v0.47.2

Release Notes — v0.47.2

Patch: A Gemini thinking model keeps its footing across tool calls

On a Gemini thinking model, the second turn of a conversation that involved a
tool call would fail. The model attaches a signed token to each tool call it
makes and expects that exact token handed back on the next request; Ogcode was
dropping it, so the follow-up turn was rejected. It is now carried the whole
way — captured as the reply streams in, kept with the tool call it belongs to,
and replayed untouched on the next request. Every other provider is unaffected:
they neither attach the token nor look for it.

Other changes

• Search is quicker and better aimed: only your query goes to the search engine,
  not the surrounding instructions; ranking is skipped when there are already
  few enough candidates; page fetches share a fifteen-second budget and a quorum,
  so the slowest page no longer holds up the answer; and the summary falls back
  to the top-ranked pages when it needs to.
• Tavily fetches a page's content as markdown in the same request and caches it,
  so a page you settle on is usually already in hand; the built-in reader parses
  each page once instead of twice.
• Safari search waits out page rendering only where it matters and polls the page
  more cheaply while it does.
• The Windows browser backend tidies up after itself — it reaps its headless
  process cleanly and removes its temporary profiles — and reports no error when
  it simply finds nothing.
• The model catalogue knows the newest Google flagship, Gemini 4 Argon, along
  with the -latest names Google rotates (Gemini Pro, Flash, and Flash-Lite).