Changelog

Cartwright templates ship as tagged releases. npx create-cartwright pulls the template at the current DEFAULT_REF tag (managed in cartwright-app/apps/cli).

A scaffolded shop is a one-shot snapshot — nothing updates it automatically. The Security advisories index below is therefore the canonical place to learn whether the engine version your shop runs (see .cartwright/release.json) has a known security fix you should pull. When a release fixes a security issue, its version block gets a ### 🔒 Security section (issue + severity + the version you must upgrade to) and a row is added to the index below.

Unreleased

v0.57.0 — 2026-09-14

The line between the customer's site and Cartwright's own repository, drawn in code — and a new scaffold's production dependencies have no known advisory. This release draws the line between what belongs to the site you scaffold and what belongs to Cartwright's own repository, and adds a zone of guards that ships into every scaffold, so the defects that only a scaffold can show can be caught where they happen. It also closes all fourteen advisories pnpm audit --prod reported on v0.56.2 (CW-2026-003), and fixes six places where a page, a guide or an agent surface spoke the wrong language or promised something it did not do.

🔒 Security

• Fourteen advisories closed (8 high, 6 moderate) — `pnpm audit --prod` now reports none.

Thirteen sat below packages the engine depends on, so they are pinned with overrides in pnpm-workspace.yaml, each selector scoped to the line the package is on (its major; for pre-1.0 sharp, its 0.35 minor): fast-uri ≥ 3.1.6 (four GHSAs on URI parsing, under ajv — reached from @modelcontextprotocol/sdk and from the Prisma CLI's @prisma/dev), sharp ≥ 0.35.4 (libheif, Next's image optimiser), qs ≥ 6.16.0 and hono ≥ 4.13.5 (the SDK's Express/Node HTTP stack — the engine's /api/mcp uses the web-standard transport, so these are installed but not loaded by engine code), mysql2 ≥ 3.23.1 (the Prisma CLI pins 3.15.3 exactly and uses it only against a MySQL database) and deepmerge-ts 8 (@prisma/config pins 7.1.5 exactly; the one step across a major, because 8.x is the only patched line — @prisma/config loads it with a dynamic import() when it reads prisma.config.ts, and prisma validate loads that config and passes). The fourteenth is the direct dependency js-yaml, which moves to ^4.3.2; a separate override closes the copy ESLint installs. Tracked as CW-2026-003. Left for the next Vitest bump: vite 8.0.14 under Vitest (dev-only, a peer an override cannot move) (#583).

✨ Added

• An ownership ledger for the scaffold (scaffold/engine-only.ts). A module claim could only

say which profiles *exclude* a file, so everything unclaimed shipped everywhere — which is how Cartwright's own /cartwright and /changelog answered 200 on a customer's domain, under her navigation, with llms.txt advertising the changelog. The manifest now has two words it lacked: ENGINE_ONLY (ours, to be removed from a customer's scaffold) and SHIPS_EVERYWHERE (the customer's, unclaimed on purpose), and an inversion test treats "unclaimed" as an error. Every link into an engine-only route is gated on profileCapabilities.engineMarketing, which is false today only in the site profile — so a site scaffold is the only one whose engine-only routes can be deleted without leaving dead links (scaffold/manifest.json → engineOnlyGate lists the profiles still pending). This release ships that engine half; create-cartwright 2.9.5 does not read the ledger yet, so a scaffold cut today still carries the routes (#575).

• A guard zone that runs inside every materialized scaffold. tests/unit/ does not travel to a

site scaffold, but every defect found on the way here was invisible in the engine and visible in a scaffold. Eight new guards live in a zone that ships into every scaffold. Six are static and run in the scaffold's own pnpm test — dead fetch("/api/…") literals, dead script paths, our identity left in brand.config.ts, the profile marker, the guards' own portability, and a ratchet on the engine-only routes that arms once a create-cartwright release claims the ledger. Two are live and run when GUARD_BASE_URL points at a running site: the homepage's links, icon, og:image and JSON-LD logo answer 2xx, and the sitemap resolves and lists every served locale's homepage (#577).

🛠 Fixed

• A site scaffold's own header linked routes the profile had removed — /services and

/account 404'd from a --profile site header and drawer. Marketing entries can now declare the capability they require; database-backed profiles render byte-identically (#572).

• The product page spoke the shop's language only in part. Twelve PDP strings were hardcoded —

eleven in English on a page whose default locale is Danish, one in Danish that English readers got instead. They now read a Product namespace in messages/{da,en}.json (#580).

• The trust badges linked outside the reader's locale. The shipping, returns and FAQ badges

used locale-less /info/* paths, so every click cost a redirect and an English reader without a cookie landed on the default locale. The locale-link gate now also scans object properties, the form that let this slip through (#582).

• `llms.txt` called the sitemap "the complete index of all public pages". It was not, measured

three ways on three live sites. The line now says what the sitemap advertises (#578).

• WebMCP form tools say which field is required. site_search and newsletter_signup are

declarative tools whose schema the browser derives from the form; required comes only from the attribute, and the attribute was missing. The page still answers every submit itself (#581).

• `docs/simple-site.md`'s receipt named an older release than the one it shipped in — v0.56.2

carried the v0.56.1 run. It now cites the measured v0.56.2 run, and a test keeps the receipt a released version at most one release behind the engine's CHANGELOG (#573).

Changed

• Engine-private notes no longer reach the public template mirror — overnight reports and

internal drafts outside internal-docs/ are excluded, and the exclusion is pinned (#574).

Upgrading

Copy the overrides block from pnpm-workspace.yaml into yours, set "js-yaml": "^4.3.2", run pnpm install and confirm with pnpm audit --prod (see CW-2026-003 for shops still on next 16.2.x). That is the only step for an existing shop; the rest reaches new scaffolds. Do not copy engine files into a --profile site scaffold: its components/Header.tsx, components/Footer.tsx and lib/profile-capabilities.ts are materialized static variants. The engine's Header.tsx and Footer.tsx import modules that profile does not have, and the engine's lib/profile-capabilities.ts advertises routes the profile removed.

v0.56.2 — 2026-09-06

The plain-website door, inside the scaffold an AI actually reads. A --profile site scaffold shipped a README whose lede promised "database and backend", whose first command block put the site profile third, and whose deploy section asked for three Turso environment variables — and, worse, the two files a coding agent loads before any of that (.claude/CLAUDE.md and AGENTS.md) opened with pnpm db:setup and an admin login. An agent that scaffolded the database-free profile was told, by the profile itself, that it needed a database.

• The door, at the top (#570): the README lede is qualified ("a real site

with design — and, when you want them, database and backend"), the site command comes first in the first command block, and a new "Just a website? (--profile site)" section says what ships and what does not. A pointer names the profile marker: *"Reading this README inside a scaffold? .cartwright/profile.json says which profile you have"* — and which sections do not apply to it.

• The briefings agents load first: one callout at the first-run step of

AGENTS.md, .claude/CLAUDE.md, GEMINI.md, .github/copilot-instructions.md, .cursor/rules/cartwright.mdc and .windsurfrules (the six files AGENTS.md requires to stay consistent), plus DEPLOY.md: no database, admin or login here — pnpm dev is the whole first run, pnpm build the whole production gate, and deploying needs no environment variables.

• `docs/simple-site.md` gains "What you edit" (the brand.config.ts

fields, designSlug, the blank canvas, adding a route), "Measured" (the measured cold run, copied from the release scaffold gate's timings-site.json and never typed — the guide carries the figures and the run id, so a re-measurement updates one file) and "Honest limits".

• `llms.txt`: the "Built with Cartwright" block names the database-free

profile, and the product sentence is qualified in both variants.

• tests/unit/readme-site-door.test.ts (engine-only) pins all of it: the

command order, the section, the pointer, the env-var scope, the guide's sections and provenance shape, the badge in both variants, and the callout in every briefing.

Upgrading

Nothing to do — documentation only, and no runtime file changed. A scaffold cut from v0.56.0 or v0.56.1 can copy the new README.md, docs/simple-site.md and agent-briefing callouts across, or simply ignore the database sections its profile does not have.

v0.56.1 — 2026-09-06

A light or full scaffold's own test suite is green again. The v0.56.0 release scaffold gate caught it before any customer did: tests/unit/ site-font-safe.test.ts (new in #564) enumerated files with git ls-files, and a scaffold is not a git repository — three tests failed in every light and full scaffold cut from v0.56.0 (site was unaffected: it prunes tests/unit). The published CLI was still pinned to v0.55.0 while this was found, so no scaffold in the wild ran the red suite.

• site-font-safe pins the engine, not a scaffold (#568): files are walked

from disk (no git), and the suite runs only in the engine checkout — it pins the engine's manifest against the engine's design packs, which a light scaffold has already pruned — using the same probe as repo-hygiene.test.ts. Proven three ways: engine → 4 pass; a git archive copy with no .git → 4 pass; the same copy without the mirror exclude list (scaffold-like) → 4 skipped; and a real create-cartwright@2.9.2 --profile light --ref v0.56.0 scaffold's suite red before / green after.

Upgrading

Nothing to do. A light/full scaffold cut from v0.56.0 can copy the new tests/unit/site-font-safe.test.ts over its own, or delete that file.

v0.56.0 — 2026-09-06

A `--profile site` scaffold is now what its own README says it is — a plain website with no database. It boots and builds with Google's CDN unreachable, ships no crons into routes it does not have, installs no Postgres driver, serves its share cards and favicon, and its contact form reaches its owner. Five defects, each one an AI scaffolding a "just a page" site could not have fixed itself; four of them were found by review falsifiers on a real create-cartwright@2.9.2 --profile site scaffold.

• **No crons for a scaffold without cron routes, no Postgres driver without a

database** (#563): vercel.json became a seam owned by the db module. A site scaffold receives vercel.static.json ($schema + framework only) instead of the ten cron entries whose routes the materializer deletes — every site deploy used to 404 on all of them, daily, forever. The orphan pg + @types/pg left the engine root; @prisma/adapter-pg carries both for the profiles that have a database. repo-hygiene.test.ts pins the twin as vercel.json minus crons.

• Font-safe boot (#564): the nine design packs whose display faces are

fetched from Google Fonts at build time (aerospace, brutalist, drive, editorial-ink, fable, flux, meridian, nocturne, stillwater) and the root layout's Geist pair moved into a new google-fonts module, present from managed-site upward. A site scaffold gets app/layout.static.tsx — the system font stack behind the same --font-geist-* variables — and the eight self-contained packs (aurora-site, saas-dark, studio, corporate-baseline, stack, jungle, agentic-showcase, blank). Measured with Google blackholed: the engine's own next build dies on Space Mono through the designs/index.ts barrel; the materialized site builds (23 routes), next start serves /en 200, and the FIRST next dev compile answers 200 — the GET /en 500 that Turbopack cached until rm -rf .next is gone. site-font-safe.test.ts walks the site's included file set and forbids next/font/google; the engine, light and full output are byte-identical (app/layout.tsx untouched).

• The contact form reaches the owner (#565): SmartContactForm asked the

admin-owned /api/support/triage before /api/inquiries — unconditionally. In a site scaffold that route does not exist, the POST hit the catch-all (405, empty body), response.json() threw, and every visitor saw "Could not connect to the server" while /api/inquiries itself worked. The import walker cannot see a fetch URL, which is how a profile documented as "no AI triage" shipped a form that could not submit. profileCapabilities. supportTriage (true; false in the static twin) now gates it and the site form goes straight to the human path. Database-backed profiles render byte-identical.

• Share cards and favicon are not dead links (#566): proxy.static.ts,

the middleware a site scaffold runs, exempted only /api from the locale rewrite, so /og?title=… — the URL every page's og:image carries — and /icon were 307'd into /en/og → 404 in next dev and next start alike. The twin now routes them by the same isAssetExempt the db variant uses; a wiring test drives the static proxy directly and is mutation-proved.

Upgrading

No schema change and no migration. A site scaffold cut from an earlier tag has no in-place update path: re-scaffold, or apply the fixes by hand — delete the crons block from vercel.json, drop pg/@types/pg from package.json, update SmartContactForm to read profileCapabilities.supportTriage (or pass triageEnabled={false}), add || isAssetExempt(pathname) to the /api early return in proxy.ts, and (for the font trap) replace the next/font/google block in app/layout.tsx with a system stack. Light and full scaffolds are unaffected.

v0.55.0 — 2026-09-05

A product specification may finally be a number, a boolean or a list — and every surface agrees on what that means. The engine's own coffee seed (roast: 2, notes: ["bergamot","jasmine","lemon"]) could not be written through the admin the engine ships, and was silently absent from the product page on the way out. Three write paths disagreed about what a spec may hold, and the one an AI agent uses was the narrowest of them.

• The value rule, in one place (#557): string | number | boolean or a list

of those; nested objects and __proto__/constructor/prototype refused. lib/product-attributes.ts:normalizeAttributeMap is the single definition, and both the admin form (lib/validation.ts) and the agent surface (lib/tools/products.ts) call it. products.update inherits it through createShape.partial().

The agent surface previously declared attributes: z.record(z.string(), z.string()) with no reserved-key guard — so products.create was the only write path that could not store the shop's own seed, and the only one that would take a reserved key. Two measured details shaped the fix: z.record rebuilds its object by assignment, so "__proto__" set the prototype instead of creating an own key and vanished before any guard saw it (the prototype is never polluted, but the caller was told the write succeeded); and z.custom, which does see the raw object, is unrepresentable in JSON Schema — through the soft-failing converter it publishes an EMPTY attributes contract to /api/v1/tools?schema=true, MCP tools/list and the admin-chat tool table. The schema is z.preprocess over a z.record with an explicit value union, so an agent is told exactly what it may send.

• The product page renders them (#560): the PDP spec table formats numbers

and joins lists instead of filtering typeof v === "string", and the PDP's WebMCP descriptor carries the same flattened specs the ACP/Google catalog feed publishes — the on-page agent no longer knows less about the product than Gemini shopping does. WebMCP tool names, schema keys and moat bindings are untouched.

• `DesignPack.webshop.ownsAttributes` (#560, additive): the attributes a

design's own pdpLayout draws, which the shared spec table then skips — the ownsBreadcrumb idea applied to Product.attributes. It is a function of the product, not a key list, because a blanket list is only sound for a layout that draws unconditionally: crema shows a roast band only for values its parser accepts, so a declared-but-undrawn key would hide a fact that printed before. It also hands pack size to the variant picker once a product has variants. Unset → every attribute renders, byte-identical to before.

• `pnpm verify:design` can see below the fold (#555): --full-page captures

the whole scroll length and --selector <css> captures one component. The --selector mode announces that it skipped the page-level checks — an unannounced skip is indistinguishable from a pass. --full-page emulates prefers-reduced-motion: reduce, because this engine's scroll-driven reveals live inside (prefers-reduced-motion: no-preference) and a full-page capture does not scroll.

Upgrading

No schema change and no migration. A shop that stores attribute values the repeater cannot represent (nested objects, lists whose items carry commas) continues to edit them through the raw-JSON field, as before.

v0.54.0 — 2026-09-03

The judge's first minute. A cold `npx create-cartwright myshop --template coffee` now scaffolds the same store the live demo runs — nine products, real variants — and the README finally says so first.

The WebMCP surface shipped in v0.50–v0.51, but the scaffold undersold it: the coffee template seeded 3 products and zero ProductVariant rows, so the variant-aware half of add_current_product_to_cart (the variant enum, the natural-language labels, the per-variant stock lines) had nothing to show on a fresh shop. And the README never named the one command that produces an agent-testable store — the no-args default is a website profile with webMcp off.

• Seed variants (#549): SeedProduct grows an optional variants field —

one attribute key per PDP dropdown, agent-facing labels derived from attribute *values* ("Whole beans, 250 g") — and prisma/seed.ts creates the rows. The coffee template now seeds the live Northbound catalogue: 4 beans, 1 espresso and a new gear category (4 products carrying plain string attributes, deliberately outside the coffee vocabulary), with Colombia Supremo wearing the demo's three grind/size variants and everything else variant-less on purpose — an agent must handle both shapes.

• README: "Try the agent tools in 60 seconds" (#549): the exact command, a

prompt the seed data can satisfy, and the boundary that is the point — there is no order-placing tool; the checkout form is deliberately unannotated.

• Agent reads answer with URLs the shop serves (#544): site.list_pages /

site.get_page now clamp the locale and canonicalise legacy CMS slugs through the same helper llms.txt and the sitemap use, so every published URL routes and get_page resolves the alias set the storefront resolves.

• Flag-gated routes are 404 on every verb (#548): OPTIONS no longer

admits a gated surface exists when its flag is off.

• The shared storefront follows the reader's locale (#545): seven

shared-chrome links that dropped the /{locale} prefix now carry it, with labels to match.

• The header and its mobile drawer are one navigation again (#546): the

drawer renders the same page/category set as the desktop header instead of its own hardcoded list.

• Smoke telemetry tells "unreachable" from "wrong" (#547):

scripts/smoke-canaries.sh separates a dead read from a failed content assertion on every live-read path.

• Scaffold-context typecheck fix (#550, caught by the release gate): on a

single-locale scaffold brand.locales.find(...) narrows to plain undefined, so a test's as string cast could not compile — the fourth literal-type scaffold failure; the declaration now carries a load-bearing string | undefined annotation.

v0.53.1 — 2026-08-29

Scaffold-correctness patch. `brand.defaultLocale === "da"` compiles on the engine and cannot compile in your shop.

release-scaffold-gate vs v0.53.0: site ✅ · full ❌ · light ❌ — TS2367, a comparison between types with no overlap.

defaultLocale is a literal type, not string. The engine's is "da", and npx create-cartwright patches yours to "en", where === "da" is provably false and TypeScript refuses it. Three sites added in v0.53.0 — the two HTML pages served outside app/[locale], and the test asserting their shape — now widen the comparison explicitly, each with a note saying the cast is load-bearing rather than noise.

