Lhy723/dsh-self-evolution 预览 preview

Lhy723/dsh-self-evolution

Plugin插件 Native原生 ⭐ 3 MIT Dev Workflow开发与代码工作流

基于基准的DeepSeek Harness自我进化 · 在冻结基准上的Agent Profile自我进化:评测 → 候选 → 严格接受/回滚

Project Overview项目介绍

dsh-self-evolution is a native plugin built exclusively for DeepSeek Harness (DSH) and the Cordis framework. It enables benchmark-driven self-improvement for DSH agent profiles, including AGENTS.md, custom skills, runtime configuration, and general profile settings. To install the plugin, you can run a single DSH CLI command: either install from the NPM registry, directly from the GitHub repository, or from a local tarball that you have downloaded or built yourself. After installation, you can verify it by dumping your DSH profile config to check that the plugin bundle is correctly injected.

When you run the plugin, it follows a strict three-step workflow: first run the target agent against the frozen benchmark, then have a private evaluator agent score the output, then have an isolated optimizer agent generate candidate changes. Only changes that produce a strictly better score than the current baseline are accepted; all other changes are rolled back to the last verified snapshot. It is designed for DSH developers who want to iteratively and reliably improve their agent profiles over time without manual intervention.

This plugin is released under the permissive MIT open source license, and requires Node.js 22 or newer to build and run. It requires an initialized DSH profile, pnpm available in your system PATH, and the official default spawn provider to work correctly. The plugin only provides context and file transaction boundaries, not an OS-level sandbox, and it does not support distributed locking for cross-host concurrent runs. All core functionality is type-checked against DSH rc.6 but end-to-end live testing with model credentials is not yet complete.

这是一款专为 DeepSeek Harness / Cordis 开发的原生插件,用于实现 Agent 配置文件的基准驱动自我进化。插件会让 AGENTS.md、技能集、runtime.json、配置文件在已冻结的基准测试集上迭代,只有改进后性能严格优于当前版本的变更才会被接受,否则会精确回滚到已验证的快照版本。它原生集成 DSH 接口,是真正的 Cordis 服务,并非 shell 脚本包装。

整个工作流分为目标子代理运行、私有评估子代理打分、隔离优化子代理生成候选变更三个核心环节,全程严格保证隐私边界,私有评测标准不会暴露给目标或优化子代理,所有变更全程可追踪,每个子代理都有独立可复盘的会话 ID。它适合需要持续优化 Agent 配置和能力的开发者使用。

本插件遵循 MIT 许可开源,要求 Node.js 版本不低于 22,需要已初始化的 DSH profile、pnpm 环境和官方默认的 spawn 提供者才能运行。它仅提供上下文和文件事务边界,不提供操作系统级沙箱,跨主机并发运行不支持分布式锁,当前端到端带模型凭据的验证还未完成。

Pre-install check安装前体检Compatibility · Security兼容性 · 安全性 1 warning1 项注意
  • Only 3 stars - very few users, little community feedback星标只有 3,几乎没人在用,遇到问题缺少社区反馈
DSH walks through these 9 checksDSH 会逐条核对这 9 项

Compatibility兼容性

  • DSH, Node, OS and profile requirementsDSH 版本 / Node 版本 / 操作系统 / profile 是否满足要求
  • External dependencies and runtimes (Electron / Python / Docker, ...)外部依赖与运行时(Electron / Python / Docker 等)是否齐备
  • Conflicts with installed plugins: command names, skill / tool names, ports, duplicate MCP registration与已装插件是否冲突:命令名、skill / tool 重名、端口占用、重复 MCP 注册

