🦋 New version release

[brentrager]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@smooai/smooth-operator@1.59.0

Minor Changes

• df9778c: SMOODEV-3342: every server now defaults its main turn to gpt-6-luna and its conversation-workflow judge to groq-gpt-oss-120b.

  Behaviour change — defaults only. A deployment that sets SMOOTH_AGENT_MODEL / SMOOTH_AGENT_JUDGE_MODEL (or an org / agent / per-turn model) is unaffected. An unconfigured server used to request claude-haiku-4-5 for both; on the Smoo AI gateway those calls were failing over to whatever LiteLLM's fallback chain picked. The Smoo AI gateway's policy is now that every default LLM call uses the gpt-6-luna family or a Groq alias.

  • Main-turn default claude-haiku-4-5 → gpt-6-luna in the Rust server + Lambda + dev-support example, and the Go, Python, .NET and TypeScript servers (plus the Helm values, Argo CD example, SST example and examples/.env.example, which shipped gemini-2.5-flash). The Go and Python servers now pass the default to the engine explicitly instead of leaving the model empty, so the engine's own built-in fallback model is never what goes on the wire.
  • The judge has its own default, groq-gpt-oss-120b (DEFAULT_JUDGE_MODEL in Rust, DefaultJudgeModel in Go, WORKFLOW_JUDGE_MODEL in Python, ServerEnv.DefaultJudgeModel in .NET, DEFAULT_JUDGE_MODEL in TS). It no longer follows the main-turn model. SMOOTH_AGENT_JUDGE_MODEL still overrides it (the .NET host now also reads that canonical name, keeping SMOOTH_JUDGE_MODEL as an alias).
  • Judge output cap raised to 512 tokens in every port (Rust was 16, the others 200). The new default judge is a reasoning model, and reasoning tokens count against max_tokens: a small cap was spent entirely on reasoning, the reply came back empty, and the workflow never advanced. The prompt still asks for a one-word JSON verdict.
  • Evals: CHEAP_MODEL (the default agent + judge model) is now gpt-6-luna; the nightly matrix grades groq-gpt-oss-120b and gpt-6-luna, both judged by gpt-6-luna. Live E2E tests run on gpt-6-luna.
  • GET /admin/settings for an org that never saved settings reports gpt-6-luna in the Go, Python, .NET and TS servers.

@smooai/smooth-operator-web-chat-example@0.0.124

Patch Changes

• Updated dependencies [df9778c]
  • @smooai/smooth-operator@1.59.0