Make web research pull Wikipedia, retry failed pages, and match inflections

[devbadya]

A 0.8B model will not recover from a firewall page, a login wall, or a German noun that only appears inflected. research already did the fan-out in code; this makes the pages it opens the ones that actually answer.

What one call does now

1. Searches the live web and Wikipedia together (MediaWiki is free of the reader budget; skipped when Wikipedia is already the provider).
2. Puts the Wikipedia hit first when one appeared, then one hit per remaining site ranked by whether the snippet already bears on the question. Login walls and download_pdf galleries go last.
3. Reads three in parallel. A 429, a Sucuri/Cloudflare interstitial, or a page with no prose is replaced from the remaining hits (up to five reads) rather than quoted. The search snippet only stands in when nothing else could be opened.
4. Scores passages with inflected forms counting as the same term, so Bundeskanzlers answers Bundeskanzler.
5. Keeps Wikipedia's one-line incumbent claim. A live run of Wer ist der Bundeskanzler? used to quote the etymology of Kanzler and never Friedrich Merz — the name sits in Amtsträger ist seit dem 6. Mai 2025 Friedrich Merz (CDU)., shorter than the prose floor and at the end of a 1,500-character lead. Dated claims now join the paragraph they follow, and a dated lead is trimmed to a window that keeps the year.

lookup-term is unchanged — a name still does not need three sources.

Eval

Who wrote Dune? and What's the population of Tokyo? already routed to research-question; the scenarios now expect research rather than web_search.

Verification

pnpm check — 1,210 tests.

Live against the network (no GPU, so this is the tool, not the model):

• Wer ist der Bundeskanzler? (Wikipedia) now quotes Amtsträger ist seit dem 6. Mai 2025 Friedrich Merz (CDU).
• Who is the current secretary-general of the United Nations? (DuckDuckGo + Wikipedia) returns Wikipedia, un.org, and names António Guterres.

The model itself is not exercised here: this environment has no GPU adapter. Whether the skill actually reaches research still needs /?eval in Chrome on a real GPU.