Workflow
1. Search — docfork:query_docs
library: Start with a short name or keyword (e.g., nextjs, react router) — this triggers multi-library search with server-side reranking. Once a result is returned, extract the owner/repo from its URL and use that exact form for any follow-up calls to pin to the verified repository.
query: Specific and descriptive — e.g., "server actions in Next.js App Router" not "next.js actions". Include version if specified (e.g., "React 19 concurrent features").
Each result returns a title, description, and url. Prefer canonical/official repos over forks. Retry with a broader or different query if results are empty or off-target.
3 calls maximum across the entire request.
2. Fetch - docfork:fetch_url
Search results are summaries. Call fetch_url to get the actual content - this is the default next step after a relevant result, not a last resort.
Two URL modes, both derived from search result URLs:
Exact chunk - pass the URL as-is, keeping the line anchor (#L40-L85):
https://github.com/vercel/next.js/blob/main/docs/routing/middleware.mdx#L40-L85
Table of contents - strip the filename and anchor, pass the parent path:
https://github.com/vercel/next.js/blob/main/docs/routing/
Returns a TOC with previews across that directory - use when you need broader context before committing to a specific chunk.
If a chunk's description is truncated or vague, fetch it - descriptions are previews only.