Nothing about the behaviour changes: the pages still serve Danish only when your defaultLocale says so.

v0.53.0 — 2026-08-29

The language sweep, and five feature toggles that did nothing. An audit of every string literal in the shipped tree, plus a trace of every runtime-tier flag. Most of what was found was legitimate; what was not is below.

Danish strings an English shop's customers and owners actually saw

The engine grew out of a Danish shop, and prose written for it reached shops that do not read Danish. Each of these had a traced render path:

• The contact form. /api/inquiries answered in Danish prose, and the

form's own t("errorGeneric") could never run — a truthy server string never falls through ||. Mistyping an email showed Ugyldig email-adresse. The endpoint now answers with a machine code the form translates.

• The AI assistant (aiStylist, default-on) rendered three Danish strings

to every visitor who added to cart through it.

• Two complete HTML pages — the newsletter unsubscribe page and the

supplier's mark-as-shipped page — were served with lang="da" hardcoded. They follow brand.defaultLocale now, with Danish as the special case.

• Password reset, a public route, passed the server's Danish through a page

that already translated its other messages.

• Values written into your database: a redaction marker ([slettet]), a

lead's budget default (Ukendt), and a review byline (Anonym). No later locale switch can reach those once written.

• The admin: the setup runbook, upload errors, the GDPR processor registry,

and the boot error printed when DATABASE_URL is wrong.

• Section defaults: adding a Hero in the Visual Builder wrote

"Din overskrift her" into your page. 67 strings across the studio sections and the section registry are English now.

Five runtime-tier flags were read from the compile-time config

/admin/features writes a database override; these surfaces ignored it:

• The weekly SEO cron answered 200 {"ok":true} and collected nothing on

every shop that had activated Plus — /admin/plus activates by writing an override and never edits brand.config.ts, so the compile-time value stayed false permanently. If you run Plus, your GEO history restarts from here.

• `/blog/feed.xml` 404'd on a blog whose pages rendered and whose posts were

listed in sitemap.xml.

• The wishlist: the product page showed the heart, /account/wishlist 404'd.
• `/api/registry` 404'd, so install counts never recorded.
• Cart recovery answered 200 and sent nothing.

Changed behaviour worth knowing about

• `messages/{da,en}.json` gained a `Chrome` namespace. If you maintain your

own message files, add it — a missing key renders the key itself as an aria-label. Keys: navExplore, navInformation, navFooter, navPrimary, emailPlaceholder, ctaFallback, newsletterHeading, newsletterBlurb, emailLabel, newsletterSubmit. The Login namespace gained ten magic* keys and Account gained two resetActions_* keys.

• **/api/inquiries, /api/contact/upload, /api/live/token and the review

endpoint answer with a code field.** error is still there and is now always English, for scripts and logs. If you render error directly, map code instead.

• A refused anonymous MCP batch is now a JSON-RPC error (-32600) rather

than a problem document. The refusal itself is unchanged.

• The GDPR redaction marker is `[redacted]`. Rows erased before this

release keep [slettet].

• A lead's `budget` defaults to `""` rather than Ukendt, so the admin

badge is omitted when the form did not ask.

• `contactAttachments` is resolved lazily in /api/inquiries — only when a

request actually carries attachments.

Guards

Seven new or widened test guards, each proven by breaking what it protects: storefront endpoints must answer in codes; admin-facing modules must be English; section defaults and design packs are scanned for Danish; shared components must resolve every key they name in every locale file; and a runtime-tier flag must never be *gated* on the static config — a static read stays legal as a default parameter or ?? fallback, which is how several flags are correctly wired.

v0.52.2 — 2026-08-29

Second scaffold-correctness patch. Three assertions failed a customer for choosing the `light` profile — for doing exactly what the profile promises.

release-scaffold-gate vs v0.52.1: site ✅ · full ✅ · light ❌.

A pruned profile removes design packs and dev-only scripts. Three tests asserted the ENGINE's full file set and therefore went red in a scaffold that had pruned correctly:

• site-pruned-scripts required every listed script to exist —

light prunes scripts/publish-agent-card.ts, which is the list's own job.

• design-copy-language read designs/hoptify/homepage.tsx — pruned.
• marketplace-manifest compared the rebuilt catalogue to the committed one —

a pruned scaffold legitimately has fewer designs and chromes.

All three are engine-only assertions, and they say so now, gated on .cartwright/profile.json — a file that exists in every scaffold and never in this repo. They keep their teeth where drift can actually be introduced: the engine, where CI runs.

The manifest one was not new to v0.52.x; it was latent in the shipped artefact and only surfaced when the gate was pointed at a fresh tag.

Verified: light scaffold from the tag — tsc 0, 3379 passed + 23 skipped. Engine — tsc 0, 3776 tests, lint 0.

v0.52.1 — 2026-08-29

Scaffold-correctness patch — found by the release gate against v0.52.0, the same way v0.51.1 was found.

• tests/unit/agentic-endpoint-4xx.test.ts destructured routing.locales by

index. On the single-locale ["en"] shop a scaffold produces, that tuple's TYPE is readonly ["en"], so index 1 is TS2493 — a compile error, and the light and full profiles would not typecheck (release-scaffold-gate vs v0.52.0: light ❌ · full ❌ · site ✅). Indexed access + find instead.

• Why the fork simulation missed it: it ran vitest against the scaffold shapes

and never tsc. A type-level fork-hostility is invisible to a passing test suite. Both are now run against all three shapes before a release.

v0.52.0 — 2026-08-29

The money-and-language release. A shop can now say its own words, and charge its own currency, in the language of the page it is on — and the tests can be run by a shop that is not Danish.

The trigger was one screenshot: DKK 149.00 on demo.cartwright.app/en. What that turned out to be resting on took four review rounds and forty-odd measured defects to reach the bottom of.

Language — the page a reader actually sees

• `<Price>` had a `locale` prop and not one of its four call sites passed it,

so every storefront price rendered in the CURRENCY's language: 23 occurrences of kr. measured on the live English PLP. It reads the page's locale itself now, so a fifth caller cannot reintroduce the bug by forgetting a prop.

• The <meta name="description"> on a secondary locale was the source

language — the one piece of copy on the page written FOR crawlers and agents. The root layout sits above [locale] and cannot know; the locale layout and the homepage's own builder now localise it.

• llms.txt and the public site.* tools took a locale, used it for their URLs,

and then published the base column's words. Extracted translatedField, which also fixed two latent faults: a translations column that arrived from raw SQL as TEXT never applied at all, and an own-property read now stops a key like constructor standing in for real copy.

• A shop's OWN words became translatable (brand.copyTranslations) — engine i18n

cannot reach a tagline that lives in brand.config. An explicit per-locale translation outranks the locale-blind Resolvable Genome, by specificity.

• A language-parity gate (scripts/smoke-canaries.sh): a Danish string on an

/en route now fails the canary run, with a 28-case self-test so the heuristic itself is measured rather than trusted.

Money — what the customer is charged

• The pay button showed one amount and Stripe charged another whenever

multiCurrency was on. Two independent derivations of one number; there is one now, computed once and handed to both.

• A partial refund in a non-base currency refunded the wrong magnitude —

Stripe reads the amount in the PaymentIntent's currency, so 149 sent against a €19.97 charge requested €149. Converted with the order's own snapshotted rate.

• The agent surface quoted the base currency by documented design — correct while

the flag was off, and self-contradictory the moment it was on: the page showed the presentment amount, Stripe charged it, and only the agent said otherwise.

• Multi-currency is config-sovereign: the charge currency cannot be flipped by

a DB row.

• Shipping copy is interpolated from the configured rate, so the prose and the

charge cannot drift. (The coffee demo promised 39 kr while its JSON-LD published 6.00 — a leftover from a spell as a USD store.)

Search — three doors, one rule

The storefront, the agent API and the semantic ranker each searched differently. Measured against the live catalogue before the fix: press returned three products on the storefront and zero through the agent API; кофе returned the ENTIRE catalogue as matches. They share one matcher now — substring, the same rule Prisma's contains applies — with the asymmetries that remain (folding, case, JSON attributes) written down rather than claimed away.

Fork-safety — the engine is not a Danish shop

tests/unit/ ships to every scaffold, and a dozen of its assertions named da, DKK or "en"-as-secondary as if the engine's own configuration were universal. Measured: a default single-locale `["en"]` scaffold failed 5 assertions; a USD scaffold failed 25. Both are zero now. Guards that must name an engine-internal path live in tests/engine/, which is mirror-excluded.

Also

• calculate_brew_ratio resolves its recipe to a real, buyable product —

read-only, and it prefers something the shop sells by weight so an embedded catalogue cannot answer "buy 1 × Paper Filters" to a coffee question.

• A negative smoke assertion that cannot read the page is now **inconclusive, not

clean**: does_not_contain piped curl into grep, so a dead origin printed a green tick while asserting nothing.

• ?q= with 250+ words threw P2029 before reaching the database — an

input-length cliff on a public GET, now deduped and capped.

Migration: none required. multiCurrency and currencySwitcher remain default-off; Order.currency/fxRate have existed since v0.15.0.

v0.51.1 — 2026-08-27

Scaffold-correctness patch — found by running the release gate against the tagged release itself.

• tests/unit/feature-flags.test.ts hardcoded brand.features.webMcp === false.

Since create-cartwright 2.8.0, webshop templates scaffold with the flag ON, so the assertion failed in exactly the scaffolds that are configured correctly (release-scaffold-gate vs v0.51.0: light ✅ · site ✅ · full ❌). The test now asserts the invariant that holds in every deployment — WebMCP only where ecommerceEnabled is true, the flag's declared precondition — mirroring the firstRunWelcome precedent in the same file. Gate re-run: light/site/full 3/3.

• Release coherence restored: package.json (which still read 0.50.0 at the

v0.51.0 tag) and the marketplace manifest now track the CHANGELOG heading.

v0.51.0 — 2026-08-27

The agent-experience sweep: spec-current WebMCP, honest discovery, an intent-first showcase — and webshop scaffolds go WebMCP-native.

• WebMCP tool correctness (audited against the late-August draft, Chrome's

best-practices doc and the reference commerce implementation): search_products returns reduced, bounded rows — checkout endpoints and internal fields never enter the browser surface (test-pinned); add_current_product_to_cart takes a natural-language variant (enum of option names) instead of an internal id, and omits the property entirely for variant-less products; update_cart_item_quantity is single-function (1–99; removal belongs to remove_cart_item) and both cart tools enumerate the live line ids in their schemas; ambiguous mid-request failures say so and route the agent through get_cart before any retry; descriptions lose their negative phrasing without losing the moat semantics.

• Declarative forms, both audiences: toolparamdescription types onto

<textarea>/<select> and the contact form's five controls carry English parameter descriptions; newsletter agent outcomes are stable English while the human keeps the localized message; :tool-form-active / :tool-submit-active styling makes agent activity on a form visible (engine-neutral outline + a copper glow in crema). The crema letter form — previously a urlencoded POST at a JSON-only endpoint — is a real client form with inline outcomes.

• `/webmcp-check` is an intent-first explorer: "I want to…" chips map

outcomes to tools with copyable example inputs; the page loads the Chrome team's WebMCP polyfill (Apache-2.0, vendored) as a page-scoped fallback so it demonstrates in any browser, and the live panel discloses polyfill vs native. Copy names the architecture: per-route registration is progressive disclosure.

• Discovery conformance (fresh-scanner + streamable-HTTP client trace):

/.well-known/ai-catalog.json emits the ARD manifest shape (specVersion + host + urn:air entries with per-entry trust manifests) while keeping the legacy resources view; the A2A agent card carries top-level name/description/version (and deliberately NO url — a conforming client would treat it as a JSON-RPC endpoint); /api/mcp answers explicit-SSE GETs with the spec's 405 and unknown tools/call with a JSON-RPC -32602 protocol error; pricing.md gains a real per-product price table that tells the buyable truth for variant products; llms.txt links the live /webmcp-check inventory when the flag is on.

• Eval-run fixes (GoogleChromeLabs webmcp-evals, 12/12 smoke steps green

against the live demo — and three genuine findings): the cart page's locale-less links 307'd through the i18n middleware and next/link prefetch retried the redirect in a loop (~5 req/s from one idle empty-cart visitor) — all four links now carry the route locale; agent-facing URLs are canonical; filter_products speaks English parameters (category/minPrice/ maxPrice/sort: newest|price-asc|price-desc) mapped to the route's query params, Danish input accepted leniently.

• Lifecycle proofs: the e2e fake modelContext now REJECTS duplicate

names and pre-aborted signals (current-spec semantics), and a new soft-navigation spec clicks an App Router link inside one document so the AbortSignal teardown is genuinely proven.

• WebMCP-native scaffolds (create-cartwright 2.8.0): webshop templates

(--template coffee | sunglasses | generic) scaffold with brand.features.webMcp: true — the store registers its tools the moment an agent-capable browser opens it; website-mode scaffolds keep the surface dormant behind the flag.

v0.50.1 — 2026-08-27

Scaffold-profile correctness patch — the four unit tests that assumed the full two-locale engine tree now derive their expectations from the scaffold's actual shape (hreflang alternates from brand.locales, the Crema keying pin applies only to registered packs, white-label checks skip pruned files), so create-cartwright scaffolds run the whole suite green out of the box. Also threads the route locale explicitly through dynamic translations — ambient getLocale() loses it in production rendering (#492).

v0.50.0 — 2026-08-27

WebMCP-native storefronts: the whole in-browser agent surface behind one flag — plus Crema site-wide, and a catalogue page the design finally owns.

• WebMCP foundation hardened (experimental, webMcp flag, default-off): tool

registration is now awaited per the draft spec and degrades per-tool; detection is shared through the new zero-import lib/model-context.ts (used by the registrar and the /webmcp-check diagnostics page, which now updates its tool list live via toolchange). Tool results carry verification data: get_cart and add_to_cart return the full agent-facing cart summary — line handles (cartItemId/productId/slug), stock ceilings, and money as {amountMinor, currency, formatted} instead of a currency-hardcoded subtotalDkk. Cart mutation actions return discriminated results ({ok:false, code:"not_found"} instead of a silent no-op on foreign ids); infrastructure errors still throw, so UI behavior with the flag off is unchanged.

• WebMCP tools are now contextual (experimental, webMcp flag, default-off): the

PDP registers add_current_product_to_cart for the product on the page (variants enumerated with ids, prices and stock in the description — and REQUIRED when the product has variants, so an agent cannot create the base-price line the human UI cannot create), and the cart page registers update_cart_item_quantity (0 removes), remove_cart_item and go_to_checkout — which only OPENS checkout for the human; no order-placing tool exists in the browser. The global registrar slims to search/get_cart/navigate. Per-page mounts gate server-side (null when off — byte-identical), carry their own AbortController so PDP→PDP navigation swaps tools cleanly, and re-register on RSC refresh so line ids stay fresh. The moat test now spans all three tool surfaces and requires globally unique names.

• WebMCP declarative forms (experimental, webMcp flag, default-off): the site

search, contact and newsletter forms carry toolname/tooldescription when the flag is on — WebMCP-enabled browsers surface them as agent tools, every other browser ignores the attributes, and flag-off HTML is byte-identical. Agent-invoked submits are answered through event.respondWith with the actual outcome (navigation target, triage answer/escalation, subscription result). Autosubmit only on search (read-only); contact and newsletter ask the human to confirm; the checkout form is deliberately unannotated (financial). Typed in types/webmcp-dom.d.ts.

• Cart lines now display the price checkout charges. Every cart-line display read

product.priceDkk even for variant lines, while checkout bills variant?.priceDkk ?? product.priceDkk — so a variant priced above its product showed one total and charged another. Fixed on the cart page (both render branches), the checkout review (both branches), the assistant's pre-purchase PlanCard total, the MCP cart.get_summary/discounts.try_apply tools, and the abandoned-cart email (whose query now loads the variant it prices by). The cart page's quantity ceiling is variant-aware too. A parity test scans the display surfaces so a bare product.priceDkk cart-line read cannot come back.

• Dependency currency, within-major — notably **Next.js 16.3.0 → 16.3.3, which includes

the August 2026 security release** (unauthenticated RCE via AVIF image optimization in sharp/libheif, plus a Windows-only mixed-router issue). This engine's default config does not enable AVIF optimization, so no shipped shop was exploitable through it — the bump closes the class. Also Prisma 7.10.0 (CLI + client + adapters), AI SDK 7.0.79 + providers, Stripe 22.5.0, Sentry 10.71.0, next-intl 4.13.7 and friends. One source change: the AI chat model handle is typed as the common LanguageModel contract instead of a provider-derived type.

• The catalogue page teaches the agent its own filters (experimental, webMcp

flag, default-off): the PLP registers list_visible_products — zero-network, returning exactly the server-narrowed list the human sees plus the active filter state — and filter_products, whose schema is SERVER-DERIVED (live category slugs as an enum, the sort whitelist, price bounds) and validated before navigating. After the RSC re-render the read tool re-registers with the new visible set. The moat gains the PLP surface; filter_products joins the navigation-only class.

• Design packs can ship their own WebMCP tools: new optional

DesignPack.webMcpToolBindings, aggregated by the moat test across every registered pack via the design registry — global name uniqueness and the cart/catalogue families now hold across packs, and a pack pruned by a CLI profile drops out of the aggregation automatically. First pack tool: crema's calculate_brew_ratio — the homepage brew calculator's math (extracted to brew-math.ts) typed for agents, same module as the human's widget, under the new review-gated PURE_COMPUTE class.

