fix(docs): serve agents real pipe tables and unbroken code blocks

[IgorShevchik]

Ports 12722057, partially. Cursor moves to it, which is nuxt/ui v4 HEAD.

Why a commit on skipped infrastructure was not a skip

The subject is upstream's agent-discovery surface, which this fork deliberately does not have: 9ef3ee39 built it, ae243116 refactored it into nuxt-agent-discovery, and both are recorded as skips. The reflex answer here was a third skip, and it would have been wrong.

Reading the diff rather than the subject splits it in two. Half is the module — the ^0.1.1 → ^0.4.0 bump, the prerenderRoutes() deletions that depend on it serving those routes instead — and none of that exists here. The other half fixes two defects in minimark/stringify, which this fork uses through the same call in docs/server/routes/raw/[...slug].md.get.ts.

The defects, measured on our own dependency

node	what comes out
table	a literal <table> block — the format has no pipe-table handler
pre whose code holds a fence	the handler always opens three backticks, so the inner pair closes it early: the rest becomes prose plus a second block

Not hypothetical here. docs/content/ carries markdown tables in 28 files and four-backtick fences in 7 (the typography pages documenting code blocks), and all of it reaches /raw/**.md — which is what skills/b24-ui-nuxt/references/components.md points agents at.

The port

docs/server/utils/markdown.ts is taken as upstream wrote it. The two passes are written against our transformMDC.ts rather than applied as a patch — ours is 1202 lines and diverges in its core (castMap rather than CAST_TEMPLATES, our own generateComponentCode, generateThemeConfig, convertIcon, processLinks). Upstream's other additions in the same commit belong to its own MDC vocabulary and were left alone.

The ordering is load-bearing here and is not upstream

Upstream absolutises links after stringifying, inside the module, so its table pass can sit anywhere. Ours does it on the tree — processLinks rewrites a.href in place — and pipeTable flattens an a node into [text](href) text. Render the tables first and those hrefs are strings before anything rewrites them: relative, and pointing nowhere from a raw file served to an agent.

Confirmed in built output rather than argued — a generated cell reads [TanStack Virtual](https://tanstack.com/virtual/latest/docs/api/virtualizer), rendered and absolute.

The llms.txt half turned up a gap of our own

Widening the filters is not the cosmetic tidy-up it looks like. LIKE '/docs/components/%' matches everything under the section but not the section's own page, and ours exists — docs/content/docs/2.components/0.index.md. The listing carried all 120 component pages and not the page introducing them.

section	before	after
Components	120	121
Typography	absent entirely	20

The Typography gap is ours independently of upstream: 20 pages under docs/content/docs/4.typography/ matched none of our four sections, so the prose components were missing from llms.txt — the only occurrence of the word in the whole file was inside another page's description.

Not taken: the nuxt-agent-discovery bump, the prerenderRoutes() deletions, and a minimumReleaseAgeExclude entry for nuxt-agent-discovery@0.4.0 — a supply-chain exception for a package this fork does not install.

Tests

test/utils/raw-markdown.spec.ts, twelve cases, running against the real stringify rather than a description of it. Two assert the premise — that minimark writes a table as HTML and breaks out of a self-fencing block — so if the dependency grows a table handler the workaround is reported as dead weight instead of quietly staying wrong.

Five mutations verified, each failing its own assertion: fence fixed at three backticks (3 red), pipe escape dropped in a code span, link collapsed to its text, short row not padded, header separator dropped.

A dependency this repository had not declared

Importing the real stringify needed minimark at the root — docs/package.json declares it, the root did not, and the root is what typechecks test/. The gate caught it from two sides, each easy to misread alone:

• typecheck → TS2307: Cannot find module 'minimark/stringify'
• test → one file failed with zero failed tests — the shape of an import error, in the vue project

Declared as minimark: ^0.2.0, the range docs already uses. The package was in the tree already, so the lockfile gains three lines (the importer entry) and resolves nothing new. dep-parity.json needs no reconciliation: it compares only packages both trees declare in the same section, and upstream declares minimark under docs.

Verification

lint · typecheck · build (3.86 MB) · test — 342 files, 7838 passed, 6 skipped · test:module · install --frozen-lockfile.

docs:generate was run separately against these changes — it is where the /raw and llms.txt numbers above come from.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JS8ypVfQSFzYVZzkTHhURb

---

Generated by Claude Code