deepseek-harness-retry

When a DeepSeek Harness model request fails due to a transient issue, overload, or interruption, blindly retrying can become unsafe. deepseek-harness-retry addresses this by joining the Host request-error waterfall, applying bounded exponential backoff after built-in recovery policies, and supporting provider-neutral error codes, exact route allowlists and denylists, Retry-After awareness, and abort-aware waits.

Beyond ordinary retries, it handles interrupted-session recovery after a DSH restart. If a retry did not finish, a model step lacks an assistant message, or a durable tool boundary is closed before the next model step, the plugin can continue only provably unfinished work. It never removes, reorders, or replaces an existing durable inbox item, and it keeps retry evidence in session events rather than model-visible history.

This plugin suits developers running DeepSeek Harness across provider routes and wanting fewer lost failed requests with clear diagnostics. It extends built-in recovery without requiring browser steps or plugin-specific environment variables, making it a practical resilience layer for long-lived Harness deployments.