shell-manifest

Push item IDs to a handle during the prerendered render, then have a loader await ctx.rendered() and read the replayed IDs with ctx.get() to issue one batched query. The handle data is stored with the payload and replayed on every cache hit.

Use the shell manifest pattern: the cached shell records which items it rendered via a handle, and a single live loader reads that collected list after the render barrier and fetches all dynamic data in one batched query instead of per-component fetches.

It works with both. Prerender pushes handle data at build time, while cache() pushes on the cache miss; in both cases the handle data is stored with the Flight payload and replayed on every hit, so the loader logic is identical.

Calling ctx.get(handle) before awaiting ctx.rendered() throws in a loader, with an error telling you to await the barrier first. The push form ctx.use(Handle) is legal anywhere in the loader body without the barrier.

ctx.rendered() is experimental and DSL-loaders-only. It throws in fetchable or standalone loader calls, in handler-invoked loaders (a detected deadlock), and in loaders registered with { ssr: false }, and it serializes the loader after the shell render.

No. Manifest handle data is baked into a shared cached artifact, so it must never contain data derived from cookies() or headers(). IDs, slugs, slot names, and variant keys are safe; anything user-specific would leak across requests.