dsh-run-guard

When using DeepSeek V4 Flash for long-running tasks, Agents often encounter the dual challenges of reasoning loops and premature stops: models can get stuck in endless, meaningless reasoning cycles lasting minutes, or complete their thinking without taking action, leaving tasks unfinished. The dsh-run-guard plugin is designed to address these issues. It employs a guard mechanism that monitors reasoning streams in real-time, using sliding window repetition detection and hard limits to quickly interrupt loops with automatic retries. Simultaneously, the continue mechanism injects prompts to resume tasks when unfinished or when the model stops early, preventing premature halts. Additionally, the recovery feature handles upstream request failures, and pause_work allows models to manually pause for flexible control. This plugin is ideal for all developers building Agents with DeepSeek Harness, especially for complex, long-duration task scenarios. By ensuring a stable runtime rhythm, it helps Agents complete tasks reliably without getting stuck or stopping prematurely, thus enhancing task efficiency and robustness.