dsh-llm-governor

When building complex AI services, simple model calls often fall short in multi-user environments where cost control and permission management are critical. As teams grow, resource allocation can easily become unmanageable without a unified governance layer. dsh-llm-governor is designed to address this pain point. As a native extension for DeepSeek Harness, it injects powerful governance capabilities into the model invocation chain without proxying actual requests.

The plugin's core strength lies in its flexible smart routing and strict access control. It supports various routing strategies, including manual selection, Quality First based on model performance, and Credit First based on quotas, and can automatically handle fallback rerouting to ensure service stability. By assigning monthly Credits and fine-grained model profiles (Lite/Balanced/Pro) to each user, administrators can precisely control costs and performance. All decisions are recorded in the Governor Trajectory, providing a traceable audit log.

Whether you are a developer needing local configuration for a single user, a small team connecting via Header mode, or an enterprise integrating enterprise-level IdPs using JWT, this plugin offers flexible adaptation. It focuses on the policy layer and does not interfere with credential management, keeping DSH focused on inference itself. For enterprises looking to build scalable, secure, and easily auditable AI services on top of their existing architecture, this is an indispensable governance tool.