v3.25.0 — What's new

Every model the plugin chooses is now a current one. Nothing else changed in
what the agents do; what changed is which generation answers when they ask.

• The whole fleet moved to the Claude 5 family. Fifty-seven of seventy
  agents pinned their advisor — the one-shot second opinion — to
  claude-opus-4-8 or claude-sonnet-4-6 by exact id, and six pinned their
  main model the same way. Those ids still resolved, so nothing broke; but a
  plugin that says it is built for Claude Fable 5.1 while every advisor asks the
  previous generation is making a claim about the harness, not about itself.
  Advisors now name claude-opus-5 / claude-sonnet-5; main-model pins moved
  the same way; claude-haiku-4-5 stays, because Haiku 4.5 is the current Haiku.
  Same price as 4.8 at the Opus tier, cheaper at the Sonnet tier.
• The same move everywhere a model is chosen rather than priced: the
  frontier judge default, the eval runner's actor and judge, the product-owner's
  four-model debate panel, /review and /digest, the pm-planning default —
  and archetype-review-base/reviewer-template.md, which stamps every new
  reviewer; left alone, the next reviewer would have been born on 4.8. Price
  tables keep their 4.x rows: a cost meter must price what ran last month, not
  only what runs now.
• Agent postures, from 3.24.0, now reach the reference docs — the regenerated
  agent table carries the current models.

Full notes, including every fix and the reasoning behind it

Full notes for this release — what changed, and why each change was made.