lib-docs

llms.txt is author-published — no curation lag, no enrichment layer that can hallucinate — so prefer it over context7.

1. Probe: bash scripts/llms-probe.sh <docs-domain> — outputs TSV kind \t url \t size (index = llms.txt page list, full = llms-full.txt whole corpus). Paths vary per site; probe instead of hardcoding URLs.

2. Fetch: full ≤ ~500 KB → WebFetch it directly. Larger or size ? → WebFetch the index, then fetch only the matching page links, in parallel. Cloudflare's llms-full.txt is ~46 MB — a blind full fetch blows the context window.

3. Probe failed → context7 via the ctx7 CLI (one-time bunx ctx7@latest login):

   bunx ctx7@latest library <name> [query]      # resolve library ID first — IDs aren't guessable
   bunx ctx7@latest docs <libraryId> "<query>"  # specific queries rank far better than broad ones
   

   Retry with --research only when the default answer is shallow — it spawns sandboxed agents and costs more.

Most Mintlify/GitBook-hosted docs publish llms.txt; Tailwind and most pre-1.0 or community libraries don't — those go straight to context7.