use a non-reasoning model for session recaps

[kevinjosethomas]

• session recaps often showed a dangling "...text" because the nemotron model used for them is a reasoning model that streamed chain-of-thought as its answer and frequently ran out of tokens before closing the recap tag.
• switched the background summarizer to the qwen3-30b instruct (non-thinking) variant, which is self-hosted on prime inference and closes the tags reliably.
• validated against 30 real local sessions: the old model produced a usable recap 18 of 30 times, the new one 30 of 30 with far fewer tokens.

---

Note

Low Risk
Limited to background dashboard recap text and model registry metadata; no auth, session, or user-facing agent path changes.

Overview
Daemon session recap generation now calls qwen/qwen3-30b-a3b-instruct-2507 on prime-inference instead of nvidia/nemotron-3-nano-30b-a3b, so background summaries use a non-reasoning instruct model that can close the required <recap> / <status> tags instead of burning tokens on chain-of-thought.

The new model is added to the prime-inference curated catalog in generate-models.ts and the generated models.generated.ts entry (262k context, 65k max output, reasoning: false).

^{Reviewed by Cursor Bugbot for commit fa986ab. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Switch daemon session summarizer to non-reasoning model qwen/qwen3-30b-a3b-instruct-2507

Updates SUMMARY_MODEL_ID in daemon-session-summarizer.ts from nvidia/nemotron-3-nano-30b-a3b to qwen/qwen3-30b-a3b-instruct-2507, a non-reasoning model. The new model is registered in the generated models list under the prime-inference provider with a 262k context window.

^{Macroscope summarized fa986ab.}

[linear]

ENG-4365 Recap often simply outputs " ...text"

Seems like the recap might be provided with missing context or something of the sort?

Review in Linear