Why can we no longer disable the "thinking" of models from OpenRouter since version 2.0? #19729

@Archibaldr Thanks for reporting this. I checked it against v1.9.13 and the current main branch: this is a regression bug, not an intentional removal.

In v2.0, reasoning controls moved from the old UI heuristics to per-model data from the provider registry. OpenRouter describes two separate facts: supported_efforts lists the levels available while reasoning is enabled, and mandatory says whether reasoning can be disabled. For example, deepseek/deepseek-v4-flash currently reports mandatory: false but lists only xhigh/high. Cherry currently derives the whole selector from supported_efforts, so Off disappears and an old Off selection is omitted from the request, leaving OpenRouter to use the model default.

The fix plan is to:

• derive the Off/toggle capability from mandatory != true, independently of supported_efforts;
• serialize Off through the OpenRouter disable form while keeping mandatory models unchanged;
• regenerate the provider catalog and add regression coverage for optional versus mandatory OpenRouter models plus the final request payload.

This also explains why the regression affects many OpenRouter models rather than only DeepSeek V4 Flash. Thanks again for the clear report.