Follow-up to #1302.

Summary

get-page now derives its url from the incoming request origin, which is correct for proxied and deployed environments. The remaining problem is that it still fetches its markdown body by issuing an HTTP request back to /raw${path}.md.

That works locally, but on Cloudflare Workers the same-worker subrequest returns 404, so list-pages works while get-page fails with Failed to get page.

This keeps the request-origin URL behavior but removes the HTTP self-fetch. Instead, get-page reads the page body directly from the Content DB, serializes the minimark AST, and reconstructs markdown-like content with frontmatter.

Reproduction

Verify

Broken worker:

curl -sS \
  -H 'accept: application/json, text/event-stream' \
  -H 'content-type: application/json' \
  -X POST https://docus-mcp-get-page-repro.je-cf9.workers.dev/mcp \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"get-page","arguments":{"path":"/test"}}}'

Fixed worker:

curl -sS \
  -H 'accept: application/json, text/event-stream' \
  -H 'content-type: application/json' \
  -X POST https://docus-mcp-get-page-repro-fix.je-cf9.workers.dev/mcp \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"get-page","arguments":{"path":"/test"}}}'

Local source verification:

pnpm typecheck