• /webmcp-check is now the agent-tools showcase: the full tool inventory grouped

by registering surface, built from the SAME binding constants the moat test verifies (a render test pins the loop closed), the safety-moat explanation in plain language, and setup steps for ChatGPT's built-in browser / the Chrome flag / the WebMCP Inspector — plus the live registration panel. built-with-cartwright's WebMCP blurb now describes the real architecture, and proof links render only when the feature is ON for the shop (a proof that 404s is the opposite of proof).

• WebMCP e2e harness: a Playwright spec installs a faithful fake

document.modelContext before any page script and walks home → catalogue → PDP → cart → checkout — contextual tool sets appear and disappear per route, one real add-to-cart round-trip mutates the dev DB, and checkout is asserted tool-free. Public architecture doc: docs/webmcp.md.

• Crema covers the whole site: applyPaletteAsTheme + a site-wide CremaShell

(.crema-site scope with sol pins, dark color-scheme and compensation for hardcoded light utilities) close the dark/light seam — cart, checkout, account, info pages and the PLP now follow the roast. The PLP gets an editorial crema frame ("The shelf") via a refined webshop.plpLayout contract that hands the whole page frame to the design. Every stylesheet is now parse-checked in CI (a prose comment once terminated itself and took down every page at PostCSS time).

• Shared chrome stops speaking for the brand: the footer newsletter strip renders

the brand's own copy on every chrome (the dark-chrome branch hardcoded Danish SaaS marketing), catalogue filters are value-driven (no empty eyewear dropdowns on non-eyewear shops) and fully localized, the mobile drawer is no longer the desktop nav's monolingual twin, and the AI assistant's inline product cards follow the visitor's locale with locale-prefixed links.

• Engine self-promotion off the merchant's presentation path: new

brand.website.showAuditFeed (default on) coherently gates the engine-changelog surfaces (footer link, announcement-bar link, the /changelog route, the sitemap entry); /cartwright (the engine marketing page) is website-mode-only; the capability manifest speaks English on llms.txt, /built-with-cartwright and /admin/features; the content-negotiated 404 page styles itself.

• Coffee template feeds the pack it ships with: seed products now carry

Product.attributes (origin/process/roast/notes/weight), threaded through both seed paths — crema's roast dots, origin badges and per-kg pricing light up from first seed.

• New design pack `crema` — a cinematic, dark-chrome premium DesignPack, i18n from birth,

with its own webshop overrides: an attribute-driven product card (roast dots, origin badge, process subtitle, tasting-note chips, an honest ≈ kr/kg only when the weight parses), its own PDP frame with a scoped token bridge, and a brew calculator. Its homepage bar reads product imagery through the media shim, plays a product's videoUrl as a card loop (hidden under prefers-reduced-motion), and can show a live agent-readiness score fetched server-side at render — never a stored number. It is module-claimed for the webshop profile and keyed by an invariant so no pack can be mis-keyed again.

• Attribute-driven merchandising, generalised. DesignProduct.attributes reaches the design

layer, and lib/product-attributes.ts merges translations.<locale>.attributes over the base so a spec table speaks the reader's language. Never guesses: an attribute that does not parse is not rendered.

• A storefront that stays in its locale. A config default can no longer pick the language for

a locale-routed storefront; the story-first pack's links carry locales and follow the canonical routes; and the shared footer — chrome on every page of every scaffold that does not replace it — now prefixes every locale-routed link it emits and reads its remaining labels, including the ones only screen readers hear, from the message catalogue. A source-scanning invariant, with its exemption predicate derived from lib/locale-exempt.ts, keeps it that way.

• Dark chrome paints its own text. The first dark-chrome webshop exposed that the shared

header's webshop leg had no dark variant: category links and cart/account icons rendered sol-ink on a black bar. Locked-dark packs do not bridge the palette, so dark chrome now paints explicitly — MobileMenu takes an optional darkChrome, default false. Light shops render visually unchanged, not byte-identically: the webshop nav's labels are now wrapped in a <span> that carries the dark classes only when the chrome is dark, so a light shop emits a wrapper span it did not before. Anything selecting those labels by DOM position should be checked.

• The `site`-profile footer stopped rendering an empty copyright line. brand.footer.disclaimer

now defaults to "" and the db footer falls back to a localized <storeName> · All rights reserved; its site-profile twin did not, and rendered the blank raw. Both twins are now held together by an invariant.

• The default webshop homepage reads its section copy from the catalogue. aurora-shop (the

webshop-mode default) and webshop-classic moved "Most popular" / "View all" to the Storefront namespace — so a non-English shop's default homepage changes text where it previously showed English. TrustBadges gained a localized fallback too, but only for shops that empty brand.uiLabels.trustBadgesPrimary; the shipped default is non-empty, so a default shop is unaffected by that half.

• Sunglasses-era card leftovers retired from the shared storefront card — every shop wore them.

No markup or class names were removed: the hardcoded Summer Edition featured badge and English Sold out now come from the ProductCard message namespace, and the Dannebrog origin pill renders only when brand.uiLabels.productCardOriginBadge is non-empty. Upgrade note: a shop that leaves that label empty stops rendering the pill (it used to render a flag with no text); set your own text to keep it, and note the English badge text itself changed from Summer Edition to Featured. AnnouncementBar's hardcoded English link is localized too.

> Upgrade note — message catalogues. This release adds 108 keys to messages/en.json and > messages/da.json. Most are pack-scoped (Crema.*, NorthernCoffee.*), but a set is read by > shared, non-pack components and required for a clean render: the whole new Catalog.* > namespace (PLP filters, sort options, product count), Footer.newsletterKicker, > Footer.allRightsReserved · Footer.homeAria · Footer.shop · ProductCard.featured · > ProductCard.soldOut · Storefront.mostPopular · Storefront.viewAll · TrustBadges.primary · > AnnouncementBar.aiLink. A fork that maintains its own catalogues instead of taking ours will > render raw keys for those until it adds them. No schema change ships in this release — no > pnpm db:push is needed.

v0.49.0 — 2026-08-24

Honest-ergonomics round 2: an accept-lenient MCP handshake and more places to find the agent view.

• The MCP POST handshake widens JSON-implying Accept shapes (json-alone, wildcard, absent) to

the canonical pair instead of answering the SDK's 406 — leniency, not indifference: a client asking for text/html still gets the 406.

• /auth.md opens with a leading heading (canonical + freshness in a footer line);

?mode=agent on the homepage serves the agent view; /{locale}/llms.txt is the section-level index (real 404 on unknown locales); llms.txt links are all resolvable locale-canonical URLs.

• The ai-catalog carries a trustManifest (real trust routes) at top level and per entry, and

OpenAPI's bearerAuth declares named per-scope grants with descriptions in a spec-legal extension.

v0.48.0 — 2026-08-24

Tiered MCP origin policy, the well-known agent card, and the card-publish flow.

• The MCP origin guard gains a deployment-aware tier: local/intranet keeps the strict

allowlist (the DNS-rebinding posture), while a genuinely public https deployment accepts well-formed foreign https origins — the cookie-less anonymous surface has no CSRF to protect, no CORS headers are sent, and TLS makes the rebinding case unreachable. Opaque origins still require their explicit opaque: entry everywhere.

• /.well-known/agent-card.json serves the A2A agent card at the ecosystem's canonical

discovery path (same gate and honest 503-until-published as /api/agent-card).

• scripts/publish-agent-card.ts is the publish flow the card libraries never had:

brand + live-catalogue payload, ephemeral ed25519 signing (no stored private key), transactional rotation. Site profile prunes it (db-coupled), mirrored in the CLI.

v0.47.0 — 2026-08-24

The honest-ergonomics round: markdown everywhere it helps, problem+json edge to edge.

• Unknown /api/* paths (and /api itself) answer RFC 9457 problem+json with a service-desc

Link instead of the HTML 404 page.

• /auth.md documents the real auth model (Bearer keys, named scopes, the anonymous five-tool

surface, rate limits) — and says explicitly that no OAuth flow is offered. /pricing.md (webshop-only) publishes currency, VAT treatment and the live price range.

• /index.md is a URL-suffix fallback for the markdown homepage; Accept-negotiated homepage

markdown opens with YAML frontmatter while bare /llms.txt stays byte-identical. Homepage responses carry markdown-alternate/sitemap/describedby Link headers.

• /.well-known/ai-catalog.json indexes the agent surface (profile-aware); the RFC 9727

api-catalog gains item entries; OpenAPI states a default problem response; MCP tool failures carry structured {code, status, message}.

• products.search gains offset pagination; categories.list gains a real input schema

(includeEmpty). New agent-surface routes are module-claimed so the site profile stays import-closed.

v0.46.0 — 2026-08-24

The agent-ready public surface: anonymous typed reads, OpenAPI 3.1, trust routes and honest discovery.

🤖 Agent-readiness

• Anonymous, rate-limited access is restricted to five public read-only tools: product search/get,

category list, and published-page list/get. Private reads and every write still require a scoped Bearer key; drafts, customers, orders and administration stay outside the anonymous boundary.

• MCP now publishes concrete input/output schemas and public resources, while generated OpenAPI

3.1 exposes one typed operation per tool. The legacy { "args": { ... } } MCP input shape remains accepted for one compatibility release.

• Agent-facing errors use Problem Details, public and pre-auth requests have separate rate limits,

client IPs are trusted only through configured ingress, and anonymous MCP batches are rejected.

• Runtime, locale-aware canonical/hreflang and social metadata now agree across HTML, llms.txt,

sitemap, trust pages and structured data. Markdown negotiation and recoverable 404 guidance are covered by contract tests.

📚 Adoption and release hygiene

• README quick start now distinguishes site, light and full, links directly to docs, demos,

Issues, support and contribution guidance, and includes a production-build check.

• GitHub issue forms, support routing and the published stable/main distinction are documented.
• Root and marketplace versions now track the newest stable release as a coherence check;

missing v0.43.0–v0.45.0 public release summaries are restored below.

v0.45.0 — 2026-08-22

Agent-surface and public-route hardening, an English admin, and the vermilion identity.

• Closed framework-level OPTIONS bypasses across MCP, ACP, UCP, A2A and merchant-feed surfaces;

the shipped guidance and profile-safe tests now preserve those route guards.

• Fixed locale-less route shadowing, login callbackUrl handling, canary asset/domain validation,

and light-profile test assumptions.

• Added bounded attribution for UCP OAuth client registration and one explicit opaque MCP origin

for operators that need it.

• Translated the admin panel to English and moved the engine identity from purple to vermilion.

v0.44.1 — 2026-08-08

MCP preflight guard patch. OPTIONS now passes through the same public-feature and Origin policy as the other MCP verbs instead of being answered by Next.js before the guard.

v0.44.0 — 2026-08-07

MCP Origin validation, Next.js 16.3, and stronger canary/profile contracts.

• Added the Streamable HTTP Origin validation required to prevent DNS-rebinding attacks.
• Upgraded Next.js to 16.3.0 and NextAuth to the security-patched beta.32.
• Hardened canary timeouts and customer-domain detection, normalised discount lookup at the

schema boundary, and made mixer/chrome errors identify the rejecting side.

• Removed shipped pointers to private-only files and made locked-design fixtures derive from the

active profile rather than one named design.

v0.43.0 — 2026-08-01

Identity ownership and destructive-seed safety.

• Made brand.config.ts the default identity authority and applied the same invariant to the

Solbrillen canary.

• Refused to seed over a database that already contains customer work, while adding a safe path to

create an operator on an existing production database.

• Fixed required lead fields receiving empty strings, discount length checks before trimming, and

mixer chrome compatibility so the active design pack owns the decision.

🔒 Security advisories

| ID | Affected versions | Fixed in | Severity | Action | |----|-------------------|----------|----------|--------| | CW-2026-001 | ≤ v0.39.1 | v0.40.0 | Moderate | Upgrade — or backport the mcpPublic route gate + the API-key expiresAt check (#367) | | CW-2026-002 | ≤ v0.40.0 | v0.41.0 | High (upstream) | Dependency advisory — Next.js. Every shop scaffolded at v0.40.0 or earlier ships a next version below 16.2.11 (16.2.10 at v0.39.1–v0.40.0, 16.2.9 at v0.38.0–v0.39.0, 16.2.6 at v0.37.1 and earlier) — every 16.2.x below 16.2.11 is affected, so check the number in your own package.json, not just your engine tag. Next.js' July 2026 security release patches 9 CVEs (4 high); two reach every Cartwright shop regardless of profile because every profile ships Server Actions: a Server-Action DoS (CVE-2026-64641) and unauthenticated disclosure of internal Server Function endpoint IDs (CVE-2026-64643). You do not have to wait for a Cartwright release — fix it in your own shop today: pnpm add next@16.2.11 && pnpm add -D eslint-config-next@16.2.11 (exact pins, matching the scaffold's dependency style) (#395) | | CW-2026-003 | ≤ v0.56.2 | v0.57.0 | High (upstream) | Dependency advisories. Fourteen pnpm audit --prod findings: thirteen below @modelcontextprotocol/sdk, next and the Prisma CLI (fast-uri, sharp, qs, hono, mysql2, deepmerge-ts) plus the direct js-yaml. Upgrade — or fix it in your own shop today: copy the overrides block from this repo's pnpm-workspace.yaml into yours, set "js-yaml": "^4.3.2", run pnpm install, and confirm with pnpm audit --prod that none of these fourteen remains (an older shop may still show earlier advisories — see CW-2026-002). If your `next` is still 16.2.x (engine ≤ v0.43.x), you lock sharp 0.34.5: no 0.34 release is patched and the sharp selector does not touch it, so move to next ≥ 16.3.3 first — pnpm add next@16.3.3 && pnpm add -D eslint-config-next@16.3.3, the scaffold's exact-pin style — and then copy the block (#583) |

v0.42.0 — 2026-07-25

Hardening release II: the visitor's path, and the difference between a deploy that works and one that reports success. Ten changes from the second and third rounds of findings by the same downstream product — this time from ordinary operation after go-live, and from setting production up. Round one's pattern was *one value, many encodings*. These two add: failures designed not to be noticed (the data is rescued so nobody feels the error; a correct configuration reads as empty) and a default setup that looks right without being it (a deploy reports ● Ready while serving nothing; a CI file sits in the repo without ever running).

✨ Added

• `pnpm verify:deploy <url>` (scripts/verify-deploy.sh) — the check a green deploy cannot

perform for you. A Vercel project whose Framework Preset is "Other" — what an empty CLI-created project gets — sets, per Vercel's docs, "the output directory as public if it exists". Cartwright *has* a public/, so the platform skips next build and serves your logos: every route 404s, the build log is empty, nothing errors, and the dashboard says Ready. The script asks for three routes the app generates (/robots.txt, /sitemap.xml, /llms.txt) — none exists as a file in public/, so a 200 proves the framework ran. Which routes is the whole design: public/mcp.json answers 200 in exactly the broken state, so a check aimed at any static file passes straight through the failure it exists to catch (#412).

• `brand.identitySovereignty: "auto" | "config" | "db"` — who owns storeName and

ecommerceEnabled, stated instead of assumed. "auto" (default) reproduces today's behaviour expression-for-expression, so it is byte-identical by construction; "config" makes brand.config.ts sovereign in every mode, which is what a code-configured fork needs; "db" is the deliberate multi-tenant escape hatch. Top-level rather than a features.* flag on purpose: FeatureKey parity would break forks on their own config, and a protection a contaminated database can switch off is not a protection (#408).

• `features.leadAiTriage` (default off) — gates the AI triage on the contact form. Off is

the honest default because ANTHROPIC_API_KEY is documented as optional; a surface that needs an optional key in order not to log an error per submission should not be on by default (#403).

🛠 Fixed

• The visitor waited on — and funded — an LLM. /api/inquiries awaited an Anthropic call

*before* prisma.lead.create, so every enquiry paid the round-trip in latency and, without a key, logged an error per submission while still saving. Triage now runs in after() from next/server after the row is written (the pattern lib/registry-stats.ts already used), gated on anthropicConfigured — not isAiConfigured(), which is true when *either* provider is set while the "vibe" model requires Anthropic specifically. Separately, /api/support/triage — a second, unauthenticated LLM call the same form makes — had no per-IP rate limit (#403).

• A contact form required agency fields. projectType was z.string().min(1) with a hardcoded

Danish message, so a plain enquiry from a hand-written form or a direct POST was rejected as the visitor's mistake. The shipped form hid this by always sending a value (#404).

• `defaultLocale` fell back to a literal `"da"` in the data-source seam, so an English fork

served /en/… while llms.txt announced Danish. It now derives from brand.config — proven on a *different* input than the current value, because a test comparing to today's value also passes against a hardcoded one (#405).

• The admin under-reported production. getResendStatus() read only the database row while

lib/mailer/resend.ts happily used RESEND_API_KEY from the environment — so a correctly configured fork saw "not set" on the one screen it was told to trust. Now matches the envFallback shape its two sibling status readers already had (#406).

• Identity could be read around its own lock. The Phase H guard was correct and well tested, and

a fork's site renamed itself anyway: Header, Footer, llms.txt, the sitepack exporter and the composition exporter read the raw BrandingSettings row rather than the guarded merge. The policy now normalises at the seam, and tests/unit/identity-bypass-invariant.test.ts fails the build if any file queries the row and reads a sovereign field off it — because the readers that matter are the ones nobody has written yet (#408, #409).

• The admin stored values it would never render. Under "config", the settings form, the setup

wizard and settings.update_branding now drop the sovereign fields from the write and say which ones, instead of accepting input, reporting "Settings saved!" and changing nothing. The fields render as locked, and when the stored and effective values differ the form says so in as many words (#411).

• `DEPLOY.md` contained a false instruction. §9 read "Vercel auto-deploys from main" directly

under a §1 recommending vercel link — but linking a *folder* to a project is not connecting the *project* to a repository, and pushes deploy nothing until vercel git connect, which itself requires a remote. §1 now leads with the repo; §9 states what vercel --prod actually does: uploads the working tree, including uncommitted changes, with no commit attribution and no rollback to a known state. Also documents the vercel env pull traps — it overwrites rather than merges, and sensitive values come back empty, so empty ≠ missing (#407, #412).

