Markdown mirror is actually markdown
Agents trust a `Content-Type: text/markdown` response to be markdown. Serving the rendered HTML page with a markdown content type defeats the purpose of having a mirror — the agent still has to run an HTML parser and extract the article, and the size-reduction win is gone.
How the check decides
The check fetches the markdown mirror and looks for two failure shapes:
- The body starts with an HTML prologue —
<!doctype html>,<html, or<?xml. Fails immediately. - More than 30% of the body’s characters live inside HTML tags. Fails with the measured ratio.
A body that contains occasional inline tags (<br>, <sup>, embedded <img>) easily passes. Returns N/A if no mirror was published at all.
How to implement it
Serialise the source markdown directly rather than converting from rendered HTML. If your build pipeline can only post-process HTML, run a real HTML→markdown converter (turndown, pandoc) rather than just changing the response’s content type.
Pass
---
title: Install a14y
---
# Install a14y
Run `npm install -g @a14y/cli`, then `a14y https://example.com` to audit a site.
- Reads `llms.txt` if present.
- Falls back to crawling the sitemap.
Fail
<!doctype html>
<html>
<body><h1>Install a14y</h1>...</body>
</html>