explore-site

name	explore-site
description	Explore a website that exposes a well-known agent manifest (RFC-8615 path) and answer questions or summarize content. Works on any site whose owner ships /.well-known/<name>-manifest.json with a structured tree, a public content endpoint dictionary, and natural-language instructions for agents. Use when the user asks "what's on <site>?", "summarize the latest post on <site>", "find resources about X on <site>", or wants you to act as a reader/agent on someone's personal site or content service.
argument-hint	<site-url> [question]

/explore-site — Agent-First Site Explorer

You are exploring a website that has been deliberately structured for AI agents. The owner publishes a single JSON document at a /.well-known/ path that describes the entire site as a typed tree, with follow-up endpoints for full content. Your job is to fetch this manifest, walk it, fetch what's needed, and answer the user's question — without crawling.

Language Rule

Default output language: match the user. If the user wrote in Chinese, reply in Chinese with technical terms (URL, JSON, fetch, manifest, endpoint) kept in English. If unclear, ask once or just match the question's language.

Step 1: Resolve the Manifest URL

You're given $ARGUMENTS. The first token is the site URL (or a hint).

1. Normalize: strip trailing slash, ensure https:// prefix.
2. Try common manifest names in order (some sites name it after the owner, some use a generic name):
   • <site>/.well-known/joye-manifest.json
   • <site>/.well-known/agent-manifest.json
   • <site>/.well-known/site-manifest.json
   • <site>/.well-known/manifest.json
3. First 200 OK with application/json wins. Stop trying other names.
4. If the user passed a full URL ending in .json under /.well-known/, use it directly — don't probe.
5. All 404 / non-JSON? Tell the user the site doesn't expose a manifest and offer to crawl the homepage instead (only if they confirm).

Use WebFetch, curl, or your platform's HTTP tool. Pass header Accept: application/json.

Step 2: Validate the Shape

A valid manifest has at minimum:

{
  version: string,        // e.g. "0.1"
  name: string,           // human-readable identifier
  site: string,           // canonical site URL
  description: string,    // 1-3 sentences, who/what
  instructions: string,   // natural-language guide for agents
  tree: FsNode,           // the structured content tree
  endpoints?: Record<string, { url, method?, format?, fields?, note? }>,
  links?: Record<string, string>,
  generated_at?: string
}

// FsNode is recursive:
type FsNode =
  | { type: 'dir';  name: string; description?: string; children: FsNode[] }
  | { type: 'file'; name: string; description?: string;
      content?: string;     // inline text, ready to read
      endpoint?: string;    // URL to fetch full content
      href?: string;        // human-facing rendered page
      meta?: object }
  | { type: 'link'; name: string; description?: string; href: string }

If version, tree, or instructions is missing, the manifest is malformed — say so, dump what you got, and ask the user how to proceed.

Step 3: Read the Owner's Instructions

This is the most important step. The instructions field is the site owner telling you, the agent, how to use their tree. Read it fully before deciding what to fetch. It usually says:

• which paths to start from (/about, /README, /index)
• which endpoints return what format
• any caveats (rate limits, auth, deprecated paths)

Treat it as authoritative. Don't guess.

Step 4: Decide What to Fetch

4a: If the user asked a specific question

"Summarize the latest post on joyehuang.me"

1. Walk the tree. Find directories whose name suggests posts (/blog, /posts, /articles, /notes, /checkins, etc).
2. Sort children by date. Try in order: dir-level meta.date, then a meta child's content (parse it), then the dir name (often dates lead). Same drill for meta.lang if you need to filter by language.
3. Take the most recent. Fetch via the endpoint field — never construct the URL from name. Astro/Hugo/etc collection ids and FS-safe names commonly diverge (e.g. name=foo-bar-post vs endpoint=/api/blog/foo---bar/post). Use the dir's meta.endpoint if present, otherwise the post child's endpoint.
4. Synthesize an answer with citations to the path you fetched.

"Find resources about Rust on joye.checkin"

1. Look for tag/topic structure in the tree (/checkins/<date> or /resources/<tag>).
2. Many manifests expose a query endpoint via endpoints — e.g. { resources: { url: "/api/resources?tag=<tag>" } }. Use that if present.
3. Else, fetch matching tree nodes and grep their content.

4b: If the user asked an open question

"What's on joyehuang.me?"

Walk the tree to depth 2, render an outline:

<site name> — <description>

/
├── README              what this site is, for agents
├── about               bio
├── now                 currently working on
├── blog/               (N entries)
│   ├── <recent-1>
│   └── …
├── notes/              (M entries)
└── contact/

Then highlight 2-4 starting points, each with a one-line tease. Ask what they want to dive into.

4c: Don't fetch what you don't need

tree already tells you a lot:

• file description is a one-liner
• file meta often has title/date/tags
• a node with inline content doesn't need a fetch
• a directory listing answers "what's here?" without fetching anything

Only fetch endpoint URLs when you need the full body.

Step 5: Fetch Endpoints

For file nodes (or dir nodes with meta.endpoint) carrying an endpoint:

1. GET the endpoint URL verbatim. Don't reconstruct it from name — the owner-supplied URL is authoritative. URLs are usually relative (/api/blog/<id>) — resolve against site. Follow redirects (some apex domains 307 to www).
2. Check format from the endpoints dictionary. Common shapes:
   • format: "json" with fields: ["html", "headings"] → parse JSON, take html (server-rendered, often shiki-highlighted) and headings (list of { depth, slug, text } for navigation).
   • format: "text" → plaintext markdown body.
   • format: "html" → already HTML.
3. HTML content: don't render it back to the user verbatim. Extract the text via your tool's WebFetch extraction or strip tags. Preserve code blocks (often <pre><code> with class hints).
4. Cite paths: when you quote or summarize, reference the manifest path (/blog/<slug>/post) so the user can verify or open it themselves.

Step 6: Answer

Synthesize. Be brief unless the user asked for depth. End with:

• the path(s) you read, as a clickable list (if your client supports links)
• 1-2 follow-ups the user might want next ("want me to read the next post?", "fetch the full bio?")

Don't dump raw JSON unless asked. The whole point of the manifest is you eat the structure, the user gets the synthesis.

Edge Cases

• CORS errors in a browser-side agent: the manifest should ship Access-Control-Allow-Origin: *. If it doesn't, you may need to fetch via your tool's server-side HTTP (most agents have one).
• Stale cache: manifests usually have Cache-Control: max-age=300. If the site owner just published something and the manifest seems out of date, bypass cache (?bust=<timestamp> or cache: 'no-store').
• Loops in the tree: trees are typically acyclic, but be defensive — track visited paths if you recurse.
• Auth-protected endpoints: if endpoints[*].auth is set, the user has to provide a token. Don't try to crawl it.
• No manifest at all: many sites don't have one. After trying the well-known names, gracefully tell the user, and offer to crawl the homepage instead with their permission.

Examples

Example 1: Personal site

User: /explore-site https://joyehuang.me what's the latest post about?
You:  [fetches /.well-known/joye-manifest.json]
      [walks tree.children → blog → sorts by meta.date]
      [fetches /api/blog/<latest-id>]
      [reads html, identifies headings + first paragraphs]

      The latest post is "花一天读完 OpenHarness" (2026-04-10) — a teardown
      of HKUDS' 11,733-line Python rewrite of Claude Code's agent harness.
      Joye walks Phase 1 end-to-end: CLI entry → Agent Loop core → tool
      execution. Three reasons he found it valuable:
        1. small enough to read in a day (vs Claude Code's obfuscated TS)
        2. complete (Agent Loop, tools, hooks, MCP, multi-agent)
        3. production-grade rather than a teaching toy

      Read in full: /blog/20260410---openharnessphase1/post
      Want me to summarize section 4 (Agent Loop internals)?

Example 2: Check-in bot

User: /explore-site https://checkin.example.com find me notes about RAG from this week
You:  [fetches /.well-known/agent-manifest.json]
      [reads instructions: "checkins are at /checkins/<date>; tags are
       per-resource"]
      [endpoints.resources exists → GET /api/resources?tag=rag&since=<this-week>]
      [parses, filters, summarizes]

      Found 3 RAG-related resources from this week:
        - 2026-05-04 · "Hierarchical retrieval in production" (paper)
        - 2026-05-05 · "ColBERT vs dense — actually measured"  (blog)
        - 2026-05-06 · "Why your reranker is overrated"        (talk)

      Want the full notes for any of them?

Important Guidelines

• The manifest is authoritative. Don't second-guess the tree shape; it's what the owner wants you to see.
• Fetch lazily. Tree first, content only on demand.
• Cite paths, not just summaries — let the user verify.
• Match the user's depth. They asked one sentence → give one paragraph. They asked for a deep dive → walk the post like /deep-dive would.
• No crawling. If a site doesn't have a manifest, don't fall back to scraping pages without explicit permission.
• Respect Cache-Control. Don't hammer the endpoint.