Security安全性

  • Repo matches the facts registered here; archived or abandoned?仓库是否与页面登记一致,是否归档或长期停更
  • Safety of preinstall / install / postinstall and install.sh / setup.ps1preinstall / install / postinstall 与 install.sh、setup.ps1 是否安全
  • curl|bash, download-then-execute, obfuscation, unrelated domains → stop immediatelycurl|bash、下载即执行、混淆代码、无关域名 → 立刻停止
  • Typosquatting or unmaintained packages among the new dependencies新增依赖里有没有 typosquatting 或无人维护的包
  • Requested permissions vs. what the feature actually needs申请了哪些权限、是否超出功能所需(filesystem / network / shell / clipboard)
  • Any sudo / admin requirement, plus uninstall and rollback是否要求 sudo / 管理员权限,以及卸载与回滚方式

Anything uncertain must be marked unknown with a note on how to confirm it. This site's signal screen is a static snapshot, not a security audit.拿不准的必须标「未知」并说明要我怎么确认。本站的信号筛查是静态快照,不能替代安全审计。

Or use CLI install (for developers)或使用命令行安装(适合开发者)

CLI Install命令行安装

dsh plugin --profile web add dsh-self-evolution

把 Lhy723/dsh-self-evolution 加入你的 DSH 配置(web profile)即可启用。

READMEREADME

dsh-self-evolution 中文社交预览图

English · 简体中文

dsh-self-evolution

License Node DeepSeek Harness

面向 DeepSeek Harness / Cordis 的基准驱动自我进化插件:让 Agent Profile(AGENTS.md、Skills、runtime.json、config)在冻结的 Benchmark 上自我迭代——每次改进必须严格优于当前版本才会被接受,否则从已验证的快照精确回滚。

Agent Profile v1
  ↓ frozen Benchmark
Target Subagents × Case × Run
  ↓
Private Evaluator Subagents
  ↓ score + public feedback + DeepSeek Session IDs
Isolated Optimizer Subagent
  ↓ bounded file operations
AGENTS.md / Skills / runtime.json / config
  ↓ Agent Profile v2
re-evaluate
  ├─ score strictly better → snapshot + accept
  └─ otherwise             → verified rollback

特性亮点

  • 原生集成:直接调用 ctx.subagents.start(...) 与 ctx.tools.register(defineTool(...)),是真正的 Cordis Service(ctx.evolution),不是 shell 包装脚本。
  • 确定性宿主策略:评测决策不进 LLM——Benchmark 冻结哈希、Case × Run 矩阵、严格接受门槛(> 而非 >=,平分回滚)、单调版本号、快照验证全部由宿主代码执行。
  • 三类隔离角色:Target 继承 Profile 的 persona 与工具范围;Evaluator 只见私有 rubric + 原始输出;Optimizer 只见公开证据,并带 { allow: [] } 全工具隔离,只能通过受限文件操作提候选。
  • 隐私边界:私有 rubric 不进 Target / Optimizer 上下文;公开与私有产物分目录落盘。
  • 事务化变更:候选文件白名单、路径穿越与符号链接防护、单 Profile 互斥锁、外部改动检测(漂移时拒绝覆盖回滚)。
  • 全程可追踪:每个子 Agent 都有独立 DeepSeek Session ID,Scoreboard 与运行事件日志可复盘每一轮的决策。
  • 即插即用:声明 dsh.bundle 的 profile bundle,dsh plugin 一条命令完成安装与组合。

快速开始

前置条件

  • 一个已初始化的 DeepSeek Harness profile(dsh CLI 与 pnpm 在 PATH 上);
  • profile 自带 tools / subagents 注册表与官方 spawn provider(默认 @deepseek-ai/dsh-base 即满足);
  • 一个 Agent Profile 目录(含 AGENTS.md 与 runtime.json);
  • 一个冻结的 Benchmark 目录(放在 Target workspace 之外)。

安装

本包是 DeepSeek Harness profile bundle(package.json 的 "dsh" 段指向根目录 cordis.patch.yml),安装、激活一条命令完成:

Showing the opening section of the README — the full document lives in the repository以上为 README 开头摘要,完整文档在仓库内 · View the full README on GitHub →在 GitHub 查看完整 README →

← 上一个 Prev dsh-audiogen 下一个 Next dsh-tavily →