dsh-model-failover

In deep learning model inference, unstable models or service providers can lead to request failures and disruptions. The dsh-model-failover plugin is designed for DeepSeek Harness to handle faults intelligently with a two-level circuit breaker mechanism: when a model repeatedly fails, it automatically opens a circuit and routes subsequent requests to a configured fallback model; if multiple models under a provider fail simultaneously, a platform-level circuit breaker takes over, switching the entire service. This ensures continuity and reliability in inference workflows without requiring core code modifications. The plugin also supports recovery probes and integrates with existing retry strategies, making it ideal for developers who need high-availability model services, such as teams building production-grade applications or researchers aiming to reduce downtime from model failures and improve user experience.