dsh-self-evolution：为 DeepSeek Harness 补上可边界约束的自进化记忆层
dsh-self-evolution 为 DeepSeek Harness 提供 Hermes 兼容的有界记忆、回合结束自动回顾、skill_learn/skill_doctor、可选跨会话 session_search，以及对话内学习行与输入框上方学习卡片。自动回顾默认仅追加记忆，删改需审批；写入带安全扫描与读回校验。安装：dsh plugin --profile <profile> add github:Tommy00748/dsh-self-evolution，并在 cordis.patch.yml 注册 bundle。插件以主机进程权限写文件，使用前请审阅源码。维护者 Tommy00748，MIT，0.9.0。
Read Moredsh-self-evolution: Benchmark-Driven Self-Evolution Plugin for DeepSeek Harness
`dsh-self-evolution` is an MIT-licensed plugin for DeepSeek Harness designed to safely iterate Agent Profiles on a frozen Benchmark. The plugin ensures the validity of improvements through an acceptance mechanism that strictly outperforms the current version; otherwise, it automatically performs a precise rollback from verified snapshots to prevent ties or performance degradation. Its core features include deterministic policy execution, isolation of Target/Evaluator/Optimizer roles, protection of private evaluation criteria, and multiple security safeguards (such as path traversal protection and mutexes). Each sub-agent possesses an independent session ID and supports log replay. Installation requires Node 22+ and DSH environment dependencies and is enabled via the `dsh plugin` command. Users can access `
Read More