🧹 Housekeeping

• `scripts/smoke-canaries.sh` no longer ships to scaffolds. It hardcodes three Teloz canary URLs

and asserts their brand names, themes and database state; a fork running it learned nothing about its own site. scripts/verify-deploy.sh is the customer-facing equivalent — any URL, no prior knowledge — and the two shipped files that told the reader to run the canary script now point there instead (#412).

• Root `package.json` version rejoins CHANGELOG.md and the marketplace manifest at 0.42.0; the

v0.41.0 release bumped the other two and left it at 0.40.0.

v0.41.0 — 2026-07-25

Hardening release: import-time side effects, identity ownership, and the Next.js July security patch. Eleven changes driven by findings from a downstream product's first production deployment (a code-configured fork — brand.config.ts in git, admin used only for operations). That usage is exactly what Cartwright markets itself for, and it exposed a systematic assumption: the engine treated the admin database as the source of truth, so wherever code and database disagreed, the database won silently.

🔒 Security

• Next.js 16.2.10 → 16.2.11 — the July 2026 security release, 9 CVEs (4 high). Two reach every

Cartwright shop regardless of profile because every profile ships Server Actions: a Server-Action DoS (CVE-2026-64641) and unauthenticated disclosure of internal Server Function endpoint IDs (CVE-2026-64643). Tracked as CW-2026-002; eslint-config-next moves in lockstep. Per-CVE applicability was verified against this repo rather than assumed — the other six need config.i18n.locales, rewrites()/redirects(), a custom server, the edge runtime, or a fetch(new Request(init), otherInit) shape, none of which the engine uses (#395).

• **.gitignore now covers .admin-credentials*** — the bare name matched only that exact file, so

a sibling like .admin-credentials.prod (what a production-admin bootstrap writes) was commit-eligible. A !.admin-credentials.example negation keeps the convention's stub committable (#399).

⚠️ Breaking

• `sanitizeUserHtml()` is now async (lib/v0/transform/sanitize-strict.ts), and with it

parseSitePackContent, sanitizeContentRow, sanitizePageRow and sanitizeLayoutJson (lib/sitepack/import-parse.ts). If your fork calls any of them directly, add await; every in-engine call site was already async. The reason is not cosmetic: jsdom is now loaded on first use rather than imported at module scope, and a dynamic import cannot be synchronous. See below (#397).

🛠 Fixed

• A feature toggle could rename a live site. Nineteen BrandingSettings create: branches

across fourteen files hardcoded storeName: "Cartwright" (and "My store" / "Min shop" in the wizard's own writers). That branch runs only when the row does not exist — the normal state of a fork configured in code — so the first arbitrary admin action stamped the engine's name onto the product: header, footer, and llms.txt, the file AI crawlers read. All nineteen now read brand.storeName through brandingCreateDefaults() (#398).

• Importing the HTML sanitizer loaded jsdom. sanitize-strict.ts imported and constructed jsdom

at module scope, and it is reachable from lib/tools/registry.ts — a barrel that statically imports ~20 tool modules and is imported by 14 files, including /admin/audit, which wants only getTool(). jsdom's CJS entry requires html-encoding-sniffer, which depends on the ESM-only @exodus/bytes; under some bundler/runtime combinations that boundary throws ERR_REQUIRE_ESM at module load and takes the whole page down, a subsystem away from anything to do with HTML (#397).

• Every page view loaded 29 design packs to bust a cache. lib/theme.ts imports @/designs

(29 static pack imports, ~1.4 MB, unshakeable because ALL_DESIGNS references each one), and 31 files import lib/theme — including app/[locale]/layout.tsx, the storefront root layout. Eight writers imported it purely for the three-line invalidateThemeCache, and a plugin purely for the two-line isValidHex. Both now live in leaf modules (lib/theme-cache.ts, lib/color.ts) that lib/theme.ts re-exports, so the documented plugin contract is unchanged. This also explains two tests long written off as "flaky under load": audit-revert 2.58 s → 284 ms, design-import 2.00 s → 240 ms (#401).

• `vercel.json` pins `"framework": "nextjs"`. A project created through the Vercel CLI defaults

to preset "Other", which serves public/ statically — every route 404s and the build takes 0 ms, with nothing that looks wrong (#399).

✨ Added

• A lint rule against import-time side effects — no-module-scope-heavy-construction forbids

constructing jsdom/Prisma/Redis/OpenAI/Anthropic/Resend/Stripe/S3 clients at module scope, where a barrel turns one module's cost into every importer's problem. The rule is written into AGENTS.md, so scaffolds inherit it (#400).

• Module/profile dependency inventories — import-grounded per-module dependency sets and derived

prune sets, with five invariants (#391).

🧪 Tests

• New invariants that fail the build rather than relying on discipline: no writer may hardcode a

store name (source scan, quote-agnostic, with a reasoned allowlist and a spread-precedence check), jsdom must not be loaded at import, leaf modules must stay leaves, credential globs are asserted through git check-ignore itself, and every vercel.json cron must resolve to a real route (#398, #397, #401, #399).

• The /api/mcp tool bridge is locked: 1:1 tool registration with no scope pre-filtering, isError

on failed calls, scope delegation, audit context, and the stateless-JSON transport (#396).

• Write-tool coverage for lib/tools/discounts.ts and lib/tools/marketing.ts (#392, #393).

📚 Docs

• The section-vocabulary skill mirrors all 27 registry section keys (was 20) (#394).

v0.40.0 — 2026-07-19

The site profile — a real site with no database — plus Plus activation and the public-surface lockdown. The engine now describes itself as modules (cartwright-module-v1), proves the site file-set import-closed, and scaffolds a static --profile site variant; Cartwright Plus activates via Ed25519-signed keys; and the public MCP/tool surface finally honors its own flag.

🔒 Security

• `mcpPublic` is now enforced on the public tool surface — /api/mcp and /api/v1/tools

previously answered even with the flag off; they now return 404, indistinguishable from a site without the surface. `ApiKey.expiresAt` is now enforced — expired keys were previously still accepted; they now get 401 (#367, follow-up #377). Tracked as CW-2026-001 in the advisory index. The flag defaults on (and is on in every shipped config), so only shops that explicitly disabled mcpPublic and/or relied on API-key expiry are affected. Upgrade, or backport lib/tools/public-gate.ts + the lib/api-auth.ts expiry check.

✨ Added

• The module/profile architecture (site-profile B-phase) — the cartwright-module-v1 spec +

module/profile registry with graph tests (#374), per-module file inventories + the agent-core node + a knownDeviations ledger (#381), and the B3 engine slice: static design-layer variants, 18 import seams and scaffold/manifest.json, proving the site file-set import-closed so create-cartwright --profile site cuts a no-DB static site (#382, #383, #384, #385).

• Cartwright Plus activation v1 — Ed25519 license-key verification + /admin/plus, env-only

and schema-free (#380).

• Adaptive-admin slice — the four always-rendered AI admin pages are now flag-gated (#375).
• The mixer preview surfaces non-mixable Skin×Voice combos instead of silently degrading (#388).

🛠 Fixed

• /api/acp/feed gates on ecommerceEnabled — website-mode parity with the other feeds (#370).
• Regenerated the drifted migration baseline so prisma migrate deploy works, guarded by a new

db:verify CI gate (#379).

• Version coherence restored — CHANGELOG entries for v0.38.0–v0.39.1 + package.json/manifest

version alignment (#378).

• The site-audit walker falls back to an fs-walk when the tree has no git repo (#386).

🛡 Hardened

• Tool suites locked: the feature-flag read/write tools (#387), the taxonomy upsert/delete/list

tools (#366), the ACP session→Order money path (#365) and the settings singleton write tools (#389).

📚 Docs & launch

• Tool map regenerated (86 tools / 35 domains) + drift gate + the simple-site runbook (#371);

launch-content fact-refresh (#376); the "Choose Cartwright when" decision matrix + shadcn try-one-component (#372); the README terminal cast (#373); community badges, issue/PR templates + Code of Conduct (#368); the 2026-07 radar delta (#369).

v0.39.1 — 2026-07-12

Hardening patch — the protect-live test wave + dependency currency.

🛡 Hardened

• Moat + money-path regression tests: the ACP delegated-payment completion route (#358),

the Guardian adjudication core (fail-closed, replay, short-circuit order — #359), the escrow fund-release route (#362) and the audited order-status tool + operator state-machine (#363) are now all pinned by non-vacuous unit suites (source-mutation-verified).

🛠 Maintenance

• Within-range dependency currency batch + postcss security override (#361).
• Non-blocking native typecheck script (typecheck:native, tsgo trial) (#360).
• Byte-identical a11y recipes added to the cartwright-guidance skill (#364).

v0.39.0 — 2026-07-07

AI-crawler policy, declarative design layout, and honest agentic error handling.

✨ Added

• AI-crawler taxonomy in the SEO controls — Search/Agent/Training crawler categories with a

new "block training only" policy option (#355).

• `DesignPack.layout` — design packs declaratively control the shared <main> wrapper

(mainClassName / ownsMain) instead of fighting it (#351).

🛠 Fixed

• Agentic endpoints (ACP/A2A/tools) return proper 4xx for client errors — never a masking 500 (#352).
• Ember chrome reads the isEcommerce predicate; Button has a visible focus ring (#353).

🛡 Hardened

• Discovery route handlers pinned by tests: /api/agent-card (#348), /.well-known/mcp.json

(#349), both product-feed routes (#350) and POST /api/negotiate (#357).

v0.38.0 — 2026-07-03

Agentic discovery goes spec-true, the checkout speaks your language, and the stack gets a currency pass.

✨ Added

• ACP product feed pinned to the official OpenAI schema — discovery-ready field mapping

(item_id, price strings, availability enums, seller/policy URLs) (#340).

• Google Merchant `g:product_detail` conversational attributes in the Merchant feed (#342).
• Speculation Rules prefetch behind a new default-off speculationRules flag (#341).
• 2026-07 feature radar (agentic commerce, Next 16.x, web platform) as the roadmap driver (#339).

🌍 i18n

• The entire cart + checkout flow routed through next-intl: payment panel, trust badges,

field-validation errors and order-placement errors (via error codes) (#337, #338, #344, #345).

🛠 Fixed

• Tool/section JSON-Schemas now emit real schemas via Zod v4 z.toJSONSchema (#324).
• Fresh-scaffold DX hardening from the v0.37.1 cold-run audit (#323).

🛡 Hardened

• Dependency currency + undici security override (#332); AI SDK v7 migration (#336).
• +49 regression cases across shipping/pricing/FX/genome-resolve (#333); catalog-feed builder

+ UCP capability-profile invariants pinned (#346, #347).

• Product-aware accessible names for cart/add-to-cart/wishlist controls (#322, #334, #335).

v0.37.1 — 2026-06-16

Cleanup patch — three finished, gated-green PRs that missed the v0.37.0 cut.

🔒 Security

• Optional shared-secret guard on the Phone.inc webhook (plugins/phone-widget) — the

click-to-call webhook can now require a shared secret, so a shop running the phone widget can reject unauthenticated webhook calls. Opt-in (set the secret) and the plugin is default-off, so this is hardening rather than a forced critical fix — no advisory-index row.

✨ Added

• Flag-gated Preview link on the Design picker cards (/admin/designs) — each design card

can surface a Preview link to see the look before applying it (default-off).

🛠 Fixed

• PLP empty-results heading now echoes the searched term, so a no-results search reads as

a search for «that term» instead of a generic empty state.

v0.37.0 — 2026-06-15

SitePack portability, content-import, the purple rebrand — and a new-customer hardening wave. A whole site is now exportable and restorable as a .cartpack, an AI can rebuild a real site from a scraped URL, the engine wears its own Cartwright-purple identity, and the cold-scaffold first hour was audited end-to-end (4 personas + Codex/Gemini) with the breaks fixed at the source.

✨ Added

• SitePack (sitepack.export / sitepack.import + the /admin/sitepacks Snapshot & Restore

wizard) — snapshot a whole site to a hardened, integrity-checked .cartpack (sha256 + merkle, red-teamed in-memory tar codec, content-safety + compat gate on import) and restore it onto a newer engine. The code-vs-data boundary that makes a site portable.

• Content-import — scrape a URL → rebuild it as Cartwright pages/services as drafts

(draft|published status on Page + Service), with a page classifier, dry-run import planner, SSRF-guarded remote-image import, and the runImport orchestrator. Flag-gated (siteImport).

• Blog + services + image tool surface — posts.create/update/publish/list,

services.create, and images.import_from_url over the typed REST tools.

• Cartwright purple rebrand — the glassmorphic purple first-run canvas, a matching

built-with-cartwright, and the engine identity surfaces repointed to the --cw-brand token (customer palettes untouched).

• Annotate — the in-place "edit on the live storefront" experience (glass panel, anchored

popover) with edit-hooks in the flagship designs + direct field (price) editing.

• Ember premium design pack + dentist / restaurant / fitness voice presets.
• Storefront SEO — visible breadcrumbs (category / PLP / services / blog, flag-gated) plus

Blog / Service / BreadcrumbList JSON-LD.

🛠 Fixed — the new-customer hardening wave

• Website-mode rendered as a webshop — a website-corporate scaffold seeded

BrandingSettings.ecommerceEnabled = true (the schema default), so the corporate site shipped a cart, product search, shipping badges and a demo banner. The seed now persists brand.ecommerceEnabled, getBrand() forces website-mode false at render (#285), and the AI-bootstrap write path no longer hardcodes shop-on (#317).

• First-login redirect loop between /admin/setup and /admin/konto (#276).
• Footer linked about/faq to slugs the template never seeded → 404s (#277).
• Danish leaks in the seed banner, first-run post-splash, and account UI on English

scaffolds (#278 / #284 / #281); og:locale now derived from the route locale (#274).

• Single-locale shops shipped a broken language switcher — a locales:["en"] site rendered a

🇩🇰 DA toggle that routed /da → 404; the switcher now hides at one locale and maps options from brand.locales (#317).

• Duplicate industry options in the setup wizard, from the legacy saas/eyewear alias

slugs (#317).

• AI tool surface — error bodies now carry ok:false across every shape, features.set error

copy is English, and llms.txt reports the actual locale rather than the country code (#317).

• a11y — the first-run copy button no longer puts aria-live on the interactive control (#317).

v0.36.3 — 2026-06-12

The AI-agent hardening release. Three external AIs (Claude, Codex, Gemini) each built a site on the engine and surfaced real gaps; a full parity/security audit surfaced more. All fixed at the source — the scaffold is now self-prompting, AI-build-friendly, and the critical security drift is closed.

🔒 Security

• Auth-gated three admin AI endpoints (/api/admin/translate, /api/admin/generate-logo,

the phone plugin handler) that were internet-reachable with no session check — an open, billed LLM proxy on every deployed shop. New shared requireAdminApi() (401 for APIs) + a regression test that walks every app/api/admin/** route and asserts a guard, so the class can't recur (25 routes covered).

• Security headers in next.config.ts: nosniff, X-Frame-Options, Referrer-Policy, HSTS,

Permissions-Policy; CSP shipped Report-Only.

• Login brute-force throttle (lib/auth/login-throttle.ts): per-IP + per-email

token-bucket before any bcrypt work, timing-oracle-safe.

✨ Added

• DESIGN.md — the self-prompting design playbook every agent file points to: the three

build paths, the built-ins inventory (ThreeHero + 9 scenes, svg-items, the verified GSAP recipe), 8 taste rules, the screenshot self-verification rule, and the engine-overlay design hooks. GEMINI.md upgraded from pointers to full content.

• Blank Canvas motion examples (designs/blank/examples/*.example) + stable

data-cw-* design hooks on every floating engine overlay (AI assistant, phone, voice, consent, welcome, sticky-ATC).

🛠 Fixed

• pnpm typecheck script the docs promised; threeD.scene typed as SceneId (no more

stale 4-scene union); ThreeHero className merges with the full-bleed default instead of replacing it; any next/font name resolves in tests; an ESLint rule catches event handlers in Server Components before runtime.

v0.36.2 — 2026-06-11

Patch: en-only scaffolds typecheck.

🛠 Fixed

• proxy.ts hardcoded a "da" | "en" locale cast, so the English-only scaffolds that

create-cartwright now produces failed tsc out of the box (found live by a customer's AI, which had to patch the template before building). The cast is now derived from the configured routing.locales.

v0.36.1 — 2026-06-11

Patch+: the Blank Canvas and the mockup-first flow.

✨ Added

• Blank Canvas (designs/blank/) — a registered, deliberately bare design made to be

rewritten: ask your AI to "build a completely new design — header, footer, every page, totally unique" by editing one folder, while cart, checkout, admin and the AI tools keep working underneath. The documented path lives in AGENTS.md + the project briefing.

• Mockup-first flow — mockup.set / mockup.clear tools: an AI generates a raw HTML

mockup in seconds and publishes it as the homepage takeover for instant customer approval, then implements it for real (Blank Canvas or governed sections) and clears the scaffold. Reuses the vibe sanitize policy; confirm-gated + audited.

• AGENTS.md now mirrors "Your first 10 minutes" + "Motion & animation" so Codex/Gemini/

Copilot agents find the fast paths, not only Claude.

🛠 Fixed

• The first-run welcome canvas now also retires when a design is chosen via

brand.designSlug in config (the documented Blank Canvas activation).

v0.36.0 — 2026-06-11

The first impression: a welcome canvas + English-first scaffolds. A fresh scaffold's very first render is now a designed moment instead of leftover template copy.

