This site's model

Every page of this site publishes its own ai-vision model. Open the site in Persephone’s built-in browser and your agent finds it at pages["<id>"].editor.app. From there it reads the page by section, searches the site and moves between pages. It doesn’t have to parse the page snapshot.

It’s a small model on purpose. The site is static documentation, so there isn’t much for an agent to do here. What it shows is how little code a site needs to plug into the same call tree that Persephone’s editors and boards use.

Try it

Open https://andriy-viyatyk.github.io/ in Persephone’s browser and ask your agent, for example:

• “What does this site say about installing boards? Open that page and show me the steps.”
• “Which Persephone boards open .torrent files?”
• “Switch this site to the light theme and show me where the theme picker is.”

What the model exposes

Path	What it is
app	Summary: the current page, the number of pages and boards, the colour theme.
app.page	The page in this tab: title, headings, clips, text, section(heading), scrollTo(heading).
app.pages	Every page with its path; index by position or by path, pages["/persephone/"].
app.boards	The Persephone board catalog; index by id, boards["torrent-viewer"].
app.search(query, limit?)	Full-text search over the site, with an excerpt for each match.
app.open(path)	Go to another page. The tab reloads and the new page publishes a fresh model.
app.theme	The colour theme, "auto", "dark" or "light"; writable.
app.elements, app.highlight(name, message)	The search box, theme picker, sidebar, table of contents and demo clip, pointed at on screen.
app.helpSearch(query)	Find a member or control by what it’s for.

A first call returns the summary and a hint that lists these members:

call pages["<id>"].editor.app
→ { kind: "Site", page: "Workspace", path: "/persephone/workspace/", pages: 21, boards: 11, theme: "auto" }

call pages["<id>"].editor.app.search  args ["install a board", 3]
→ [{ title: "Installing boards", path: "/persephone/boards/installing/", excerpt: "…" }, …]

How it’s built

It’s one script loaded on every page. It loads a small index of the site, describes the model, and calls expose():

import { expose } from 'ai-vision/remote';

const root = {
  aiVision: {
    kind: 'Site',
    summary: 'The andriy-viyatyk.github.io site: docs and demos for Persephone, ai-vision, av-grid and boards.',
    members: [
      { name: 'page', kind: 'property', node: true, summary: 'The page open in this tab.' },
      { name: 'search', kind: 'method', signature: 'search(query, limit?)', summary: 'Full-text search over the site.' },
      // …
    ],
    summarize: () => ({ kind: 'Site', page: document.title }),
  },
  get page() { return page; },
  search,
  // …
};

expose(root); // publishes window.__aiVision; Persephone's browser finds it

Outside Persephone nothing reads window.__aiVision, so the page works as usual. Persephone treats the model as page-authored content. It labels results as written by the page, keeps them under .app, and never probes Incognito or Tor tabs.

The full source is src/scripts/site-model.ts. If you can’t change a site’s code, a Site Extension adds a model to it from the outside.