Allow using a different model for /refine

[callmehanu]

Area

AI providers and models, TUI

Problem

I noticed compaction already supports using a cheaper/faster model via the
session_before_compact extension hook (the custom-compaction.ts example is great). It would be
nice to have the same for /refine.

Proposed direction

Right now planRefinement and reviewAutoRefine always use this.model, so if you're on an
expensive reasoning model you're paying premium rates for what's essentially "read the
conversation and output some JSON edits." A deepseek-v4-flash or similar could handle that just
fine at a fraction of the cost.

Two ways this could work:

• A session_before_refine hook (same pattern as session_before_compact)
• Or simpler: a refineModel setting in settings.json (no extension needed)

Either would be great. I know refine_complete exists but that's after the fact — by then the
tokens are already spent.

Alternatives considered

No response

Additional context

No response