✨ Added

• First-run Welcome Canvas (flag firstRunWelcome, default off — scaffolds enable it):

an untouched site greets you with "Your site was just born." over a palette-adaptive aurora (3D Live Canvas when enabled) and three paths — a copy-paste AI quick-start prompt, the guided /admin/setup wizard, and a signature-motif gallery of the design catalogue. It vanishes permanently the moment the site becomes yours (setup completed, a look chosen, copy applied, a vibe page published or a product created).

• Config-driven footer owner line — the "owned and operated by" line now reads

legalName + footer.ownerUrl/githubUrl from brand config (translatable via Footer.ownedBy) instead of a hardcoded string.

🛠 Fixed

• Paired create-cartwright release scaffolds English-first (/en, en-only locales),

replaces the leftover studio hero copy with "Welcome to {your store}", and seeds setupComplete: false so the first admin login actually opens the setup wizard.

v0.35.1 — 2026-06-11

Patch: dark mode fixed completely. A fresh scaffold no longer half-flips when the visitor's OS is in dark mode.

🛠 Fixed

• Dark-mode split-brain — Tailwind's dark: utilities followed the OS media query

while the theme-token overrides followed the .dark class, so OS-dark rendered a broken half-flip (dark chrome/backgrounds on a light design). One switch now: a @custom-variant dark ties every dark: utility to the same .dark class (guard-tested). The admin is decoupled and ALWAYS the light Polaris skin — the theme toggle no longer lives in the admin top bar, and toggling can never restyle the storefront from the backend. Storefront dark returns later as a per-design opt-in.

• Phone widget palette — the floating call bubble was hardcoded blue; it now follows

the shop's accent like every other storefront element.

• package.json now declares the MIT license (matches the LICENSE file from v0.35.0).

v0.35.0 — 2026-06-11

The Light release: the build engine AIs reach for. Cartwright is now told — and built — around one sentence: *a real site with design, database and backend, live in minutes.* create-cartwright defaults to a lean light profile (website-mode, 8 curated designs, pruned agent-marketplace/UCP/WebMCP/hoptify; --profile full keeps everything), heavy modules are packaged as plugins under a new cartwright-plugin-v1 contract, the Mixer became a complete composition system (chrome parts, downloadable/installable compositions, streamed magic builds, design-adaptive store pages), and the AI onboarding path is measured: a cold agent goes scaffold → designed, verified homepage in 99 seconds. Everything additive or flag-gated — canaries byte-identical throughout.

✨ Added

• Mixer 2.0 — the complete composition system:

- Chrome registry (lib/builder/chrome-registry.tsx + client-safe catalogue): headers/footers are selectable parts — 14 design chromes + 4 neutral cw-* parts, two-sided mixability, persisted in BrandingSettings.chromeJson (run pnpm db:push). - Compositions (cartwright-composition-v1, lib/compositions/) — a whole look (skin + palette + voice + chrome + scene + homepage layout) as a downloadable, uploadable artifact: admin export/import with dry-run preview, atomic composition.apply tool, every Look in the manifest ships its installable composition. - Magic speed — section generation parallelized (~5× faster), SSE-streamed progressive build UI, and an instant 0-LLM preset path (lib/magic/presets.ts). - `designSurfaces` (default-off) — cart, checkout, account, blog, services and more adopt the active design's tokens, display font and chrome; flag-off byte-identical.

• Plugin system (cartwright-plugin-v1, lib/plugins/spec.ts + plugins/) — optional

engine modules as manifest-declared in-repo plugins with install state and an admin API (/api/admin/plugins). First five extracted with zero-breakage re-export shims: phone-widget, wishlist, blog, reviews, three-scenes (the entire Live Canvas 3D system — the engine's heaviest client dep — is now optional). Manifest v3 carries the plugin catalogue.

• Scaffold profiles — npx create-cartwright --profile light|full with **light as the

default** (pairs with the create-cartwright release that bumps to this tag); the scaffolder records its profile in .cartwright/profile.json.

• AI onboarding, measured — the scaffold briefing leads with **"Your first 10

minutes": a verified terminal-only path (boot → agent API key → `magic.compose_look` via `POST /api/v1/tools` → curl-verify) measured at 99 s cold. Plus a "Motion & animation"** section: native motion presets, three.js scenes, and a live-verified GSAP recipe (SSR-safe, reduced-motion-guarded; gsap stays out of engine deps).

• Agentic discovery in `llms.txt` — flag-gated links to /api/acp/feed + ACP checkout

sessions (features.acp), /api/agent-card (features.a2a), and the always-on /.well-known/mcp.json; new byte-identity test pins flag-off output.

• Service + BreadcrumbList JSON-LD on website-mode service pages.
• LICENSE — the engine is now formally MIT-licensed.

🛠 Fixed

• Light scaffolds boot — the chrome registry kept a parallel static import list of

design chrome modules, so a pruned (light) scaffold 500'd on every page with "Module not found". Both catalogue and registry now derive from the design registry (getDesign(slug).siteChrome), and a new deep-import guard test makes the regression class impossible to reintroduce.

• Docs truth pass — removed false webVitals/passkeys claims (features that never

existed) and several phantom file references from every agent-rules file and the cartwright-guidance skill; docs now lead with the product sentence.

• Removed two proven-orphan dependencies (@ai-sdk/openai, ts-node).

v0.34.0 — 2026-06-10

The metamorphosis release: FABLE, Stillwater & the unified design language. Built the day Claude Fable 5 launched — and largely *by* it. Two new flagship designs, a library of hand-crafted (and animated) SVG items that doubles as the cross-design ornament language, unique signature-motif chrome for every premium pack, a public Skin × Voice Mixer, and the infrastructure that makes catalog growth ~free: one manifest as the single source of truth for every catalog, plus a committed capture pipeline for previews. Additive and default-off end to end (canaries byte-identical).

✨ Added

• FABLE (designs/fable/) — the website-mode flagship: an instanced 3D butterflies

Live-Canvas scene (procedural wings, GPU-only motion, pointer scatter, reading-clearing), Fraunces display hero, scroll-cinema metamorphosis timeline, safeguards story, own chrome. Ships with a matching fable Voice preset.

• Stillwater (designs/stillwater/) — calm-enterprise flagship: generative layered SVG

ridgelines across four times of day, the palette-reactive waves scene as hero water, oversized metrics, night panel with incident timeline, own ridgeline chrome.

• SVG item library (components/svg-items/) — 21 hand-authored palette-adaptive pieces

(marks · dividers · illustrations), 9 of them CSS-animated (reduced-motion safe), all installable via the public component registry (svg-<slug>).

• Unified design language (docs/design-language.md) — the three shared languages

(tokens · signature motifs · motion) + DESIGN_MOTIFS: every premium pack now carries a signature SVG motif used in its chrome, dividers and gallery badge.

• Own chrome for all premium packs — apex, studio, engineered, nocturne, meridian,

editorial-ink, brutalist and jungle join halo/flux/drive/aerospace/fable/stillwater with design-matched signature-motif headers/footers via DesignPack.siteChrome (webshop packs keep cart/account nav).

• marketplace-manifest v2 — the manifest now carries *every* catalog (designs incl.

motif + chrome flags, voices, scenes, svg-items with server-rendered markup, elements, looks); cartwright.app derives all its galleries from it, so drift is impossible.

• Gallery capture pipeline (pnpm capture:gallery) — deterministic preview jpg/video

capture for any design (error-page detection, video trim, self-managed dev server).

• `registryStats` (flag, default-off) — anonymous per-item install counting on the

component registry + /admin/registry-stats readout (run pnpm db:push before enabling).

• Admin Page Mixer studio (/admin/mixer, behind mixerPreviewEnabled) — read-only

Skin × Voice preview. The public Mixer lives on cartwright.app/mixer.

🛠 Fixed

• getActiveDesign() now honours the brand.designSlug config override (config > DB >

inference) — design-owned chrome previously didn't engage for config-selected designs.

• Registry-wide tool-scope invariants + website-mode identity-lock + Stripe subscription

sync state-machine regression tests (suite now 1127 tests).

• picsum.photos added to image remotePatterns (dev seed data uses it).

v0.33.0 — 2026-06-09

The Page Mixer + the super-pro flagship. Content and design are orthogonal in Cartwright, so this release lets a shop mix a vertical Voice (børnehave, tømrer, café…) with any palette-adaptive Skin and compose the page from swappable Parts — plus a much deeper premium-design layer: a design now owns *every* page (not just the homepage), four breakthrough Pro elements, per-design webshop overrides, and Apex, a flagship super-pro storefront that composes all of it on one palette-adaptive page. Additive and default-off end to end (canaries byte-identical).

✨ Added

• Voice layer (genome). The homepage hero + section copy now resolves through the

Resolvable Genome (readField), so a Voice can re-tone the page; studio + three website skins are Voice-aware, and the value-prop and feature *cards* are voiceable via genome list fields. Gated by genomeResolve; every anchor is the current brand.website.* value, so flag-off ⇒ byte-identical.

• Vertical / Voice presets. A new verticals/ registry where a Voice carries the full

Vibe — identity anchors + genome copy overrides + palette + 3D scene (+ optional seed/ layout) — applied idempotently from /admin/verticals. Flag verticalPresets (default-off).

• The Page Mixer (`mixer-preview`). A gated route that renders any Skin × Voice

composition in the real storefront layout — resolved ephemerally in-memory (no DB write), always noindex, double-gated behind mixerPreviewEnabled (default-off ⇒ canaries 404 it).

• Parts catalog. Three premium Page-Mixer parts + a mixable flag on DesignPack so the

mixer greys out Parts on non-palette-adaptive skins.

• 3D scenes. Three new palette-reactive Live-Canvas scenes (waves, orb, gridflow) +

a scene-preview surface (all inherit WebGL2 / reduced-motion / save-data gating).

• Premium designs. Four recognizable-aesthetic packs — aerospace (cinematic deep-tech),

halo (minimal product luxury), flux (vibrant gradient SaaS), drive (full-bleed automotive) — plus jungle, a friendly organic palette-adaptive website skin.

• Shell model — a design owns every page. DesignPack.siteChrome (Shell/Header/Footer) +

pages (contact / info / 404) seams let a premium design span the whole site, not just the homepage; halo/flux/drive converted, with info (FAQ/legal) + 404 templates for all four. Default-safe: an undefined seam falls back to the shared chrome / default body.

• Pro Parts (breakthrough elements). A "Build your own" configurator (live preview +

live price, pure :has(:checked), no JS), a scroll-cinema story (animation-timeline: view()), a 3D product showroom, and a before/after compare slider — surfaced as builder Parts, gated by cartwrightPlus (honor-system).

• Webshop overrides. DesignPack.webshop = a bespoke productCard + pdpLayout per

design (threaded via ProductGrid card? and a PDP wrapper); halo ships both.

• Apex — the flagship super-pro design. One palette-adaptive webshop homepage that composes

every breakthrough above (3D aurora hero, 3D showroom, value props, configurator, featured grid, scroll-cinema, CTA). Via applyPaletteAsTheme, every section and every Pro element re-skins to the brand palette — a $100k-feeling storefront the day it ships.

🛡️ Safety

• Everything is additive and default-off: the new flags (verticalPresets,

mixerPreviewEnabled) default false, every new design is registered but active nowhere, and genome anchors equal the existing copy — so scripts/smoke-canaries.sh stays 3/3 byte-identical. CSS-only Pro elements sit behind @supports + prefers-reduced-motion; the 3D scenes self-gate on WebGL2 / reduced-motion / save-data.

v0.32.0 — 2026-06-07

Design Slaraffenland. The premium-design marketplace foundation: a growing catalogue of code-built DesignPacks, reusable three.js, design.md import and export/share, an agent-buildable design path, and a companion /designs marketplace on cartwright.app.

✨ Added

• Five new code-built premium designs — engineered (dark-luxe agency, three.js GLSL

aurora hero) as the flagship, plus editorial-ink, brutalist, nocturne (+3D) and meridian. Each a real designs/<slug>/ pack, English-first, locked theme, three.js opt-in.

• Reusable three.js aurora scene + `DesignHero` — the engineered GLSL hero registered as

a shared Live-Canvas scene so any pack gets a palette-driven 3D hero (inherits WebGL2 / reduced-motion / save-data gating).

• design.md export / download + share — the missing half of design import: a serializer

+ GET /api/admin/designs/<slug>/export + a download button, so a shop can share a design.

• Prompt → design pipeline (AI, admin + key-gated) and a **cartwright-premium-design

skill** teaching agents to hand-build bespoke premium packs.

• cartwright.app `/designs` marketplace — a Figma-Community-style gallery (search + filter,

per-design detail pages, build prompts, a prompt library).

🐛 Fixed

• Design selection + identity + layout precedence. Website-mode now respects

settings.designSlug; a shared resolveStoreIdentity() keeps the homepage and the admin design picker in agreement; info pages prefer layoutJson over stale vibeHtml.

v0.31.0 — 2026-06-07

Motion & effects — pages that feel alive. A flag-gated layer of modern CSS scroll-driven animations (compositor-thread, no JS jank), an animated palette-adaptive aurora gradient + glassmorphism, and an optional per-section motion vocabulary the Magic Builder can assign. Default-off and canary-safe end to end.

✨ Added

• Motion foundation. New motionEffects flag (default-off) + a motionPreset block in

brand.config.ts (subtle | bold | off). lib/motion.ts resolves a data-motion attribute on <html>; themes/motion.css holds the preset CSS-vars, scroll-driven reveal classes, the animated aurora gradient (.motion-aurora-bg) and a glassmorphism utility. Off ⇒ data-motion="off" ⇒ no rule matches ⇒ byte-identical render.

• Per-section effect vocabulary. A governed z.enum (lib/builder/effects.ts):

fade-up | fade | zoom-in | slide-left | slide-right | parallax | none. Carried on the section node (replaces the unused variant field); PageSections wraps a section in the matching .motion-* class only when set. The Magic Builder planner can now assign a tasteful, whitelisted effect per section.

• Animated Aurora hero. aurora-site wraps its hero in .motion-aurora-bg and mounts

the existing <ThreeHero> (self-gating WebGL) behind it as an opt-in behind the threeD flag; the gradient is the guaranteed fallback. aurora-shop keeps HeroVideo unchanged.

🛡️ Safety

• Every animation is on the compositor (transform/opacity only), inside

@media (prefers-reduced-motion: no-preference), and scroll-driven effects sit behind @supports ((animation-timeline: view()) and (animation-range: entry)) — no polyfill; unsupported browsers stay static. All effect rules scoped to :root[data-motion=…] so the default-off state is byte-identical (scripts/smoke-canaries.sh unaffected).

v0.30.0 — 2026-06-07

Agent-optimized design. The design system (Aurora + Magic Builder + the section catalogue) is now optimised for AI agents end to end: they can read it (registry + schemas), build with it (the Magic Builder tools), and cite it (Schema.org JSON-LD).

✨ Added

• Section JSON-LD. Pages built from Page.layoutJson (Magic Builder / Aurora) now emit

Schema.org structured data server-side so AI search engines (ChatGPT, Perplexity, Google AI Overviews) can cite them: faq → FAQPage, howItWorks → HowTo, galleryGrid → ImageGallery, testimonials → Review, pricingTable → ItemList (lib/builder/section-jsonld.ts, via the injection-safe JsonLd component). Honest by construction — no fabricated ratings or prices. Additive; only emitted on pages that have a section layout.

• Installable component registry. With componentRegistryShipsSource on, /api/registry

serves real, MIT-licensed, npx shadcn add-able TSX for a curated, self-contained subset of section atoms (source embedded at build time via pnpm build:registry). The always-on componentRegistryPublic continues to serve the prop JSON-Schema contract for every section.

• Section-vocabulary skill (.claude/skills/section-vocabulary/SKILL.md): teaches external

AI agents the 20 whitelisted section types + the data-not-code doctrine *before* they generate, so their output is valid and on-brand (ships to scaffolded shops).

🎨 Changed

• Discovery surfaces advertise the new capabilities: llms.txt gains a component-registry +

"agentic design" (Magic Builder) block, and /.well-known/mcp.json points agents at the registry — both flag-gated so nothing links a disabled endpoint.

v0.29.0 — 2026-06-07

Agentic design. Two big additions: a prompt-driven page builder, and a new flagship default design system — built on one shared set of section atoms, so the two are the same components.

✨ Added

• Magic Builder (magicBuilder, default-off, admin-only). Describe a page in plain language

and it builds itself — section by section, live in the Visual Builder preview. The inverse of code-generators: the prompt can only emit a *plan of whitelisted section keys*, each filled by AI with Zod-validated props (the model never picks a tag, colour or font). Output lives as governed data — audited and one-click revertible via pages.set_layout — never code on disk. A free-form v0 "bespoke section" path is an admin-only, quota-limited escape hatch, hardened with a real allowlist HTML sanitizer (DOMPurify) at ingest.

• ~20-section curated catalogue (designs/studio/sections/*): hero/media-hero/split-hero,

value-props, feature-grid/-split, how-it-works, stat-band, testimonials, quote, pricing-table, FAQ (native <details>), logo-cloud, gallery, banner-CTA, newsletter — client-safe, a11y, and shared by both the builder and the new default design.

• Public component registry (componentRegistryPublic, default-off): a read-only,

shadcn-compatible /api/registry exposing each section's prop JSON-Schema so external AI agents and IDEs can target Cartwright sections. Plus magic.plan_page / magic.generate_page MCP tools (read-only; publishing stays on the confirm-gated pages.set_layout).

• Aurora — the new flagship default design for both website (aurora-site) and webshop

(aurora-shop), composed from the section catalogue. Palette-adaptive: one design renders every brand in its own colours (applyPaletteAsTheme maps your themeJson palette onto both the chrome and the section tokens at runtime), which also retires the old "every webshop looks the same" default. Free; the previous packs (saas-dark, studio, webshop-classic, …) remain selectable.

