dsh-custom-provider-reasoning

In DeepSeek Harness, users of custom model providers often encounter a limitation: the composer model selector only displays reasoning effort options for models with pre-defined metadata, leaving custom routes without this crucial control. This gap prevents users from fine-tuning the inference behavior of powerful models like DeepSeek V4, even though they can freely add custom providers.

This plugin bridges that gap by automatically injecting reasoningEfforts metadata into each eligible model, enabling options like Off, Low, Medium, and High in the model chooser. It ensures these selections are effectively transmitted to the backend. Additionally, it intelligently fills in missing maxTokens and contextWindow parameters to avoid output truncation and manages protocol-specific compatibility settings, such as thinkingFormat and supportsDeveloperRole, to prevent errors from unsupported roles or format mismatches across different API protocols.

Designed for developers and researchers using custom model providers, especially those needing to adjust inference depth for performance or cost optimization, this plugin works seamlessly after installation. It empowers custom routes with the same reasoning effort control as built-in providers, streamlining development workflows and enhancing the usability of advanced models in the DeepSeek Harness ecosystem.