Repository navigation
Releases
v6.2.8 Project isolation and verified skill trials
v6.2.8 修复跨项目技能候选合并、宽松成功归因与 Agent/Session 标识关联问题。
- 候选聚类和 ID 纳入项目、任务、根因及策略身份；不同项目和不同策略分别积累证据。
- 取消“工具无错误即技能有效”的自动反馈。learning_skill_feedback 改为 begin/complete 试验流程，由维护者本地契约固定策略调用、后续独立检查及结构化结果断言；真实 session 事件与验证凭据共同决定是否计数。
- 注入、工具观察与反馈统一按 Session ID 关联；正确读取原生工具结果中的 callId 和嵌套错误标记。已激活技能按项目过滤后再应用数量和 Token 预算。
- 旧版候选先备份到 skill_candidates.pre-evidence-v2.json，再按项目重建；旧版宽松成功计数不继承到新版技能。插件来源的上下文不会被当成用户偏好学习。
验证：本地 25 项测试通过；公开 GitHub Actions 在 Windows/Linux 的 Node.js 24 环境均通过 25 项测试、npm audit --omit=dev（0 漏洞）和发行包检查。
CI：https://github.com/326sun/runtime-self-learning-harness/actions/runs/41048515705
升级注意：技能反馈接口已收紧。没有维护者配置的验证契约时，插件继续观察、记忆和生成候选，但普通成功回合不会触发技能晋升。配置与调用示例：
https://github.com/326sun/runtime-self-learning-harness/blob/v6.2.8/docs/SKILL-VERIFICATION.md
本次属于集成验证补丁；尚未完成真实 Harness 宿主安装、多轮模型交互与项目切换的端到端验收。验证可靠性也取决于维护者提供的独立检查本身，不能据此宣称技能在所有任务中都有因果收益。
附件包含相同包内容的 TGZ、ZIP 与 SHA-256 校验文件。原生 bundle 安装仍可使用 dsh plugin --profile web add github:326sun/runtime-self-learning-harness。
v6.2.7 Bind skill evidence and inject active skills
Fixes candidate-bound verified skill feedback, active skill injection through agent/pre-step, and documentation for the package.json dsh.bundle.patch plus cordis.patch.yml contract. Validation: 7 tests passed; npm audit reports 0 vulnerabilities.
v6.2.6 Complete Harness governance and skill feedback
Fixes unified scoped retrieval with token and governance gates, disables the auto-action pipeline for Harness, uses agent/created, preserves pre-step decision fields, records success and regression feedback for skill candidates, and removes active skills after regression. Validation: 7 tests passed; npm audit reports 0 vulnerabilities.
v6.2.5 Fix Harness turn lifecycle and project scope
Fixes native DSH turn-end handling, tool failure mapping, global session observation, automatic skill evolution, per-agent dataDir/project scope, and lifecycle cleanup. Validation: 7 tests passed; npm audit reports 0 vulnerabilities.
v6.2.4 Automatic skill evolution feedback loop
Native DSH self-learning loop with automatic observation, scoped pre-step memory injection, repeated failure reflexion records, governed skill evolution candidates, 17 native tools, and 6 passing full-flow tests. npm audit reports 0 vulnerabilities.
v6.2.3 Complete native self-learning loop
Native DSH self-learning loop: automatic session observation, scoped agent pre-step memory injection, governed skill evolution, 17 native tools, and 5 full-flow tests. npm audit reports 0 vulnerabilities.