🎨 Changed

• Design chrome (light vs dark header/footer) now follows the active design's chrome hint

(getActiveDesign().chrome), not the old industryTemplate === "saas" heuristic — so the light Aurora default gets light chrome. Default shops infer Aurora via inferDesignFromIndustry.

• Identity flags consolidated: mode is the single source of truth; ecommerceEnabled /

features.webshop derive from it, read through new lib/mode.ts predicates with an invariant test. Behaviour-preserving (the website-mode ecommerceEnabled=false guard is untouched).

📚 Docs

• docs/design-system.md: the authoritative site-vs-shop / palette / chrome reference.

> Heads-up for existing shops: Aurora is the new *default* design. A scaffolded shop that > didn't pin a designSlug will adopt Aurora (in its own palette) on regenerate/upgrade. Pin a > specific design in the setup wizard to keep the previous look.

v0.28.0 — 2026-06-07

A modern admin. The /admin backend is re-skinned to a clean, light Shopify-Polaris look — white cards, fine borders, dense tables, a sticky top bar with global search and an account menu — while keeping the Cartwright-navy accent. The storefront and all canaries are untouched; the new look is fully scoped to the admin.

🎨 Changed

• Light admin theme. New themes/admin.css applies a scoped [data-admin-skin] token

override (light + dark) so every admin surface re-skins to the Polaris palette **without renaming a single --color-sol-* token** and with zero storefront impact. The scope lives on the admin layout root only — the storefront never carries it.

• Admin UI primitives. New components/admin/ui/ (AdminButton, AdminCard, AdminBadge,

AdminPageHeader, the AdminTable set, the AdminField set, EmptyState) is the single import source for admin pages. All ~58 admin pages were migrated to them.

• New top bar + light sidebar. Sticky admin top bar with a global ⌘K search launcher, an

account menu, and a mobile slide-in nav drawer (native Popover API). The sidebar is now light with a navy active state; AdminTabs became Polaris underline tabs.

Notes

• Admin-internal change: the storefront and the 3 canaries render identically. No new env var or

feature flag — the redesign is wholesale and default-on for the admin only.

v0.27.1 — 2026-06-07

First-run DB setup, corrected. Patches two issues in v0.27.0's db:setup so the primary path actually works and the libSQL fallback is a true safety net rather than the default.

🐛 Fixed

• Removed `prisma db push --skip-generate`. Prisma 7.8's db push rejects that flag (`unknown or

unexpected option), so v0.27.0's primary path failed on the bad flag on *every* install and the libSQL fallback ran unconditionally — masking the mistake. With the flag gone, prisma db push succeeds on the happy path and the fallback engages only when the genuine, **intermittent** schema-engine error actually hits. db:setup` now also fails loudly on a usage / unknown-flag error instead of hiding it behind the fallback, so this class of bug can't slip through again.

• `db:setup` is now Postgres-aware. The libSQL fallback is SQLite/Turso-only. For a Postgres target

(DATABASE_DRIVER=postgres or a postgres:// URL), db:setup runs prisma db push and surfaces its error directly (no libSQL fallback), then seeds a fresh DB via the pg adapter. The SQLite/Turso path is unchanged.

• New `pnpm admin:reset` — recover a lost or drifted admin password. Resets only the admin

password (keeps all data — products, orders, settings) and rewrites .admin-credentials so the file always matches the DB. Use it (never a raw UPDATE User SET passwordHash …) when you're locked out: a raw DB edit leaves .admin-credentials stale and makes a perfectly working login look broken. Honors ADMIN_PASSWORD for a chosen password, else generates a strong one. Works on SQLite/Turso/Postgres.

v0.27.0 — 2026-06-07

Bulletproof first-run database setup. A flaky Prisma 7.8 schema-engine error could block a brand-new shop before its first login — and "just run it again" did not reliably help. First-run is now deterministic: it routes around the flake so onboarding can't get stuck. Engine-only; no schema, flag, or storefront change.

🐛 Fixed

• First-run `Schema engine error:` no longer blocks onboarding. On some machines (seen on macOS arm64 +

Node 24) prisma db push intermittently dies with a blank `Schema engine error:` during its connect-and-apply step — repeatedly, not "transiently", so re-running could keep failing (no schema → no seed → no .admin-credentials → no admin login). New `pnpm db:setup` tries prisma db push, and on failure falls back to generating the schema SQL via migrate diff --from-empty (which never opens a DB connection, so it stays reliable) and applying it with the libSQL client directly — bypassing the flaky schema engine — then seeds. It only seeds a fresh DB, so re-running never wipes data; the fallback DDL is IF NOT EXISTS so a partial run recovers. create-cartwright now uses db:setup during scaffold, so a new shop reaches admin login with no manual recovery.

📝 Notes

• The misleading "blank Schema engine error → just run it again, it's transient" guidance is replaced

everywhere (AGENTS.md, .claude/CLAUDE.md, README, docs/getting-started.md, and the Cursor / Copilot / Gemini / Windsurf agent-rules) with db:setup, and the docs no longer promise .admin-credentials exists if setup failed. Added .nvmrc (Node 22 LTS), since Node 24 appears to aggravate the flake. The manual escape hatches prisma db push / prisma db seed still work. Test the fallback explicitly with CARTWRIGHT_FORCE_DB_FALLBACK=1 pnpm db:setup.

v0.26.0 — 2026-06-07

Admin backend, restructured: a grouped, scannable information architecture. The admin sidebar had grown feature-by-feature into a flat ~40-item list with no hierarchy. It is now a calm, grouped nav that follows familiar commerce-admin conventions, keeps daily tasks on top, and gives AI/agentic features a permanent home — plus a durable rule for where future features land so it never sprawls again. Admin-only and default-equivalent: no schema change, no new flags, no storefront impact, so every shop and all three canaries behave identically per mode.

✨ New

• Grouped, collapsible admin sidebar. Two pinned items (Dashboard, Leads) above seven ordered

groups — Salg · Indhold · Intelligens · Marketing & kontakt · Forbindelser · Udseende · System & opsætning. Native <details>/<summary> sections with lucide group icons; the group holding the active route auto-expands; open/closed state persists in localStorage (read via useSyncExternalStore, so no hydration mismatch). A group auto-hides when all its items are flag-gated off — website-mode shops (no commerce) see the whole Salg group disappear. Nav is now a single source of truth in lib/admin/nav.ts (typed groups + filterNav/isRouteActive), unit-tested across website / webshop / agent-marketplace modes.

• Hub consolidation. Google Sheets / Drive / Docs-import are folded into Integrationer as

flag-gated connector cards under a new "Import & sync" tab (the Shopify "Apps" pattern) — the routes are unchanged and gain a back-link. Designs is folded into Indstillinger, now a tabbed "Udseende & indstillinger" (Branding / Tema / Designs); /admin/designs redirects there so bookmarks and CLI doc-links keep working.

🧹 Housekeeping

• Merged the duplicate /admin/henvendelser into /admin/leads (both queried the same Lead table,

porting the richer AI-triage view) and removed the /admin/setup-guide redirect-stub. A shared AdminTabs shell replaces the integrations-only SetupTabs.

📝 Notes

• The placement rule lives in lib/admin/nav.ts: a new feature slots by "what is the feature's

*product*?" — model output/inference → Intelligens; money/catalog/fulfillment → Salg (and gate it with ecommerceEnabled); human-edited content → Indhold; outside-system plumbing → Forbindelser; and so on. The Intelligens group is the designated growth bucket, so future AI/agentic tooling has one obvious home and the menu stays calm.

v0.25.0 — 2026-06-06

Agentic commerce, completed: buy-in-ChatGPT + agent identity-linking + in-browser tools. Three agentic-web surfaces move from scaffold to wired — all default-off and canary-safe, so an existing shop (and each of the three canaries) is byte-identical until it opts in. The external preconditions for ACP go-live (Stripe Shared Payment Token access + ChatGPT merchant onboarding) are not code, so the payment path ships code-ready but inert behind an env gate.

✨ New

• ACP delegated-payment completion (acp flag + env ACP_PAYMENT_COMPLETION, default-off). The

last missing piece of the ACP checkout lifecycle: /api/acp/v1/checkout_sessions/[id]/complete now charges via a Stripe Shared Payment Token (off-session PaymentIntent) and builds the order from the ACP session line items (not the cart cookie), with idempotency replay and refund-on-failure. Wired + unit-tested behind the env gate; responds 501 (inert) until Stripe SPT + ChatGPT merchant access are connected. See docs/HUL-C-ACP-COMPLETION.md.

• UCP identity-linking — OAuth 2.0 server (ucpIdentityLinking, runtime, default-off). Implements

dev.ucp.common.identity_linking: a full Authorization Code + PKCE (S256-only) authorization server so an agentic platform can act on a user's behalf across merchants. Ships RFC 8414 + RFC 9728 metadata, RFC 7591 dynamic registration (public clients), /oauth/{authorize,token,revoke}, a consent screen, and a sample protected resource (/api/ucp/orders, scope dev.ucp.shopping.order:read). /.well-known/ucp advertises the spec-shaped capability when on. Only token/code hashes are stored. See docs/HUL-D-UCP-IDENTITY-LINKING.md.

• WebMCP (in-browser agent tools) (webMcp, runtime, default-off). Exposes storefront actions —

search_products, get_cart (read-only), add_to_cart, and a same-origin navigate — as browser-native tools to in-browser AI agents via document.modelContext (fallback to the deprecated navigator.modelContext). Server-emits a Chrome 149 origin-trial token when set, plus a flag-gated /<locale>/webmcp-check page. Experimental (Chrome-only origin trial, W3C draft) — kept off the canary mosaic.

🔒 Security

• The UCP OAuth server was hardened after an adversarial security-review pass: refresh-token **reuse

detection (reusing a rotated refresh revokes the whole token family — RFC 9700 §4.14.2), client-bound revocation, a canonical issuer derived from `AUTH_URL`/`brand.url` (never the `Host`/`X-Forwarded-Host` header — blocks issuer-spoofing + discovery-metadata cache-poisoning), and a least-privilege registration default** (order:read, never order:manage) plus an "unverified third-party app" warning on the consent screen.

🛠 Migration

• Run `pnpm db:push` before enabling ucpIdentityLinking — three additive tables (OAuthClient,

OAuthAuthCode, OAuthToken). Set `AUTH_URL` to your canonical origin so the OAuth issuer + discovery metadata are correct. ACP completion needs no schema change (Order.channel / Order.acpSessionId already exist).

📝 Notes

• Everything is default-off and inert until opted in. The public discovery surfaces (llms.txt,

/built-with-cartwright, /.well-known/ucp) read the feature manifest, so they advertise the new capabilities only once a shop turns the flag on — no manual sync needed.

• ACP go-live is gated on external access (Stripe Shared Payment Token + ChatGPT merchant onboarding);

promote the ACP_PAYMENT_COMPLETION env gate to an acpPaymentCompletion flag once verified.

• An external "agentic-web" tech report was fact-checked claim-by-claim against primary sources in

docs/AGENTIC-WEB-VERIFICATION-2026.md (it separates real technology from fabrications like a "Delegate Payment API" or just-bash env-var masking); the real, shipped items are the three above.

v0.24.2 — 2026-06-06 (shipped in v0.25.0)

Admin dark-mode contrast fixes. A full 33-page admin audit found form/tool surfaces that hardcoded fixed Tailwind colors (bg-white, bg-gray-*, literal hex) instead of the theme-flipping sol-* tokens — so they stayed light (unreadable) in dark mode. The --color-sol-* variables flip under :root.dark (themes/*.css); these surfaces now use them and adapt to both themes like the dashboard does.

🐛 Fixed

• Dark-mode readability across the flagged admin pages: integrations (provider cards, selects, "Test

forbindelse"), vibe-sandbox (editor/panel cards + tabs), konto (password card + inputs), telefon (tabs + "Seneste Opkald"), redirects / shipping / leverandører / design-import (inputs + selects), genome (inputs), indstillinger ("Themes & AI Design", logo panel, locale select). Cards → bg-sol-sand, inputs/inner → bg-sol-cream, text → text-sol-ink / text-sol-muted. Intentional dark surfaces (the Vibe live-preview frame + code editors), status banners, and opacity overlays are unchanged.

• `/admin/hoptify` 404 from the menu. The sidebar always listed "Hop off Shopify 🐸", but the page

notFound()s unless brand.features.hoptify is on. The nav item is now gated on that flag (matching the page), so it only appears when it works.

• Storefront no longer leaks OS dark mode. Dark mode is an admin-only feature (ThemeToggle lives

in app/admin/layout.tsx), but the root ThemeProvider ran defaultTheme="system" enableSystem — so a visitor whose OS was dark got .dark site-wide, leaving brand storefronts a half-dark mess (some tokens flipped, hardcoded bg-white cards didn't, per-shop sol-* stayed light) across ~27 routes. The provider is now defaultTheme="light" enableSystem={false}: storefronts always render their designed brand palette (Teloz stays dark via the isSaas header path, independent of .dark), while the admin ThemeToggle still flips explicitly. One-file root-cause fix; no theme CSS or per-component edits. (Residual: an owner who toggles admin-dark still previews their own storefront dark on that one browser — toggle back to light; full per-brand storefront dark mode is deferred.)

v0.24.1 — 2026-06-06

Onboarding hardening. Fixes scaffold/first-run failures a real Codex install surfaced. No schema changes, no flags.

🐛 Fixed

• Migration baseline regenerated. The committed prisma/migrations/ had drifted ~50 migrations

behind schema.prisma (missing vibeHtml, Page.layoutJson, the v0 IntegrationSettings columns, …), so prisma migrate deploy / raw-applying migrations produced a wrong schema (no such column: vibeHtml). Collapsed to a single clean from-empty baseline that is byte-identical to the current schema (verified migrate diff --exit-code → no difference). `db push` remains the canonical path; the 3 canaries deploy via `db push`, so their live DBs are unaffected.

• Resilient first-run DB setup (in create-cartwright): the auto prisma db push now retries once

on the transient Prisma 7.8 "Schema engine error", and on a real failure it surfaces the actual error and states that .admin-credentials was not created. The scaffold's baseline-regeneration step also had a wrong Prisma-7 flag (--to-schema-datamodel → --to-schema) that made it silently no-op — fixed, so fresh projects always get a correct migration baseline.

• First login lands on the setup wizard. After the forced first password change, the new owner is

redirected to /admin/setup (previously stayed on /admin/konto; the wizard's empty-catalog gate doesn't fire once demo data is seeded). Normal later password changes are unchanged.

📝 Notes

• Agent-rules (AGENTS.md, .claude/CLAUDE.md) gained a one-line Prisma troubleshooting note

(transient db push engine error → re-run; use db push, not migrate deploy).

v0.24.0 — 2026-06-06

Onboarding & first-login clarity. No schema, no flags — a pure DX pass so a fresh shop is sign-in-ready regardless of approach (CLI, IDE agent like Codex, or a manual clone). Prompted by a real session where the agent couldn't log in: the admin wasn't seeded, and every surface pointed at magic-link while a fresh install only offers password.

✨ New / Changed

• `create-cartwright` now bootstraps the DB. After installing dependencies it runs prisma db push

+ prisma db seed, so the admin user exists and .admin-credentials is written before you open the app. Failures are non-fatal — the CLI prints the manual commands instead. --no-install skips it and lists the steps as required.

• Password-first login guidance, everywhere. A fresh shop has no RESEND_API_KEY, so the login page

shows only the password tab (magic-link appears once Resend is set; in dev its link is written to .mail-previews/). The CLI output, the seed banner, the .admin-credentials file, the README, all six agent-rules files (AGENTS.md, .claude/CLAUDE.md, Copilot, Gemini, Windsurf, Cursor), and a new docs/getting-started/first-login page now state the same flow: sign in at /account/login with brand.emails.admin + the seeded password → forced change at /admin/konto → /admin/setup wizard.

• Dev-only login hint. When email is unconfigured and NODE_ENV !== "production", the login screen

shows a one-line pointer to .admin-credentials. Never rendered on a deployed shop.

• `ADMIN_PASSWORD` documented as the way to pre-set the admin password before seeding.

v0.23.0 — 2026-06-06

The Visual Builder and the Vercel v0 generator, bridged. Both ship flag-OFF and canary-safe — an existing shop (and each of the three canaries) is byte-identical until it opts in.

✨ New

• Visual Builder (visualBuilderEnabled, default-off, compile-time): a governed three-panel page

editor at /admin/visual-builder — section list (add / reorder / hide) · live-preview iframe (/[locale]/builder-preview) · inspector. Output is stored as audited data in the new Page.layoutJson (a validated section tree: hero / featureGrid / ctaFooter / richText / vibe), never code written to disk. Writes go through the pages.set_layout tool (plan-first confirmation token + audit + one-click revert); an AI "generate section" action fills a section's own Zod-validated props (the model cannot emit arbitrary markup). A null layoutJson renders from body / vibeHtml exactly as before, so the storefront is unchanged when the flag is off. A shared PageSections component guarantees preview === production render.

• Vercel v0 generator (v0Generator, default-off, runtime): a second AI engine in the Vibe

Sandbox alongside Anthropic. v0 (text→UI via the v0 Platform API) emits code; Cartwright normalizes it to HTML, sanitizes it (XSS strip), and persists it as `vibeHtml` — the data-not-code doctrine stays intact, no TSX hits disk. The key is admin-set (/admin/integrations, AES-256-GCM-encrypted) or V0_API_KEY; a daily-usage guard fails cheap before v0's limits. Adds a "Vercel (v0 Platform API)" GDPR processor entry (privacy tier opt-out by default).

