web-research

End-to-end web research pipeline: search → fetch with a free-first cascade → summarise.

Tool routing policy

Search (find candidate URLs): Tavily /search (keyless by default). Defuddle is not a search engine and is not used here.

Fetch (one URL → clean markdown): cascade in this order, stop at first success.

defuddle (CLI, fully free, no rate limit) — preferred when installed. Skip entirely if command -v defuddle returns non-zero. Do not install defuddle automatically; tell the user once if it is missing.
Tavily /extract (keyless, free but rate-limited) — fallback when defuddle is missing or fails on a URL. Header X-Tavily-Access-Mode: keyless is sent by scripts/tavily_extract.py.
mcode in-app browser (MiniMax Code's built-in browser, "Browser Use") — last resort for login-walled, JS-rendered, or Cloudflare-protected pages that the two layers above cannot parse. The pipeline writes a  marker; the model sees the marker and re-fetches that URL with the browser tool.

Each output file is named NN-<slug>.md and prefixed by a tag noting which layer produced it ([defuddle], [tavily extract], or [browser required]).

Default flow

If the user did not specify constraints, run the full pipeline in one shot: