Recursive find (rfind)

Use when the question is "where on this site / in these docs / among these listings does X appear?" One rfind call replaces dozens of web_fetch calls and returns only hit pages with short snippets, not page bodies. Headless port of the Recursive Find Chrome extension (github.com/grtwrn/recursive-find).

Skip it for: reading one known page (web_fetch), JavaScript-rendered sites or logged-in pages (browser), or general web questions (web_search).

Run

rfind <url> "<query>" [--depth 1] [--within /docs/] [--max-pages 100] [--mode text|word|regex] [--case] [--main] [--first] [--json]

(rfind is scripts/rfind.mjs in this skill, symlinked onto PATH; Node 18+, no dependencies. rfind --help lists every flag.)

1. Start from the most specific page you know (a docs index, a category page), not a site's home page. Big home pages link to hundreds of pages and burn the page budget before reaching the right section.
2. Narrow with --within <path-prefix> whenever the content lives under one path (e.g. --within /docs/). Keep the default --depth 1; use --depth 2 only with --within or a small site.
3. Use --mode word for short words (cat won't match category), --mode regex for patterns, --first when one hit answers the question.
4. Add --main when the question is which pages discuss X: it searches only the main content (<main>, role="main", <article>), so "Next topic" links, sidebars and menus that merely name X don't count as hits.
5. Read the summary line: if it says hit --max-pages, results are incomplete. Narrow (--within, a deeper start URL) before raising --max-pages (max 1000). It also flags pages over 32 MB that were only partly searched. Skipped non-HTML links are normal, not failures.
6. Matches show as [[match]] inside snippets. Cite the page URL; web_fetch it if you need more than the snippet.

Behavior

• Same site only by default (--any-site to leave it); honors robots.txt (--ignore-robots only when the user asks); 4 parallel fetches; 10 s per page; the whole crawl stops after 120 s (--max-seconds), flagged in the summary.
• Text comes from server HTML only (no JS), no cookies/login. Skips PDFs, images and other non-HTML links.
• Exit 0 = crawl ran (even with 0 hits); 1 = start page failed; 2 = bad arguments.