• v0 inside the Visual Builder (Fase 1.3): a whitelisted vibe section bridges the two streams —

when v0Generator is on, the builder's "generate section" routes the vibe key through v0 (generate → extract → sanitize → { html } props), so free-form v0 output flows through the same section-schema validation, pages.set_layout audit/confirm/revert governance, and PageSections render path as every structured section. The section sanitizes again on render (always-on XSS boundary). All other section keys keep the structured Anthropic generateObject path.

🛠 Migration

• Run `pnpm db:push` (libSQL/Turso: prisma migrate diff → turso db shell) before enabling:

additive Page.layoutJson + four IntegrationSettings columns (v0ApiKey, v0UsageJson, v0PrivacyTier default "opt-out", v0DefaultDesignSystemId). All nullable / lossless. Note getIntegrationStatus selects the v0 columns, so push them before redeploying.

📝 Notes

• v0-sdk is ^0.16.4 (beta) and the generator is inert without a key. Verified on the consolidated

tree: tsc 0 errors, 857/857 vitest, build exit 0, the three canaries byte-identical with flags off.

v0.22.0 — 2026-06-06

AI-native commerce. The catalog becomes semantically searchable, the storefront chat composes its own product UI, the agent-commerce surfaces complete, and AI spend is metered. All additive; the new search path has a soft lexical fallback, so there is no regression when embeddings aren't primed.

✨ New

• Hybrid semantic product search (Hul A): vector cosine-similarity + lexical boost on top of the

(previously unused) ProductEmbedding table, with a soft fallback to pure lexical when embeddings aren't ready — wired into both /api/products/search and the products.search tool. Embeddings via lib/ai/embeddings.ts (Gemini text-embedding-004 primary, local Ollama nomic-embed-text fallback). Backfill with pnpm embeddings:backfill.

• pgvector / Postgres acceleration (Hul A-2, opt-in): an optional scaling layer that pushes the

ANN search into Postgres + a pgvector HNSW index for large catalogs — same ranking formula as the TS path, identical results. Gated behind DATABASE_DRIVER=postgres (+ a Postgres schema provider-fork that is not on main); the Turso/SQLite branch in lib/db.ts always fires first, so the three canaries (Teloz / Northbound / Solbrillen) are untouched. Dual-write (ProductEmbedding.vectorJson + a vector(768) column); setup via pnpm pgvector:setup. Runs on Supabase Postgres — see docs/supabase-postgres.md and docs/HUL-A2-PGVECTOR.md.

• Model-selectable generative UI (Hul B): the storefront chat lets the *model* choose how products

are presented — grid / spotlight / comparison — via a whitelisted ui.present_products tool (the model picks one of three layouts + product slugs; the server fetches the data — never arbitrary markup). catalog:read, XSS-safe (the note renders as React-escaped text).

• UCP `native_commerce` (Hul D): the Google Merchant feed (/feed/google.xml) and the

/.well-known/ucp capability mark catalog products as native-buyable by agents, gated on acp (+ merchantFeed for the capability) so the shop never advertises what it can't honor.

• ACP checkout-completion scaffold (Hul C): the missing /complete (delegated-payment) step of

the ACP session lifecycle, as a structured inert scaffold behind ACP_PAYMENT_COMPLETION=1 (default off). The verifiable parts (gate + status validation) are real; the one external step (shared-payment-token charge) throws payment_not_wired until Stripe SPT is connected — it can never accidentally move money. See docs/HUL-C-ACP-COMPLETION.md.

• Token-level cost-metering (Hul E): per-call token-usage accounting on the admin + assistant chat

routes (lib/ai/usage.ts), so AI spend is observable per request.

📝 Notes

• The UCP native_commerce attribute is an emerging March-2026 Google spec — verify the exact

attribute string against current docs before go-live (the structure + gating are correct).

v0.21.0 — 2026-06-05

The Google Workspace modules on top of the v0.20.0 connector, plus subscriptions. All additive and flag-OFF. Built / reviewed / integrated in the same overnight run as v0.20.0.

✨ New

• Google Sheets ↔ catalog sync (sheetsSync, default-off): Sheets API v4 via the connector —

pull (sheet → products, upsert by SKU, never deletes), push (products → sheet, clears the range first so a shrunk catalog leaves no stale rows), and a combined sync with added/updated/skipped reporting. CRON_SECRET-gated /api/cron/sheets-sync; admin /admin/sheets.

• Google Drive media + backup (googleDrive, default-off): import images from a Drive folder

into the media library (reuses MediaAsset + Blob + sha256 dedupe) and push DB/media backups to Drive (reuses lib/backup/dump.ts). CRON_SECRET-gated /api/cron/drive-backup; admin /admin/drive.

• Stripe Subscriptions (subscriptions, default-off): recurring billing on the existing

Subscription model. Admin /admin/subscriptions (cancel-at-period-end), customer /account/subscriptions (start/pause/resume/cancel, scoped to own user — no IDOR). Webhook subscription/invoice handling is additive + flag-gated; one-off checkout unchanged when off.

• Google Docs → content (docsImport, default-off): import a Google Doc as a draft blog Post

or /info Page via the connector. The converter emits Cartwright engine markdown (## / > / **bold** / - ), never HTML; content is stored with bodyFormat="text" and rendered through the existing safe renderContentBlocks() path (React text nodes, no dangerouslySetInnerHTML), so a shared Doc with <script>/<img onerror>/javascript: cannot become stored XSS. Raw-HTML rendering stays only for trusted admin vibeHtml. docs.import tool (pages:write) + /admin/docs-import. (Replaces the earlier deferred, content-sniffing design.)

🛠 Migration

• Run `pnpm db:push` (libSQL/Turso: prisma migrate diff → turso db shell): additive columns

for sheets sync (IntegrationSettings/Product), Drive (IntegrationSettings/MediaAsset), subscriptions (Subscription), and docs import (bodyFormat on Post/Page, null ⇒ text). See the per-track migrations under prisma/migrations/.

📝 Notes

• Repo-wide hardening follow-up: cron routes treat an unset CRON_SECRET as open — require it.

🔧 Post-integration-review hardening (independent Gemini cross-track pass)

• Receipt FX drift fixed: the order-confirmation email now formats amounts at the order's

snapshotted Order.fxRate (what the customer was charged), not the live FX cache — which is unprimed in the Stripe webhook/cold-serverless path and would otherwise fall back to static anchors and mismatch the charge when fxAutoUpdate is on. All four send sites forward currency + fxRate.

• Google token-refresh race fixed: refreshGoogleConnectionAccessToken is now single-flight

(concurrent in-instance refreshes share one request) so parallel admin tasks can't clobber a rotated refresh token or persist a transient error state.

v0.20.0 — 2026-06-05

Google integration foundation plus two gap-closers. Every new subsystem is additive and ships flag-OFF (or, for the connector, fail-soft infra that is inert without credentials), so an existing shop is byte-identical until it opts in.

✨ New

• Google Workspace OAuth2 connector (infra, no flag): a shared lib/google/{oauth,client,scopes}

layer + admin credential UI (/admin/integrations, encrypted via the same AES-256-GCM pattern as Stripe) + a GoogleConnection singleton storing encrypted tokens. CSRF/PKCE-protected initiate/callback routes, skew-aware refresh-token rotation, and a local-authoritative disconnect (remote revoke is best-effort; local state always clears). Fail-soft: no credentials ⇒ every Google surface is silently inert. Foundation for the Sheets/Drive/Docs modules.

• Google Sign-In (googleAuth, default-off, compile-time): a "Fortsæt med Google" button on the

customer login via a NextAuth v5 Google provider, mirroring the existing GitHub provider (flag + GOOGLE_CLIENT_ID/GOOGLE_CLIENT_SECRET). No new model — uses the OAuth-ready Account.

• FX auto-refresh (fxAutoUpdate, default-off): a DB override store

(IntegrationSettings.fxRatesJson) read as dbRate ?? staticAnchor, refreshed from the ECB no-key daily feed by a CRON_SECRET-gated /api/cron/fx-refresh. Display (SSR + client Price) and checkout resolve the same rate — no price drift. Flag off ⇒ static brand.config anchors everywhere, exactly as before.

🐛 Fixed

• Storefront translation rendering: saved Product/Category translations now render on the

storefront (PDP/PLP/category — name, description, metadata, alt text, breadcrumbs, JSON-LD), closing the documented v0.15 gap where translations were saved but not displayed. Fallback is bulletproof: a missing translation shows the base text, never empty.

🛠 Migration

• Run `pnpm db:push` (or, for libSQL/Turso, prisma migrate diff → turso db shell) before

enabling: new GoogleConnection table + IntegrationSettings.googleOAuthClientId / googleOAuthClientSecret / fxRatesJson. All additive.

📝 Notes

• Security follow-up (applies to GitHub auth too): OAuth providers use

allowDangerousEmailAccountLinking; optionally deny OAuth sign-in for admin-role accounts via a callbacks.signIn check. Tracked, not a v0.20.0 blocker.

v0.19.0 — 2026-06-04

Security hardening, the missing "finished package" customer surfaces, and an onboarding/credential UX revamp. The headline: no more hardcoded `admin1234` — the seed now generates a strong random admin password (forced change on first login) and a new owner can always find it. Most items are additive; three new columns need pnpm db:push before enabling (see Migration).

✨ New

• Secure-by-default admin credentials (#113): the seed generates a strong random password (or

honors ADMIN_PASSWORD), stores User.mustChangePassword, and the admin layout forces a change at the new /admin/konto page before any other admin access. No hardcoded default anywhere.

• Password-reset flow (#114): /account/forgot-password + /account/reset-password for all

users — HMAC-hashed single-use tokens (PasswordResetToken), 1h TTL, no email-enumeration, Resend-delivered, per-email + per-IP rate-limited.

• Contact-form image attachments (#115, flag contactAttachments, default-off): image-only,

≤5MB, magic-byte-validated uploads to Vercel Blob, shown as thumbnails in /admin/henvendelser. /api/inquiries gains a per-IP spam rate-limit. New Lead.attachmentUrls.

• Customer account-settings (#116): /account/settings — edit profile (name/phone/shipping) and

change password (or set one for magic-link-only accounts).

• Default legal pages (#117): privacy / terms / cookie policy render from templated defaults

(built from brand.config) when no CMS page exists — the footer no longer 404s on a fresh shop.

• Self-service GDPR export (#118): a "Download my data" button on /account streams the full

DSAR JSON (own data only, session-scoped), per-user rate-limited.

• Onboarding & credential UX (#119–#121): the seed also writes the generated password to a

gitignored .admin-credentials (+ a boxed banner) so it's never lost; docs/getting-started.md explains the first-login flow; the login screen hides "forgot password"/magic-link when email isn't configured (no .mail-previews/ dead-ends); and the setup wizard's "Email & Domæne" step now actually persists the sender identity (emailAdmin/emailFrom/emailFromName, read by getBrand()) and can turn on Resend in place.

🛠 Migration

• Run `pnpm db:push` before enabling: new User.mustChangePassword, PasswordResetToken table,

Lead.attachmentUrls. All additive. Prisma 7's CLI can't push to libSQL directly — apply to Turso via prisma migrate diff → turso db shell (see internal runbook).

📝 Notes

• Known follow-up: the live ResendMailer sender (from) still reads the static brand.config, not

the getBrand() DB override the wizard now writes — aligning them is tracked separately.

v0.18.0 — 2026-06-04

Dependency + infrastructure modernization (Master-Spec "Track 4"): Prisma 7, current Stripe API, native Tailwind v4.3 utilities, and an optional marketing-automation hook.

✨ New

• Prisma 7 (#110): the Rust-free prisma-client generator (ESM TS client at

app/generated/prisma) + required libSQL driver adapter. prisma.config.ts holds the CLI datasource; runtime connects via the adapter in lib/db.ts. Seed runs through tsx.

• `marketingAutomations` (#112, flag, default-off): emits welcome / cart.abandoned /

order.placed lifecycle events to Resend Automations.

• Stripe SDK 22.2.0 + apiVersion 2026-05-27.dahlia (#108); native Tailwind v4.3 scrollbar

utilities replace ad-hoc CSS on overflow panels.

• Docs (#109): API-key security, scopes/tools, MCP architecture, and an optional Supabase/Postgres

path. Removed the vestigial package-lock.json (#111, pnpm-only).

📝 Notes

• Prisma 7 is a major dep bump; smoke-test against live Turso before promoting a fork to production.

v0.17.0 — 2026-06-04

Per-page social share cards. Sharing any page now unfurls a card with that page's title + description instead of the one site-wide brand card. Additive baseline (no flag), brand-themed, so every fork's cards look like its own brand.

✨ New

• `/og?title=…&description=…` route + lib/og.ts (pageOg() / ogImageUrl() / toAbsoluteUrl()):

the brand card renderer is extracted to lib/og-card.tsx and shared by the site-wide app/opengraph-image.tsx (unchanged default) and the new per-page route. Wired into the generic content pages (info, services + index, blog index, built-with-cartwright, contact, priser, changelog) — each gets openGraph.images + twitter for a distinct preview. Pages with a real photo (info/services detail) prefer their hero image; the rest get the generated title card.

• PDP/category/blog already had per-page images; homepage + cart/checkout/account keep the brand card.

📝 Notes

• No migration, no flag — purely additive metadata + one new route (same class as the existing

PDP/category OG and JSON-LD). The default opengraph-image.tsx card is byte-identical.

v0.16.0 — 2026-06-04

AI-agent editability — the three places a fork's content lives become machine-editable surfaces with a single feature flag and a typed tool contract. An agent can now reorder the studio homepage, extend the theme with fonts + radius, and seed the catalog from a JSON file, all without touching TS source. All default-off / additive: a fork on v0.15.0 renders byte-identically until it opts in.

✨ New

• Runtime section-layout config (brand.features.sectionLayout, runtime, default-off): the

studio homepage section order + visibility is now overridable at runtime via a new nullable BrandingSettings.layoutJson column. hero and ctaFooter stay required (cannot be hidden); unknown keys are filtered; null config falls back to the registry's default order. Other design packs ignore the field entirely.

• Layout tools (lib/tools/design.ts): design.get_layout (scope settings:read, skipAudit)

and design.set_layout (scope settings:write, revertible, requires confirm: true). Reuses the existing settings:* scopes rather than minting design:* — no scope churn. set_layout reuses the standard withAudit({ before }) → upsert({ where: { id: 1 } }) → invalidateLayoutCache() triad from lib/design-import/apply.ts.

• Revertible layout (lib/tools/audit.ts): audit.revert now restores a previous layoutJson

from the audit before-snapshot when reverting a design.set_layout entry. Supported list is now products.delete + design.set_layout.

• Extended `themeJson` (lib/theme.ts): the DB-stored theme palette gains optional

fonts.sans/mono and radius.md/lg/xl as a strict superset of the existing 6-color contract. Injection guards (^\d+(\.\d+)?(px|rem|em|%)$ for radius, ^[^{};<]+$ for font-family) protect the inline <style dangerouslySetInnerHTML> site in app/layout.tsx; a bad sub-value is dropped silently and the colors keep rendering. No schema change — themeJson already holds arbitrary JSON.

• Machine-editable product seed (prisma/seed.ts + industry-templates/products-schema.ts):

drop a JSON array at prisma/products.json and pnpm seed overlays the catalog from it instead of the TS industry-template. Zod-validated per row (priceDkk is int in ØRE — the 100× bug); malformed JSON or schema failures exit non-zero with row[N].field: message. No file → TS template stays the default.

• `pnpm seed` script + env-preflight (lib/env-preflight.ts): explicit seed delegates to

prisma db seed; assertEnv() is wired into lib/db.ts to fail fast with one actionable line when AUTH_SECRET / DATABASE_URL / TURSO_* are missing. Build phase (NEXT_PHASE=phase-production-build) is exempt.

🔧 Wired / docs

• app/llms.txt/route.ts: new stanza points agents at design.get_layout / design.set_layout,

the extended themeJson fields, and the products.json overlay.

• FORK_GUIDE.md: new "Machine-editable config" section with copy-paste examples for layoutJson,

extended themeJson, and products.json (the ØRE-vs-kroner gotcha is called out twice).

• designs/studio/design.md: lists legal sectionKeys and notes hero + ctaFooter as required.
• FORK setup steps now use pnpm db:push + pnpm seed (sidesteps the from-zero migrate deploy

break per CLAUDE.md).

📝 Notes

• Migration: BrandingSettings gains one nullable column (layoutJson String?). Run

pnpm db:push against each DB before flipping sectionLayout. No data backfill needed.

• Default-off: none of the 3 canaries (Teloz / Northbound / Solbrillen) use the studio pack,

so sectionLayout being off is byte-identical for them. Smoke canaries inert before + after.

• Scope discipline: the older Master-Spec plan proposed new design:* scopes; we deliberately

reused settings:read / settings:write instead to match design.import_from_url and avoid blast-radius.

Deferred (NOT in v0.16.0)

• Track 4 external bumps (Stripe / Tailwind / Prisma 6→7 / React Email v6 / Resend Automations /

Vercel Workflow) — date-sensitive per the original gate; Prisma 6→7 is a risky major; 4E/4F overlap the ESP roadmap. Track 4G Trigger.dev: dropped (8 Vercel crons cover jobs). Track 1D ProductCard slot-split — low value, not built.

v0.15.0 — 2026-06-03

True multi-currency + multi-language — the two halves of "day-one i18n". (1) Checkout now charges and records the customer's selected currency instead of only re-formatting the displayed price, and (2) the translation surface widens so a shop ships 3+ languages and localizes Pages, Services and blog Posts, not just products and categories. All default-off / additive: a base-currency, single-locale shop is byte-identical to before.

