Make the Replicate page fast

[skyfallwastaken]

The Replicate page was slow for three reasons, each fixed here:

1. 22.7 MB of cover images. Cards loaded Replicate's original covers, several of them 2–6 MB PNGs and GIFs. Covers are now served from /replicate/covers/<owner>/<name> as 640 px WebP thumbnails, made with Bun.Image and cached in memory per source URL for a day. When a cover can't be re-encoded (GIF, video, fetch failure), the endpoint redirects to the original. Only covers of listed models are served, so it isn't an open proxy.
2. Full model objects in the page data. Each card shipped Replicate's whole model object, including default_example and the OpenAPI schema. The page now gets only the fields a card renders.
3. Slow first visit after a deploy. The catalog was built on the first visit, which took 2–7 s (one model fetch and one pricing fetch per model). backend.start() now warms the catalog and the cover thumbnails in the background, which covers both entrypoints.

Also fixed along the way: the aspect-video cover box grew to the image's own height, so rows were uneven. The image is now absolutely positioned inside the box.

Numbers

Local dev server, 29 cards, scrolled to the bottom:

	Before (neo)	After
Images transferred	22.69 MB	0.58 MB
Page HTML	343 KB	168 KB
Page data, first request after a cold start	1.69 s (6.6 s on an earlier run)	0.02 s
A cover, first request after start	original from Replicate's CDN	~5 ms, 7–50 KB

Screenshots

Before (neo): uneven cover heights	After

Testing

• read-model.test.ts:
  • cards carry only card fields;
  • a 1600×900 PNG cover comes back as a 640×360 WebP;
  • a cover that can't be decoded redirects to the original;
  • an unlisted model gets no cover.
• bun run typecheck, bun run check: 0 errors. bun test: 0 fail.

🤖 Generated with Claude Code