tabler/docs/pages/llms.txt.ts

// https://llmstxt.org — index of the docs, pointing at the .md mirror of each
// page (see [...slug].md.ts). Order and grouping follow the sidebar tree in
// docs.json; any page missing from that tree still lands in "Other", so the
// index can never silently omit a page.
import type { APIRoute } from 'astro'
import { getCollection } from 'astro:content'
import docs from '@data/docs.json'
import { site } from '@shared/lib/site'
import { docsUrlFromId } from '@lib/docs-pages'
import { baseUrl } from '@lib/llms'
import { markdownUrlFromDocsUrl } from '@lib/markdown-url'

export const prerender = true

type MenuNode = { title?: string; url?: string; children?: MenuNode[] }

const normalizeUrl = (url: string) => {
  const parts = url.split('/').filter(Boolean)
  return parts.length ? `/${parts.join('/')}` : '/'
}

export const GET: APIRoute = async () => {
  const entries = await getCollection('docs')
  const byUrl = new Map(entries.map((entry) => [docsUrlFromId(entry.id), entry]))
  const base = baseUrl()
  const listed = new Set<string>()

  const item = (url: string) => {
    const entry = byUrl.get(url)
    if (!entry || listed.has(url)) return null
    listed.add(url)
    return `- [${entry.data.title}](${base}${markdownUrlFromDocsUrl(url)}): ${entry.data.description}`
  }

  const pageUrls = (nodes: (MenuNode | undefined)[]) => nodes.flatMap((node) => [node?.url, ...(node?.children ?? []).map((child) => child.url)]).filter((url): url is string => typeof url === 'string' && url.startsWith('/'))

  const sections: string[] = []
  for (const group of docs.menu as MenuNode[]) {
    // Groups get a heading of their own; pages listed straight under a section
    // (products without sub-groups) are collected into one heading for it.
    const loose = (group.children ?? []).filter((child) => !child.children?.length)
    for (const section of (group.children ?? []).filter((child) => child.children?.length)) {
      const lines = pageUrls([section])
        .map((url) => item(normalizeUrl(url)))
        .filter(Boolean)
      if (lines.length) {
        sections.push(`## ${group.title} — ${section.title}\n\n${lines.join('\n')}`)
      }
    }
    const looseLines = pageUrls(loose)
      .map((url) => item(normalizeUrl(url)))
      .filter(Boolean)
    if (looseLines.length) {
      sections.push(`## ${group.title}\n\n${looseLines.join('\n')}`)
    }
  }

  const remaining = [...byUrl.keys()]
    .filter((url) => !listed.has(url))
    .sort((a, b) => (a === '/' ? -1 : b === '/' ? 1 : a.localeCompare(b)))
    .map((url) => item(url))
    .filter(Boolean)
  if (remaining.length) {
    sections.push(`## Other\n\n${remaining.join('\n')}`)
  }

  // Class reference for every page that carries `classnames` front matter, so an
  // agent reading this file alone knows which classes exist without fetching
  // each page. Grouped by kind, in the same order the page renders them.
  const groupOrder = ['component', 'part', 'style', 'modifier', 'behavior', 'direction', 'color', 'size']
  const classSections = entries
    .filter((entry) => entry.data.classnames)
    .sort((a, b) => a.data.title.localeCompare(b.data.title))
    .map((entry) => {
      const groups = Object.entries(entry.data.classnames as Record<string, { class: string }[]>).sort(([a], [b]) => (groupOrder.indexOf(a) + 1 || 99) - (groupOrder.indexOf(b) + 1 || 99))
      const lines = groups.map(([group, rows]) => `- ${group}: ${rows.map((row) => `\`${row.class}\``).join(', ')}`)
      return `### ${entry.data.title}\n\n[${entry.data.title} documentation](${base}${docsUrlFromId(entry.id)})\n\n${lines.join('\n')}`
    })

  const classReference = classSections.length ? `## Class names\n\nEvery class each component ships, grouped by kind. A name in braces stands for a family: \`alert-{color}\` means any base color, \`table-mobile-{breakpoint}\` any breakpoint.\n\n${classSections.join('\n\n')}` : ''

  const body = `# Tabler

> ${site.description}

Tabler is a free and open source dashboard UI kit built on Bootstrap. This file indexes the documentation at ${site.docsUrl}.

Every page is available as markdown by appending \`.md\` to its url, for example \`${site.docsUrl}/ui/components/button.md\`. Those files contain the same prose as the html page plus the markup of every example, as fenced code blocks. The whole documentation concatenated into one file is available at \`${site.docsUrl}/llms-full.txt\`.

The markup is taken from the rendered pages, so every example is plain html — the same markup the copy button on the page gives you.

An agent skill for building with Tabler is published at \`${site.docsUrl}/.well-known/agent-skills/tabler/SKILL.md\` (index: \`${site.docsUrl}/.well-known/agent-skills/index.json\`).

${sections.join('\n\n')}

${classReference}
`

  return new Response(body, {
    headers: {
      'Content-Type': 'text/plain; charset=utf-8',
    },
  })
}