dsh-repeat-guard: Intercepting Degradation in Streaming Output and Enforcing Tool Calls

`dsh-repeat-guard` is a plugin for fixing "thinking degradation" in model reasoning. It addresses the problem where models such as `deepseek v4.1 flash` repeatedly loop short phrases or output fragmented repetitions (e.g., "OK. Do. Start. Do."), leading to excessive token consumption and context bloat. The plugin detects and blocks short phrases on the whitelist to force-abort output. Installation requires DSH 0.1.7 or later. After truncation, the system displays a message that forces the model to proceed directly to the next step, without emitting acknowledgments or pleasantries, thereby restoring the normal agent-loop. Users can configure the blocking threshold, whitelist, and continuation instruction in the settings.