Hi Kaxil, I can pick this up if you are okay with it.
On Tue, 6 Oct 2026 at 23:28, Kaxil Naik <[email protected]> wrote: > Hi all, > > More and more people write Dags with coding agents (Claude Code, Cursor, > Copilot, Codex), and those agents read our docs as HTML. A provider guide > page is about 150 KB of HTML for about 20 KB of content, and fetch tools > often summarise it down to a paragraph, code samples included. Pydantic AI, > Ray, Camel, Astronomer and most of the popular sites & project publish an > llms.txt index plus a Markdown copy of each page. We publish neither. > > I've opened an issue with a plan and a working proof of concept, starting > with the Common AI provider docs: > > https://github.com/apache/airflow/issues/74376 > > It has three parts: > > 1. Docs build: generate page.html.md next to every page, plus llms.txt and > llms-full.txt, using sphinx-llm and a small extension. The proof of concept > builds cleanly for Common AI with no change in build time. > 2. Publishing: a couple of guards in breeze so stable/ always has the > Markdown copies. > 3. Serving: Claude Code, Cursor, Copilot and OpenCode already send "Accept: > text/markdown" on every fetch. A short .htaccess rule in airflow-site can > answer those requests with the Markdown copy, so agents get it from the > URLs they already use. I tested this on airflow.staged.apache.org; ASF's > Fastly cache keeps the HTML and Markdown answers apart. > > *I won't have time to build this or to review it, so it needs an owner who > can see it through, including reviews.* The issue lists what's done, what's > left, and the rollout constraints. It needs a committer, or a contributor > pairing with one: testing the airflow-site change on staging, publishing > docs and checking the live rollout all need committer access. > > This would be a high impact change, so thanks in advance for whoever picks > it up. > > Regards, > Kaxil >