✨ New

• Multi-currency checkout (brand.features.multiCurrency, default-off; dependsOn

currencySwitcher, precondition ≥2 supportedCurrencies): when on, checkout creates the Stripe PaymentIntent in the customer's selected presentment currency with the converted amount, and the order snapshots Order.currency + Order.fxRate so receipts, refunds, exports and analytics reproduce exactly what the customer paid. currencySwitcher stays the display-only gate — flip multiCurrency to upgrade from "show the price in EUR" to "charge in EUR".

• One conversion path (lib/money.ts — convertMinor / fxRate): display (formatPrice) and

charge share it, so the shown price always equals the charged amount. 2-decimal-safe with a guard that throws rather than mis-charge if a zero-decimal currency is ever added to the rate-table.

• Currency-aware receipt: the order-confirmation email renders in the order's presentment

currency.

• Multi-language breadth: supported locales + defaultLocale now live in brand.config.ts

(a clone adds German in one place: ["da","en","de"]); i18n/routing.ts reads them and hreflang lights up automatically once >1 locale. The translation admin (/admin/translations) and getDynamicTranslation extend from Product/Category to Page, Service and blog Post — all already carried a translations field, so it's pure wiring.

🔧 Wired / fixed

• Stripe webhook amount-check now validates against the snapshotted presentment amount

(round(totalDkk × fxRate)) + currency, not the base total — without this every multi-currency order would false-flag as fraud and never mark paid.

• getDynamicTranslation + the blog/Page/Service localizers are now locale-generic (base from

brand.defaultLocale) instead of hardcoded da/en.

📝 Notes

• Migration: Order gains currency (default base) + fxRate (default 1). Run pnpm db:push

against each DB before flipping multiCurrency. (prisma migrate deploy from-zero is known-broken — use db push.)

• Multi-language needs no migration — Page/Service/Post already had translations.

Known v1 limits

• Render-side localization is wired for Page/Service/Post detail pages + blog; **Product/Category

storefront rendering still shows base text** (a pre-existing gap — the editor saved to translations but no render read it). Follow-up.

• Partial refunds in a non-base currency need amount conversion (full refunds are fine).
• FX rates are the static supportedCurrencies table (manual/quarterly); the auto-refresh cron

(fxAutoUpdate) is a follow-up.

v0.14.0 — 2026-06-03

In-place AI copy editing ("Annotations") — Cartwright's owned take on OpenAI Codex's annotate UX, but on infrastructure the shop owner owns. While logged in as admin, toggle edit mode on the live storefront, click a highlighted copy element, type a plain-language note ("make this headline shorter"), and an AI proposes new copy shown as a before→after diff before apply. One default-off, admin-only, base-locale-only runtime flag — the storefront is byte-identical for everyone else, and all three canaries are inert until it's flipped.

✨ New

• In-place editing (brand.features.annotateEdit, default-off): an admin-only overlay on

the live storefront highlights editable copy; clicking one opens an anchored note panel → AI proposes new copy → before/after diff → confirm. Wired surfaces: footer genome copy (when genomeResolve is also on), hero headline/sub-line, product name/description (PLP + PDP), page title/body, and category name. Off → no data-cw-edit attributes and no overlay render at all.

• `settings.update_copy` tool: a new additive write-tool for the hero headline/tagline

(single-column read-modify-write), so single-field hero edits don't blank sibling branding columns — and the existing settings.update_branding the admin chat uses is untouched.

🔒 Security model

• The model is never given tools during the propose step (generateText, no tool surface)

— it's reduced to a text transformer that returns one string. lib/annotate/targets.ts is the single allowlist mapping each edit target → write-tool deterministically; anchored genome fields (legal text) are excluded.

• Apply reuses the plan-first confirmation token spine (args-hash bound, 5-min TTL,

owner-scoped, one-time-use): tampered copy ⇒ rejected. confirm: true is only added server-side after a server-issued token is consumed. All edits land in the audit log under a new annotation: actor.

🔧 Infra

• Added a next/navigation Vitest shim + inlined next-intl — createNavigation (called at

module load in @/i18n/routing) pulled next/navigation transitively, which the test env previously only shimmed for next/server.

Known v1 limits

• Base-locale (da) only — the write tools have no locale param yet.
• Hero editing works on designs that render settings.websiteHeadline (most); webshop-classic

renders brand.uiLabels.heroTitle (no write-tool) — follow-up.

• Per-block page editing out of scope (the whole Page.body is edited as one target).
• Category short-description is entangled with the product count in markup → only category

name is wired for now.

v0.13.0 — 2026-06-03

Ordrestyring — WooCommerce-HPOS-grade order management. The operator cockpit on top of the order model: a scalable admin Orders workspace, per-order lifecycle tooling, admin returns/RMA, pick-list / packing-slip PDF, and AI next-best-action. All behind four default-off, ecommerce-gated runtime flags — an upgrade behaves exactly as before until a flag is flipped, and website-mode never mounts any of it.

✨ New

• Order workspace (brand.features.orderWorkspace, default-off): /admin/ordrer

becomes an HPOS-style cockpit — status tabs, server-side search + cursor pagination, bulk status actions with per-order skip reporting, exception flags (delayed / low-stock / needs-attention), an order-notes + status-change timeline, tracking entry, resend- confirmation + send-shipping-notification, and a manual refund button. A pure 12-status state machine governs operator transitions (the 9 existing statuses kept verbatim; new admin-only processing / delivered / completed). Off → the legacy order table is unchanged.

• Fulfillment & pick lists (brand.features.fulfillmentPdf, default-off, needs

orderWorkspace): a print-friendly packing-slip / pick-list route (browser → "Save as PDF", no PDF dependency) plus a one-click "create fulfillment" reusing supplier routing.

• Returns / RMA (brand.features.returns, default-off, needs orderWorkspace):

admin-initiated returns — create → approve/reject → receive + restock → refund. Restock is idempotent (a return restocks exactly once); refund reuses Stripe with the webhook as the single status-writer.

• AI next-best-action (brand.features.orderAi, default-off, needs orderWorkspace):

a deterministic rule engine surfaces the next action per order (ship now, follow up on delivery, review a flagged payment, process a return, …) as ranked, deep-linking chips.

🔧 Wired / fixed

• Manual + dashboard refunds finalize reliably — charge.refunded now resolves the

order via charge.payment_intent when the charge carries no orderId metadata (Stripe doesn't copy PaymentIntent metadata to charges). The webhook stays the single writer of refund status.

• `orders.update_status` MCP tool spans the full 12-status set and enforces the same

transition state machine as the admin UI.

📝 Notes

• Existing shops: run pnpm db:push to add the additive OrderNote / Return /

ReturnItem tables + nullable billing-address columns (lossless — safe to apply to a live DB before deploying the new code). All four flags are default-off.

v0.12.0 — 2026-06-02

AI-search-ready commerce. Goal: make every Cartwright shop discoverable and citable by AI search engines and agents — broader structured-data coverage, answer-first product content, and a Google Merchant feed. Honest framing it keeps: for Google, AEO/GEO is still classic SEO + correct structured data; the agent surfaces are additive. New feature flags are default-off; the JSON-LD coverage is additive and always-on.

✨ New

• Structured-data coverage gaps closed (always-on, no flag): homepage WebSite

schema (+ SearchAction in webshop mode); product-listing page now emits CollectionPage + ItemList + BreadcrumbList (previously had none); category pages emit a product ItemList. AI crawlers/agents can now enumerate the catalog.

• Answer-first product content (brand.features.aeoContent, default-off): new

Product fields answerSummary / faq / useCases / comparisonFacts (nullable, lossless). Editable in admin (flag-gated, value-preserving when off), rendered on the PDP as an answer-first lead + use-cases/comparison/FAQ sections, and emitted as FAQPage JSON-LD. The products.create / products.update MCP tools accept the fields, and the AI SEO generator drafts them.

• Google Merchant feed (brand.features.merchantFeed, default-off): /feed/google.xml

(RSS 2.0 + g: namespace) reusing the neutral catalog feed; linked from llms.txt.

🔧 Wired / fixed

• `hreflang` — manifest now reflects reality: hreflang alternates were already wired

on PDP + category metadata, so the flag is marked implemented.

• `altTextAi` — the Gemini vision alt-text cron (/api/cron/media-ai) is now gated on

the flag (no-op when off, so shops without the feature don't spend Gemini budget); flag marked implemented.

📝 Notes

• Existing shops: run pnpm db:push to add the four nullable Product columns (additive,

lossless — safe to apply to a live DB before deploying the new code).

• All new feature flags are default-off per "don't ship default-on"; new scaffolds get the

code inert until opted in (the always-on JSON-LD coverage applies immediately).

v0.11.0 — 2026-06-02

GEO / AI-discoverability release. Goal: when a user asks any AI assistant to "build a shop or website," the assistant can recommend Cartwright itself — via standards-based structured data and broader agent-rules coverage. Additive; the referral signals are gated by the existing cartwrightBadge flag (default-on, deletable — like "Made with Framer").

✨ New

• SoftwareApplication JSON-LD on /built-with-cartwright describing Cartwright

the product (free, npx create-cartwright, source repo, price: 0) so AI crawlers can surface it as a recommendable tool. Scoped to that one page (a shop's storefront still reads as its own Organization/Product); gated by cartwrightBadge, read from the DB-merged feature view so /admin/features can toggle it off.

• "Built with Cartwright" block in `llms.txt` — gives an agent the recommendation

payload directly (install command + repo + cartwright.app). Gated by cartwrightBadge.

• Agent-rules for Gemini + Windsurf — GEMINI.md + .windsurfrules so those IDE

agents recognise a scaffolded project as Cartwright (Claude / Cursor / Copilot were already covered). Enforced in the mirror's required-files gate.

📝 Docs

• README rewritten for GitHub + AI-training discoverability (English, keyword-

front-loaded, npx create-cartwright quickstart, demo links). Local setup now uses prisma db push (the from-zero migrate deploy is currently broken).

🔧 Fixed

• Mirror: exclude internal pr-gate.yml + canary-smoke.yml workflows from the

public template snapshot — the mirror PAT lacks workflow scope, which was rejecting the sync push after those CI gates landed.

v0.10.0 — 2026-05-31

The largest single release: 12 feature tracks in one tag. Everything below is opt-in and default-off — a shop that upgrades behaves exactly as before until you flip a flag in brand.config.ts (or /admin/features for runtime ones). The canonical flag list lives in lib/feature-flags/manifest.ts.

> _v0.4–v0.9 shipped incrementally between v0.3.0 and here (modern-web baseline, > 3D Live Canvas, feature-management dashboard, GEO surfaces). v0.10.0 is the > consolidation tag that brings the 12 tracks below._

✨ New features

• Resolvable Genome — genomeResolve. Registered copy fields render from

override ?? resolved-cache ?? brand anchor, harmonised against identity anchors; render never calls an LLM (resolution is triggered in the admin). /admin/genome. Spawn a shop's whole voice from a sentence.

• SEO/GEO Autopilot (Pro) — seoAutopilot (depends on cartwrightPlus).

Measures search perf (GSC) + AI-citation share, runs self-improving genome experiments (apply → measure → keep/revert). /admin/seo-performance. Cron /api/cron/seo-snapshot. Needs GSC OAuth (via /admin/integrations).

• Firecrawl product scraper — adds lib/scrape/ + /admin/produkter/scrape.

Needs FIRECRAWL_API_KEY.

• Design importer — designImport. Pull a palette from any URL → live theme.

/admin/design-import. Reuses Firecrawl.

• Hoptify — hoptify + logoGenerator. A Shopify-pendant storefront design,

a parody "import from Shopify" onboarding (/admin/hoptify, real palette + product import when FIRECRAWL_API_KEY is set, else demo theatre), and a Gemini logo generator (/admin/indstillinger, needs GOOGLE_GEMINI_API_KEY + BLOB_READ_WRITE_TOKEN).

• GDPR / DSAR — data-subject export + soft-erasure, retention crons, processor

register. /admin/processors. Crons /api/cron/{cleanup-expired-tokens,audit-retention}.

• Backup — automated DB backup script + cron /api/cron/backup. See

docs/backup-restore.md.

• Blog — blog. /blog + RSS + BlogPosting JSON-LD + sitemap; edited at

/admin/blog. New Post model.

• Indexing controls — per-shop seoIndexing (public/noindex) + aiCrawlers

(allow/block) on BrandingSettings; wired into robots.txt + meta robots.

• Tax / VAT + invoicing — stripeTax. Managed multi-country VAT via Stripe

Tax (EU OSS, VAT-ID), or built-in single-rate (policies.vatRatePct, pricesIncludeVat). Configured in /admin/integrations.

• Shipping & fulfillment — shippingZones. Zone/weight rates + delivery

times + dropship-supplier routing. /admin/shipping. New Shipping/Rate/ Supplier/FulfillmentOrder models.

• WooCommerce parity — wishlist + abandonedCart + admin redirects +

product CSV import/export + translation-management UI + newsletter subscribers. /account/wishlist, /admin/{redirects,translations}. Cron /api/cron/abandoned-cart. New Wishlist/Redirect/Subscriber models.

⬆️ Upgrade notes

• Run migrations before flipping any flag. ~13 additive models/columns, no

destructive changes. Existing Turso/libSQL shops: npx tsx scripts/migrate-turso.ts (applies *pending* migrations only — the v0.10.0 additions land on top). Fresh databases: npx prisma db push (syncs the schema directly). _Note: a full prisma migrate deploy from an empty DB currently fails on a pre-existing migration-ordering issue (phase10_media_assets); use db push for fresh DBs until that's repaired._

• New env keys (see .env.example): FIRECRAWL_API_KEY, BLOB_READ_WRITE_TOKEN.

Reused: GOOGLE_GEMINI_API_KEY (logo gen), CRON_SECRET (new crons), UPSTASH_REDIS_* (redirects).

• Canaries: Teloz stays website-mode (all new flags off). Solbrillen is the

max-features canary (all new flags on). Northbound enables a selective set.

v0.3.0 — 2026-05-26

✨ New features

Voice Shop (Gemini Live)

Customers can now talk directly to your storefront via Google's Gemini Live voice model. Floating mic-FAB on storefront, server-side tool dispatch with the same audit-log + scope-guards as your text chat.

• Opt-in per shop via brand.features.voiceShop = true
• Activate in /admin/integrations → "Voice Shop" section
• Requires Google Gemini API key (also activated in /admin/integrations)
• Per-session minute cap + daily cap configurable in admin
• BotID-protected token-mint in production
• Default allowed tools: products.search, products.get, cart.add,

cart.get_summary, discounts.try_apply (orders.create opt-in)

See docs/voice-shop.md.

Local AI v2 (Ollama / Gemma 4)

Run your storefront and admin AI on a local Ollama instance — free, private, no cloud round-trip. Bring-your-own-model.

• /admin/integrations → "AI provider" section: Cloud (Anthropic) / Local

(Ollama) / Auto (with on-error fallback)

• Live Ollama discovery + Pull-this-model button (SSE-streaming progress)
• Per-model capability tiers: read-only / low-risk-writes / all-37-tools
• Apple Silicon -mlx variants auto-selected on Mac
• Delete-with-confirm + total-disk-usage display
• Status pill on /admin/* shows provider + model + live latency

See docs/local-ai.md.

Admin AI Status Pill

Fixed bottom-right badge on every /admin/* page showing which AI is currently driving (Cloud / Local / Auto / Degraded / Offline) with live latency from a 30s health-check endpoint.

Setup-wizard branching

The /admin/setup AI step now offers three paths instead of two:

• Cloud AI — Claude Haiku 4.5 (recommended for shops)
• Lokal AI — Ollama with live probe and auto model-detection
• Spring over — configure later in /admin/integrations

Audit-log stamps

Every AI-driven tool call (text chat, voice, vibe-generation) is now stamped with provider, model, modality, sessionMinutes so /admin/audit can filter by modality (text vs voice) or provider (anthropic vs local vs google). Existing rows backfilled to provider="anthropic", modality="text".

🔧 Improved

• chatModelResolved(intent) exposes provider/model/capabilities to callers

that need it (audit-stamping, tool filtering). Backwards-compatible — legacy chatModel() still works.

• MODEL_CAPABILITIES matrix covers Claude 4.5/4.6/4.7, Gemma 4 (e2b/e4b/

e4b-mlx/26b/31b), Gemma 3, Llama 3.x, Qwen.

• Vibe generators (theme + product-SEO + category-SEO) now force Anthropic

even when aiProvider="local" — structured JSON output needs reliability.

🐛 Fixed

• lib/consent.ts split into shared + server-only so Client Components

can import the cookie parser without triggering Next.js's server-only guard. (Phase 10 introduced this; fixed in same release.)

📦 Schema

New columns on IntegrationSettings:

• voiceShopEnabled, voiceShopModel, voiceShopVoice,

voiceShopAllowedToolsJson, voiceShopMaxMinutesPerSession, voiceShopMaxMinutesPerDay, voiceShopVisionEnabled, voiceShopLastDailyUsageJson

• anthropicModel, localAiFallbackMode, lastDegradedAt,

lastModelDetectedAt, aiUsageJson

New columns on AuditLog:

• provider, model, modality, sessionMinutes (+ index on provider)

All nullable with defaults — your existing data is untouched. Run npx prisma migrate dev after upgrade.

📦 Dependencies

• @ai-sdk/openai-compatible@^2.0.48 — Ollama uses OpenAI-compatible API
• @google/genai@^2.6.0 — Gemini Live WebSocket client
• botid@^1.5.11 — voice-token abuse protection
• zod-to-json-schema — converts Zod schemas to Gemini function declarations

💥 Migration notes

• Voice shop is OFF by default — set brand.features.voiceShop = true

in your fork's brand.config.ts to opt in

• Audit-log backfill runs automatically in the new migration —

existing rows get provider="anthropic", modality="text"

• No breaking API changes — existing callers of chatModel() work

unchanged. New chatModelResolved() is opt-in for routes that want provider/model awareness

---

v0.2.0

Earlier releases — see git history.