Refinement always fails with 400 on models that can't disable thinking

[pcarielo]

Hey! Hit a refinement bug on 0.9.3 and want to report it the right way this time (I opened issue #2080 first, then read CONTRIBUTING, so feel free to close that one, this is the real intake).

What happens:

Every refinement fails with my current model because the refinement call forces non-reasoning mode and the model can't disable thinking.

Error: Refinement failed: 400 This model always engages in thinking and cannot be disabled; please use low, high, or max

Where it comes from:

In dist/core/refinement/refinement.js, planRefinement and reviewAutoRefine both discard the session thinkingLevel on purpose (void thinkingLevel;) and call completeSimple without reasoning. The comment explains it keeps the output budget for the JSON object, which makes sense, but it hard fails with models that always think.

Impact:

All refinement paths break with these models: refine.run(), /refine, and the auto refine review. So harness refinement is basically unusable there.

Suggested fix:

Instead of forcing reasoning off, pass the lowest supported thinking level for the model (THINKING_LEVELS is already around) or catch the 400 and retry with reasoning enabled. Either way keeps the JSON goal without hard failing on always-thinking models.

Happy to test a fix or provide more details. Thanks!

[kevinjosethomas]

I’m taking a look at this.

[kevinjosethomas]

I’m looking into this in #2130.

[kevinjosethomas]

Merged in #2130; this fix is not released yet.