Expert references for Advanced UI and Multi-Agent, PyRIT in Security, and a raw-markdown fix

[amirkiarafiei]

Three things, one of which is a fix for a defect currently live on main.

References to keep

Thirteen links for Advanced UI (agentic and generative UI, the A2UI and AG-UI protocols, CopilotKit's three pieces, the renderers) and three for Advanced Multi-Agent (langgraph-swarm, langgraph-supervisor, and the workflow-versus-agent page).

Every URL was fetched before publishing: all sixteen return 200. Two are not what their names suggest, so the notes say what they actually are:

• AGenUI is a native A2UI renderer for iOS, Android and HarmonyOS, not anything to do with AG-UI
• the two langgraph package pages are the JavaScript/TypeScript references, unlike every other LangChain link in this corpus

These deliberately do not appear on the site. A draft renders only its first paragraph and its schedule of parts, which is why advanced_training.md's reference list has always been GitHub-only too. They are there for whoever opens the file to write the module.

advanced_multiagent_tr.md had drifted, missing two topics and the entire reference block its English sibling carried. All ten expert modules now match exactly on links, bullets, blocks and diagrams.

PyRIT

It existed only as "It is built on PyRIT" inside another tool's bullet, with no link. It now has its own entry among the red-teaming tools, placed before the agent built on it so that sentence reads as a back-reference.

Details come from the verified dossier in the scratchpad, so it carries the things that save time: the four pieces a run is assembled from, the attacks that ship as runnable classes (which ties back to this module's own jailbreak taxonomy), pyrit_scan for CI, and two traps — Orchestrator became Attack in December 2025, and Azure/PyRIT is archived rather than redirected, so a stale link still serves a page that looks current.

The fix: a draft's dek is plain text

The first paragraph of a draft becomes the sheet dek and the page's <meta name="description">, and neither is passed through the markdown renderer.

The earlier pass that replaced bare module numbers with title links put links in exactly that slot, so six expert stubs have been shipping literal brackets — on the page, and in the metadata that feeds search results and link previews. Nothing caught it: the link gate inspects href attributes and links.test.ts hunts hrefs ending in .md; this was text.

Fixed by naming the module by title in plain text there, which still satisfies the rule that pass existed to serve. A fresh build now contains no literal markdown link text anywhere in out/.

One test removed

leaves every stub under 200 words went red at 231 when Advanced UI gained its links. A stub carrying a reference list is still a stub; CLAUDE.md names a word count as the first thing an assertion here may not state; and this same file had already dropped a sibling word-count test for the identical reason. A comment records it. What it reached for is enforced properly in curriculum-file.ts, where a ready module needs a summary and minutes above zero.

Gate

Check	Result
npm run typecheck	clean
npm test	2052 passed
npm run build	58 pages, every internal link resolves
raw markdown in out/	0 files
EN/TR structural parity, expert set	0 mismatches