Manifest

A build that only writes images leaves the site to work out for itself what was generated, where it lives and how big it is, which usually means globbing for *-og.png to find the landscape variant, or hardcoding 1200 and 630 into the meta tags. All of that is known while the images are being generated, so pointing manifest at a file writes it down:

export default defineConfig({
  placement: { strategy: "public-dir", dir: "public/og", urlBase: "/og" },
  manifest: "data/colophon.json", // src/data/ for Astro, _data/ for Eleventy
});

{
  "version": 1,
  "pages": {
    "blog/my-post": {
      "images": {
        "og": {
          "url": "/og/blog/my-post-og.png",
          "width": 1200,
          "height": 630
        },
        "square": {
          "url": "/og/blog/my-post-square.png",
          "width": 1200,
          "height": 1200
        }
      },
      "widest": "og",
      "alt": "My post"
    }
  }
}

Every generator in scope reads JSON as native data. Hugo picks it up from data/, Astro imports it, Eleventy and Jekyll read _data/, and Zola has load_data. That makes the meta tags a lookup rather than a convention:

const page = manifest.pages["blog/my-post"];
const image = page.images[page.widest];
// <meta property="og:image" content={image.url}>
// <meta property="og:image:width" content={image.width}>

metaTags will do that part for you.

What is in it

• Pages are keyed by slug, which is what the site addresses a page by, and under slugStrategy: "route" is the route itself. Two pages cannot share a key, and a build that would need them to fails saying so, because a lookup returning the wrong post’s image is worse than no manifest at all.
• widest names the most landscape image by aspect ratio, ties going to the size configured first. It is what a summary_large_image card wants, and the check every site currently writes for itself. Note that og (1200x630) and square (1200x1200) are equally wide, so comparing widths would pick either.
• url is absent where the placement knows none. See Placement. The dimensions are always there.
• alt comes from the props’ title, and is absent for a page without one.
• Extra images are not pages, so they are not listed. A project that named the output path of one already knows where it is.

What it describes

The manifest describes what exists, not what a given run did. A rebuild that skips every image still writes the whole thing, so a build over an unchanged tree still produces a complete record of the site.

It is built from the planned work rather than the results, which is also why a duplicate slug fails before the tree is rendered rather than after.

Pages and sizes are sorted and the JSON is indented, so a manifest committed to a repository changes only when the build does.