Skip to main content

v1.4.2 — Search that finds things

Search that finds things, and a research prompt that asks only when you mean it.

Released 2026-10-25. Install or update: mercury upgrade, npm i -g @cosmicstack/mercury-agent@1.4.2, or the GitHub release.

  • Three web sources, the ones that work. web_search asks Parallel's public search first (keyless MCP built for agents, about a second, with excerpts and dates), then DuckDuckGo's html page, then its lite page. Brave, Wikipedia, Hacker News and Bing's web feed are gone from web search: Brave refuses automated requests and the others answered general questions badly. News keeps Parallel (article links only, section fronts dropped), then Bing News, Google News and Hacker News.
  • DuckDuckGo parsed exactly. Both pages are read from their real markup: each result's title, the real address behind DuckDuckGo's redirect, and the snippet; ads (result--ad, sponsored rows) are dropped. A challenge is retried once after a short pause, and a page that is neither results nor "No results" is reported as "page not recognised" instead of silently empty. Tests run against saved copies of both real pages.
  • News links open the article. News search used Google News first, whose links are Google redirects (news.google.com/rss/articles/…) that can't be turned into the publisher's address. Bing News now comes first and its links are unwrapped to the real article; Google News is the fallback, and its links are marked as redirects.
  • Search never depends on the model. Results come only from Mercury's own search tools, never from a provider's built-in web search, so search works the same with every model.
  • The web-search skill (1.3.0) explains the sources and tells the model to report which ones failed and why when a search comes back empty.
  • You can see what happened. Every result ends with the trail, for example Searched: Parallel: timed out → DuckDuckGo: 8 results, so "nothing found" and "everything was blocked" read differently.

Chat​

  • The research prompt asks only when you want research. It used to fire on words ("latest", "recent", "explain", "today's news"), so "convert the latest epub" or "explain how my build works and fix it" got a research question. The words now only decide whether to check; one short model call decides whether you asked for an in-depth, multi-source write-up. Instructions, quick questions and today's news never get the prompt, and a failed or slow check means no prompt.
  • Quick answer comes first. The picker starts on "Quick answer", so an accidental Enter never starts a long research run.