v0.17.0

AI assistants can read a private deployment as the person who connected
them, through an MCP server the site's Worker serves; it is off unless a
project turns it on. Search returns the passages that match, and the content
health page names documents too long for an assistant to read whole. See
ADR-041,
ADR-042
and
ADR-043;
the last two are proposed, and their numbers are starting points, each a
setting.

Breaking

• The sync section of site.config.json refuses a setting it does not
  know, as mcp does, so a misspelled line fails instead of falling back to
  its default. Remove any stray key from it.

Added

• An MCP server, with "mcp": { "enabled": true } in site.config.json,
  on a deployment whose every environment is private and has signIn.
  • Any assistant connects at https://<host>/mcp — claude.ai, ChatGPT,
    Claude Code, Cursor — by CIMD or dynamic registration, through one
    consent page and the site's own Google sign-in. The grant keeps the
    person's Google ID only, and every request needs the kb:read scope.
  • search and fetch read the person's groups from the directory snapshot
    on every request, so an assistant sees exactly what its person may open
    on the site, and a stale snapshot opens only what every member reads. A
    document is judged by the build's own list, never by the bucket.
  • search returns up to ten documents, each with up to three passages that
    matched, its folders and when it changed. A passage counts only when the
    reader may read the class it came from. fetch returns a document's
    Markdown, cut at mcp.fetchCharacters.
  • Every five minutes the Worker publishes the build's Markdown to an R2
    bucket and asks AI Search to index it; a rollback publishes its own build.
  • mcp.search and mcp.fetchCharacters tune the search per project, each
    optional and checked against the range AI Search accepts.
  • ctcdocs-sync validate requires, with the server on, each environment's
    own OAUTH_KV namespace, KB_DOCUMENTS bucket and KB_SEARCH instance,
    the global_fetch_strictly_public flag and the five-minute cron. With it
    off, the bindings may stay.
  • The access smoke test checks that /mcp asks for a token after a deploy;
    after a rollback, which may restore a version without the server, it does
    not require it.
  • The denial suite publishes the fixture build and asks both tools for
    every document as each reader.
• Long documents on the content health page. The sync measures every
  page's Markdown version, the text fetch returns, and notes a Google Doc
  over sync.largeDocumentCharacters (40,000 characters unless the project
  sets it) under Improve, as worth splitting, and one over
  mcp.fetchCharacters (100,000 with the server off) under Fix next, since
  assistants read only its beginning. A long PDF is noted, not told to split.
  The sync report records the lines it measured against, and the denial suite
  requires the documents fetch cuts to be the ones the report names.
• @ctcstack/ctcdocs-core/published-markdown, the one serializer of a
  page's Markdown version, which the site serves and the sync measures.

Changed

• The Pagefind check searches for the first document of each format before
  the rest, so a PDF's page is always among its cases.
• The rollback workflow tells the smoke test that it restores an older
  version.

Upgrade note

Bump the packages and the workflow pins. Check that the sync section holds
only generatedBy, commitBotName, defaultLocale, largeImageMegabytes
and largeDocumentCharacters. The next sync writes the length notes.

To turn the MCP server on, follow
Step 16
and Cloudflare setup: per
environment, an OAuth namespace, a documents bucket and an AI Search instance,
and the account's AI Search token once.