zeekay
89691f3697
The build writes the catalog into dist after vite: a page per listing (the shell with its own head, JSON-LD and a static body the app replaces), its twin at <path>.md, the shelves, the front, sitemap.xml, llms.txt, llms-full.txt, 404.html, and a moved page for any path the last release listed and this one does not. Sources are read whole; one that fails or answers empty stops the build. Every external link is fetched and dropped unless it answers. A price and an Offer come only from the shop. Directory and template entries say they are not shop listings. Fabricated ratings, reviews and installs, the directory's certified tax form, and the persona and template CLI lines the CLI lacks are gone. WebMCP registers search_market and get_listing, read-only. An unknown path renders Not found instead of the front page. Persona definitions move to /personas/<slug>/PERSONA.md, where the twin's path was.
3.8 KiB
hanzoai/market — hanzo.market
Vite SPA, published by hanzoai/ci's site lane as slug hanzo-market (universe infra/sites/sites.yaml, spa: true). Every read and write is api.hanzo.ai /v1.
Build
bun run build = vite build, then bun scripts/pages.ts writes the catalog into dist/ (about 36 s, most of it the link check):
- A page per listing at
<path>/index.html(the built shell with that page's title, description, canonical,<link rel="alternate" type="text/markdown">, JSON-LD and a static body inside#root;createRootreplaces the body when the app mounts), and its twin<path>.md(/->/index.md). - Shelves (
/agents.../templates), the front (/), shells for the private routes (/login,/jobs,/sell/*: noindex, empty root, a twin pointing at/auth.md). sitemap.xml(an index past 50,000 URLs or 50 MB),llms.txt,llms-full.txt,404.html(noindex app shell),retired.json.- Pure logic is
src/lib/pages.ts(under the 80% coverage gate); the script only reads, checks and writes.
Sources, read whole; one that fails or answers empty stops the build so a publish never drops pages:
GET /v1/marketplace/shop (seller listings: the only source with a price and an Offer), GET /v1/catalog (apps), api.hanzo.ai/.well-known/agent-skills/index.json (skills), src/lib/personas.ts + public/personas/<slug>/PERSONA.md (67 definitions), src/lib/registry.ts (directory agents, tools, MCP servers not in the shop), src/lib/templates.ts.
Not prerendered: the MCP registry mirror (GET /v1/tool/catalog, 51,266 servers, synced daily). One HTML per server is ~51k files re-uploaded on every JS hash change, past the 5-minute CI ceiling, and duplicates registry.modelcontextprotocol.io. Its pages stay client-rendered under spaMode; llms.txt, /mcp and WebMCP point at the API.
Truth rules: price and JSON-LD offers only for shop listings; a directory or template entry says Shop: Not a shop listing; reputation only from the platform (fabricated numbers removed); no CLI line the CLI lacks (persona and template commands removed). Every external link is fetched at build and dropped unless it answers 2xx (~180 of ~780 drop: template sources and previews, dead app URLs). Trusted without a fetch: URLs the build makes on api.hanzo.ai and persona dirs in hanzoai/personas.
Removed listings: the build reads the live sitemap.xml and retired.json; a path that left gets a moved page (instant meta refresh + canonical to its shelf, noindex) and a moved twin, carried in retired.json. A true 301 needs an edge include in universe.
Agent files (public/)
robots.txt (AI crawler group, Content-Signal: search=yes, ai-input=yes, ai-train=yes, private paths disallowed, Sitemap; no Bytespider group, as before), auth.md (hanzoskills.com's text, H1 # auth.md), .well-known/api-catalog (RFC 9727; marketplace anchor). WebMCP: src/webmcp.ts, search_market + get_listing, read-only, feature-detected; tested in e2e/webmcp.spec.ts with a recording polyfill.
Edge (universe, owned by the coordinator)
hanzo-market wants headers: [agent-link], catalog: true (linkset type on /.well-known/api-catalog), markdown on pages, and spaIndex: 404.html (unknown paths get the noindex shell, not the front page's content). A real 404 needs spa: false + notFound: 404.html with a spa route for the dynamic app paths (/checkout/, /jobs/, /sell/listings/, /l/, /mcp/). Trailing-slash misses (/x/) already answer a bare 404; every real page answers with and without the slash.
Test
bun run lint && bun run typecheck && bun run coverage && bun run build (the CI gate). bunx playwright test (dev server, platform mocked). buyer.spec.ts "browse → listing" has been red since c0d7773 (catalog counts changed with the personas shelf).