Derive responses-only routing from the model registry's targetFormat instead of hardcoding model checks, and strip thinking suffixes when looking up models in providerModels so variants like gpt-5.6-luna(high) are routed correctly to /responses.
5.9 KiB
5.9 KiB