Summary

Hi! This PR fixes an LCM summarizer edge case for reasoning-capable models.

If the initial summarizer request omits a reasoning setting, some provider compatibility layers can normalize that into a reasoning-disabled request. Some models do not support turning reasoning off, so the first request fails and only the retry succeeds.

I first noticed this with openrouter/minimax/minimax-m2.7, but the issue is more general than one specific model.

What changed

• add a small reasoningIfSupported option to the completion bridge
• keep explicit reasoning values unchanged and give them priority
• apply the fallback only when the resolved model advertises reasoning support
• update the LCM summarizer to send reasoningIfSupported: "low" on the initial request
• keep the existing explicit retry behavior unchanged
• add focused tests and a patch changeset

Why this approach

This keeps the fix narrow and model-aware.

It avoids a model-specific special case, preserves explicit caller intent, and does not force reasoning on models that do not support it.

Testing

• npx vitest run test/index-complete-model-auth.test.ts test/summarize.test.ts test/lcm-summarizer-reasoning.test.ts

Thanks very much for taking a look.