Blog

4 min readfeature

The commerce round: a storefront, a checkout, and a browser QA pass

Three SDLC rounds — planned, reviewed, browser-QA'd — and the landing page now frames two agent-built apps in a live gallery: 838 calls, one SHIP verdict.

  • #landing
  • #demo

The live-preview round asked whether the artifact runs. This round asks the harder one: can the agent carry a whole SDLC — plan, build, review, QA — on an app complex enough that each stage has real work to do? And can the landing page carry more than one proof at a time?

What got built

Sigil Bazaar — a fullstack e-commerce demo, an arcane marketplace of runes, sigils, relics and tomes:

  • a storefront with 16 products, search, category chips, four sorts, and CSS-drawn glyph monograms (no binary image assets anywhere);
  • a cart drawer with steppers, tax at 8%, and free shipping over $50;
  • a three-step checkout wizard with real inline validation (Luhn on the card, expiry-not-in-the-past, 5-digit ZIP);
  • an order history seeded with eight orders so the analytics have data;
  • an owner dashboard: KPIs, a 14-day revenue chart, top-5 products, category mix, and an inventory table with restock actions;
  • API routes with an in-memory store seeded identically to the client; localStorage-first persistence; 88 unit tests (the brief asked for 40).

How it was driven

Like a real user, across the three rounds — planned, reviewed, then browser-QA'd by the agent itself:

  1. Plan — plan mode, the 13-point plan, the approval overlay (a human in the loop before a line of code), then the build: scaffold, domain layer, components, API routes, tests, curl verification, commits.
  2. Review — my real-user pass found the API snapshotting order items as name=productId, price=0 and re-posting every seeded order on each fresh load; a VLM design critique through the bridge found the rest. The agent verified all 56 findings claim-by-claim, fixed the real ones, and rejected the misreadings with reasons.
  3. QA — the agent drove a real browser (a browser subagent) through every view, a failing-validation checkout, and the mobile layout. It found one real bug — the confirmation step closed too fast to read — fixed it, re-ran the pass, and shipped the verdict below.
The unown TUI in plan mode: the 13-point plan for Sigil Bazaar rendered above the approval overlay, with Execute the plan pre-selected
The plan-approval overlay — the moment the round stops being a prompt and becomes work. The pane is a real PTY capture; the plan is the agent's own.
The QA round's closing summary in the unown TUI, ending in the SHIP verdict with the honest follow-ups listed
The QA round's closing summary — the verdict is SHIP, and the follow-ups it lists are the ones it could not fix (server-side id-dedup, chart tooltips, mobile at desktop width only). An honest list, not a marketing one.

The receipts: 838 tool calls across the round (the ledger also counts 9 auto-checkpoints), 7 agent commits, 88 tests, zero human edits to the app's code. The full record — briefs, findings, patches — is docs/bazaar-demo-session.md.

The gallery

The landing page's live-demo band is now a gallery: two agent-built apps, each iframed from its own route in browser chrome, one switcher — the pattern v0's template gallery made the category default. Pick Sigil Focus and drag a card; pick Sigil Bazaar and buy something.

Sigil Bazaar — the storefront the agent built: an arcane marketplace with category chips, product cards with glyph monograms, and the cart in the header
The poster is a real capture of the route the band frames — the app boots seeded, so this is what a first-time visitor's storefront looks like.

Switching apps is one same-origin navigation on a persistent frame; the receipt chips swap with the tab (each app's numbers come from its own build record, pinned by test — no hand-typed counts).

The honest engineering of embedding

The integration pass found real defects and fixed them at the right layer:

  • A future-dated seed: the seed generates same-day orders at fixed times-of-day. At 00:04 UTC the integration test caught the seed dating orders into the future — future orders render ahead of "now" in the Orders view and the analytics count revenue that has not happened. A same-day time still ahead of now is now dated yesterday at the same time.
  • AA contrast through alpha: the first cut of the gallery's selected tab used a translucent background over the chrome bar. Axe computes the effective contrast through an alpha background — 1.73:1. The tab is an opaque pill now, and the a11y gate is the reason.
  • Both demos got the hardening round: CSP frame-ancestors 'self' + X-Frame-Options: SAMEORIGIN on the demo routes, the reset handshake pinned to window.origin (never "*"), a per-IP rate limit on the public order POST, a seedVersion tag so new seeds reach returning visitors, and the sandbox attribute trimmed of allow-modals.

Try to break it

The band keeps the same contract as the first demo: sandboxed frame, lazy-mounted (an idle visit never pays the demo's hydration cost — the poster carries the first impression), a working reset, and a full-screen escape hatch per app. Below sm it is the poster and a tap-through, not a 390px iframe.

If the storefront sells you nothing else, the checkout wizard is worth one pass: type an invalid card, watch the inline validation, then put 4242 4242 4242 4242 through and read your order number. It was built, reviewed, browser-tested, and shipped by the agent — the receipts are one click away.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts