Serve full page content to AI page-action buttons on long pages

[jon-neher]

What & why

The AI page-action buttons ("Open in Claude", "Open in ChatGPT", etc. in PageActions.tsx) prefill an assistant with the current page's Markdown by embedding it into the tool URL. On long pages that embedded prompt exceeds MAX_AI_URL_LENGTH (8000), so the content was binary-search-truncated — the assistant received an incomplete page.

This change stops truncating. Long pages now link the assistant to a complete, same-origin raw-Markdown file instead. Short pages are unchanged.

An alternative would be to link the assistant to the GitHub source file, but that would expose the internal org/repo name to end users and break the gray-label rule (.cursor/rules/gray-label-business-owner-docs.mdc). Keeping the fallback on the reader's own domain avoids any brand leak.

How it works — three coupled pieces

1. docusaurus/plugins/raw-markdown/ (new local plugin) — a postBuild step that writes each doc's complete source to build/<permalink>.md (frontmatter stripped, title reinstated as an H1). No new npm dependency. Runs once after HTML generation; cost is negligible.
2. nginx.conf — a location ~ \.md$ block that serves those files as text/plain; charset=utf-8 (render inline, not download) with X-Robots-Tag: noindex (never search-indexed). Placed before the SPA catch-all so .md requests hit the real file instead of the index.html fallback.
3. docusaurus/src/components/PageActions.tsx — buildAIUrl now falls back to the <permalink>.md URL (derived from window.location via the new toMdUrl helper) when the fully-embedded prompt won't fit, instead of truncating. Shared by all five AI buttons, so each gains the complete-content fallback.

Docs added: docusaurus/plugins/raw-markdown/README.md and an "AI Page Actions" section in CLAUDE.md, both flagging the gray-label constraint so a future edit doesn't swap the same-origin fallback for a GitHub-source link.

Net behavior change: only long pages change — from truncated inline to complete via same-origin .md. Short pages are byte-for-byte identical to today.

How to test

1. Build emits the .md files

cd docusaurus && npm install && npm run build
find build -name '*.md' | wc -l          # expect one per doc page (~430+)
# spot-check a long page: frontmatter gone, title is an H1, full body present
cat build/business-app/ai/ai-workforce/ai-chat-receptionist/connect-the-ai-receptionist-with-servicetitan.md

Confirm build time is essentially unchanged. Note the index-page case does not collide: ai-chat-receptionist.md (file) coexists with the ai-chat-receptionist/ directory.

2. nginx serves .md correctly (this is the part not yet verified locally — no Docker daemon in the authoring env)

Build the image and run it (mirrors the Dockerfile: build → /usr/share/nginx/html, nginx.conf → /etc/nginx/conf.d/default.conf), then:

curl -I http://localhost:<port>/business-app/ai/ai-workforce/ai-chat-receptionist/connect-the-ai-receptionist-with-servicetitan.md

Expect: 200, Content-Type: text/plain; charset=utf-8, X-Robots-Tag: noindex. Open the URL in a browser and confirm it renders inline as text (no download prompt). Confirm normal HTML routes still load.

3. Button behavior in the running site (npm run start or the built image)

• Short page → click "Open in Claude"; the tool URL still embeds the Markdown inline (unchanged).
• Long page (one that previously truncated) → the prompt now references …/<page>.md instead of a truncated slice. Open it and confirm the assistant can read the full page. Decode the q= param to inspect the prompt.

4. Gray-label check
Decode a long-page prompt and grep it for github and your internal org/repo name — neither should appear. The .md URL host must equal the page host (whatever partner/gray-label domain is serving it).

Out of scope

No changes to docs/ content or .github/workflows/.

[maryam6samadi]

Thanks for catching this bug. Long pages were sending silently truncated content to LLMs. No good, and that needs a fix. However, this solution seems to be overly complicated for the problem it's solving. To go from "truncated content" to "clean Markdown," it adds a build plugin, an nginx rule, ~430 emitted .md files, and a permalink-to-filename mapping duplicated in two files that must stay in sync with nothing enforcing it. A custom slug or the WordPress Hosting index pattern can point the button at a .md that was never written, which can cause a silent 404 nobody would notice.

Simpler fix: instead of building a parallel .md delivery pipeline, we can just stop truncating and point the assistant at the page's own URL:

"The full page is at . Please read it before helping me."

That's a few lines in buildAIUrl. The only thing we give up versus the full PR is serving pre-parsed Markdown instead of letting the assistant parse the HTML itself, which is a modest quality difference.

[jon-neher]

@maryam6samadi that works for me!

Previously, agents were not able to grab the actual content from these pages. That's why I came up with the markdown serving pipeline.

It seems like agents can fetch the comment easily now so it's not a problem anymore!

[maryam6samadi]

@maryam6samadi that works for me!

Previously, agents were not able to grab the actual content from these pages. That's why I came up with the markdown serving pipeline.

It seems like agents can fetch the comment easily now so it's not a problem anymore!

@jon-neher That's fair. Things change fast. Would you rather trim it yourself or shall I make the change?

[jon-neher]

@maryam6samadi that works for me!

Previously, agents were not able to grab the actual content from these pages. That's why I came up with the markdown serving pipeline.

It seems like agents can fetch the comment easily now so it's not a problem anymore!

@jon-neher That's fair. Things change fast. Would you rather trim it yourself or shall I make the change?

Please go-ahead and make the change! I can review later if you need a reviewer.

[maryam6samadi]

@maryam6samadi that works for me!

Previously, agents were not able to grab the actual content from these pages. That's why I came up with the markdown serving pipeline.

It seems like agents can fetch the comment easily now so it's not a problem anymore!

@jon-neher That's fair. Things change fast. Would you rather trim it yourself or shall I make the change?

Please go-ahead and make the change! I can review later if you need a reviewer.

I committed the change. It's now ready for your approval and a possible merge.