dsh-auto-evolve

Traditional DeepSeek Harness agent workflows often rely on static skill sets, which lack the flexibility needed for complex, evolving tasks. dsh-auto-evolve introduces a revolutionary concept: empowering agents to not only execute tasks but to actively evolve. By closely observing runtime signals—such as tool failures and repetitive calls—the plugin proactively identifies optimization opportunities.

The core of this plugin lies in its rigorous self-evolving closed-loop mechanism. It maintains a "genome" of evolvable assets, including skills, tool-result post-processors, and prompt sections. When anomalies exceed specific thresholds, the system leverages an LLM to propose concrete mutations. Crucially, every proposal is validated within a sandboxed environment; changes are only applied to the live system if the trial demonstrates measurable improvements over the baseline.

This is invaluable for developers and advanced users looking to build long-term maintenance and high-reliability automated systems. By leveraging this mechanism, agents can automatically fix bugs, optimize prompts, or improve tool-call strategies during runtime, achieving a true "learning while using" capability. Paired with robust version control and first-class rollback features, users can rest assured that the evolution process won't break existing functionality, offering a safe and efficient intelligent operational experience.