A quick tour under the hood.

The lookup path. A query DSL parses a wishlist line (top 5 Mew cards <= $50) into a structured request; the pricing layer hits several open sources in order until one resolves, with a stale-while-revalidate cache in front. Results stream to the SPA over Server-Sent Events, so rows fill in as they land instead of blocking on the slowest one.

The seams. Sources, the cache, and the mailer are all small interfaces, which keeps tests honest — most of them patch one seam rather than mocking the world. Architecture decisions live as ADRs in docs/adr, so the "why" is written down, not folklore.

How it reviews itself. Changes go through an "AI pit crew": one agent implements, a different agent reviews, and a human owns direction and the merge. It's documented in the repo's contributor guide — and it's why the commit history reads the way it does.

The guardrails are strict on purpose: Conventional Commits, a DCO sign-off, a maintainability gate, and a design-token adherence check all run in CI. They make a first contribution predictable — green CI means you matched the house style.

Curious about a particular corner? Reply and ask — I like these questions.

— Matt