Fresh session summaries default to OpenAI's gpt-6-luna

Fresh session summaries move from Google's gemini-2.5-flash to OpenAI's gpt-6-luna, the model Triage and most summaries already use. See Which LLM providers are called.

What's new?

• A session's first summary, when no Triage call reached it, and the one summary written after a session has been idle for 12 hours now go to OpenAI instead of Google.
• The default managed monitors no longer call Google. A self-hosted deployment needs GOOGLE_API_KEY only for graders on a google/ model: a monitor overridden in server-config.yaml, a custom grader, or the built-in Summarize and Policy Triage graders when called on demand.
• A monitors.summarize.model override in server-config.yaml still decides the model for fresh summaries.
• Watcher Cloud needs no action.

Required upgrade steps

These steps apply to self-hosted deployments only. Follow Update Watcher for the rollout itself.

1. Before upgrading, confirm that the OpenAI key or OpenAI-compatible gateway the Watcher API uses can call gpt-6-luna. Deployments that moved Triage to another provider in server-config.yaml may not have that access. If yours cannot call it, set monitors.summarize.model in server-config.yaml to a model it can call, such as google/gemini-2.5-flash to keep the previous model. A deployment that already sets monitors.summarize.model needs no change.
2. After upgrading, check the operation worker (xylon-operation-worker) logs for Operation handler failed warnings with kind summary. Seeing summaries in the Analyzer is not enough: most summaries continue a Triage call and do not exercise the fresh-summary path.