 204 while the feature is off, the way /blog/ does, and no file is written.
  • The search panel’s message under pnpm dev said a build alone would enable search there; it is pnpm build and then pnpm preview.
  • LICENSE §5 said pnpm content:reset --yes removes the demo datasets. It removes the demo listings, the demo posts and the files in public/logos/, and leaves seed/datasets/ in place; the licence now says so. Comments that miscounted the runtime secrets or the hourly jobs (.env.example, scripts/put-secrets.ts, wrangler.jsonc, src/worker.ts, src/server/jobs.ts, AGENTS.md) are corrected too.
  • The import-listings and generate-listings skills said a numeric custom field does not render and told the agent to quote it. Numbers and booleans render; only arrays and objects are dropped, which is what both skills now say.

The first buyer release. Everything listed here works today. Several subsystems ship switched off because switching them on needs a database or a credential we cannot put in a repository — the README’s “What’s on, what’s off” table says what each one still needs.

Directory core

  • Listings, categories, tags and pagination from Markdown in src/content/listings/, one file per listing, validated against site.config.ts at build time.
  • One catch-all route builds every directory page from a pure, unit-tested route table, so the URL segments (routes.listingBase, categoryBase, tagBase) are yours to rename.
  • Related listings on every detail page, and configurable per-listing custom fields.
  • Full-text search over listing pages (Pagefind), indexed during the build. /search/ is the only page in the site that hydrates.
  • Blog with its own feed at /blog/rss.xml, a site-wide latest-listings feed at /rss.xml, and about, privacy, terms and 404 pages.

Submissions, moderation and payments (features.submissions, features.payments)

  • A public two-step submission form that prefills itself from the submitted site’s own metadata, a honeypot trap, per-email rate limiting, and a signed, token-addressed status page that needs no account and no password. The submitter never sees the word “logo”: the prefill ranks every icon the page declares, and the server downloads them in order until one is storable.
  • An admin area — review queue, listing editor, featured placements and stats — open to the accounts named in admin.emails, with a Publish button and an optional rebuild on approval (providers.rebuild: Cloudflare Deploy Hook or GitHub repository_dispatch).
  • Stripe checkout for priced tiers, a signed webhook that records the payment idempotently, and an hourly cron that sweeps unpaid submissions, retries the mail outbox and expires featured placements.
  • Transactional email through Resend, Cloudflare Email Sending or a console provider, every message rendered into an email_outbox table first so nothing is lost to a failed send.
  • Outbound click counting at /out/<id>/ — <a ping> by default, so the listed site keeps the link it was sold — and self-hosted newsletter capture into D1.
  • A D1 schema with drizzle-kit migrations, a build-time D1 REST loader (LISTINGS_SOURCE=d1), and seed/export scripts in both directions between Markdown and the database.
  • pnpm db:seed --with-images stores the dataset’s own logos in your R2 bucket — an absolute https:// URL in the frontmatter is downloaded, a /logos/ file is uploaded — and --with-favicons fetches the icon each remaining site publishes about itself, so moving from Markdown to D1 does not lose a picture the Markdown named. A favicon.ico is decoded to the PNG of its largest frame, in plain TypeScript, so most entries come back with a picture; sites that block crawlers keep the initial-letter tile. Every object carries a five-minute cache-control, so a re-seed shows within minutes and needs no cache purge.

SEO

  • Programmatic SEO pages from two configurable patterns (best-category-for-tag, name-alternatives), with a thin-content guard, a page cap, and per-page copy overrides from src/content/pseo/. The demo data generates 211 of them.
  • Open-graph images rendered at build time with satori + resvg, one per listing and category, in your theme’s colours; plus a regenerable static fallback (pnpm og:default).
  • Chunked sitemap, robots.txt, canonical URLs, JSON-LD (breadcrumbs, item lists, organization, and a listing type chosen from site.kind), and a noindex rule for thin taxonomy pages.
  • A build-time SEO report that counts inbound internal links per listing and can fail the build on orphans (seo.strictLinks).

Configuration and theming

  • site.config.ts as the single configuration surface, zod-validated with ten cross-field rules, every error reported as site.config.ts › path: message.
  • Three theme presets (default, warm, mono), 14 CSS tokens overridable from the config, dark mode with a no-flash bootstrap, and self-hosted web fonts via Astro’s font pipeline.

Tooling

  • pnpm check:config — validate the config and estimate the static-file budget before building.
  • pnpm content:reset / content:use / content:check — swap between the demo dataset and examples generated from your own categories.
  • pnpm listings:import — CSV or JSON to Markdown, with column aliases, --map overrides and a --dry-run that reports every bad row at once.
  • A static-file budget check that warns at 80% and fails the build at 90% of the Cloudflare plan limit, so an over-size version is stopped locally rather than rejected at deploy.
  • 857 unit tests (pnpm test), 379 more inside workerd against a real D1 (pnpm test:workers), 30 Playwright tests over the built site (pnpm test:e2e), and a Lighthouse CI config that audits one page per template from your own route manifest (pnpm lhci), asserting ≥ 0.95 in all four categories.

Deployment

  • Static output to Cloudflare Workers assets. pnpm deploy builds and uploads. A Worker script rides along for the on-demand routes and the hourly cron, but ships with no DB or MEDIA binding and no secrets: with the default flags nothing reaches it, and a first deploy provisions nothing.

Built and tested, off by default. Each needs something this repository cannot carry for you:

  • Submissions and the admin queue (features.submissions) — a D1 database, the DB binding, and pnpm db:migrate:remote.
  • Payments (features.payments + providers.payment: 'stripe' + the priced tiers enabled) — STRIPE_SECRET_KEY and STRIPE_WEBHOOK_SECRET. Cross-rule 3 requires all of them together, so a priced tier cannot be enabled with payments off.
  • Self-hosted newsletter (features.newsletter: 'd1') — the same DB binding. features.newsletter: 'embed' needs nothing but your provider’s HTML.
  • Rebuild on approval (providers.rebuild) — a Deploy Hook URL, or a GitHub token and repo. With 'none', approved listings go live on your next pnpm deploy.
  • D1 as the build’s data source (LISTINGS_SOURCE=d1) — CLOUDFLARE_ACCOUNT_ID, D1_DATABASE_ID and a D1 Read token. Until then the build reads src/content/listings/.
  • Transactional email (providers.email: 'resend') — RESEND_API_KEY and a verified sending domain. Without it every message still renders, into the log and email_outbox.
  • Firefox does not send <a ping> at all, so features.clicks: 'ping' undercounts its visitors; 'redirect' counts them all at the cost of the link equity.
  • A local build’s OG images carry no logo. They link logos through the /media/ route, i.e. the local bucket, which the build cannot read; a production build fetches them from media.baseUrl as before.
  • listing.screenshotMax and limits.screenshotMaxBytes validate but nothing reads them: every entry in a listing’s screenshots array renders, and nothing in the app uploads one.

Node 24 (.nvmrc), pnpm 10.12.4 (packageManager), Astro 7.3.1, wrangler 4.131.1.