DeepSeek Harness Plugin Hub

Publish and manage complete Harness Profiles. Discover Plugins for your next setup.

Explore

PluginsPresetsDocsNews

Community

Publish a pluginContactReport an issue

Resources

Plugin Hub on GitHubDeepSeek HarnessSystem statusPrivacy notice
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

Independent and unofficial. Not affiliated with, authorized by, or endorsed by DeepSeek.

Skill Evolution — DSH Plugin for DeepSeek Harness
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in
← Plugins
S

dsh-skill-evolution

Skill Evolution

DSH web plugin: Hermes-style skill self-evolution. Watches agent execution traces, crystallizes reusable skills from successful turns via LLM review, progressively improves them, and manages everything in a dedicated Settings section.

The plugin will be installed here. Keep web if you are unsure.

npx -y @deepseek-ai/dsh plugin --profile web add github:VanadisGithub/dsh-skill-evolution#40f7bcafe3418ee9a2d6c391e28e6319d465b0fa
READMECompatibilityVersions
Settings tab

Compatibility and provenance

Skill Evolution is published as dsh-skill-evolution and currently resolves to version 1.1.0. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
web
Release source
github
Registry updated
9/1/2026

Versions

1.1.0stable
9/1/2026

Related plugins

Loading related plugins…

Latest
1.1.0
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
web
License
MIT
Source
github
GitHub
★ 0
Weekly downloads
0
Last push
9/7/2026
View source ↗
README badge

Click the badge to copy Markdown for your README.

Do you maintain this Plugin?Claim benefit · Priority security scan

Verify the GitHub repository declared in package.json to manage this listing. After you claim it, Hub will prioritize a security scan of the current version and publish the result when it passes.

Claim this Plugin →
Report an issue

Related plugins

More verified plugins in agents-orchestration.

Headless@deepseek-ai/dsh-headlessThe dsh one-shot bundle: a direct core Agent/Session runner over dsh-base with no Host, HTTP, or browser layerExperimental Agent Team Web Profile@deepseek-ai/dsh-experimental-agent-team-web-profileExperimental Web profile layer for Agent Teams Remote and UI pluginsSubagent Codex@deepseek-ai/dsh-subagent-codexOne-shot Codex subagent provider over the official app-server protocolSubagent Claude Code@deepseek-ai/dsh-subagent-claude-codeOne-shot Claude Code subagent provider over the official Agent SDK

README

dsh-skill-evolution

English | 中文

DeepSeek Harness(DSH)的 Hermes 式技能自进化插件:观察 agent 执行轨迹,在每个成功回合结束时按信号触发 LLM 评审,把值得复用的工作流结晶为技能并注册进技能目录;后续同类执行还会把新经验合并进已有技能。插件自带设置页管理面板,技能持久化、重启不丢。

结晶模型参考 Nous Research Hermes Agent 的自进化机制(事件驱动评审,而非纯频率统计)。

工作原理

事件驱动:每个回合结束(turn/end)时,对满足最小步数的成功轨迹检测三个信号,任一信号触发即送 LLM 评审:

信号条件对应 Hermes 触发条件
complex成功回合且工具调用 ≥ minToolCalls(默认 5)"completed a complex task (5+ tool calls) successfully"
recovered有失败步骤但回合最终成功(找到了绕行路径)"hit errors or dead ends and found the working path"
repeated同一工具序列出现 ≥ minPatternOccurrences 次且成功率 ≥ minSuccessRate重复轨迹学习(GEPA 式)

评审门禁(Hermes background-review 模型):评审 LLM 收到完整证据(信号、用户意图、工具序列、每步成败与摘要),决定 SAVE 或 NOTHING_TO_SAVE。硬性负面清单——以下内容绝不结晶:

  • 环境性失败(缺二进制、未配置凭证、未装包)
  • 对工具的负面断言("X 工具坏了"会固化成长期拒绝)
  • 重试即愈的瞬时错误(教训是重试模式,不是原错误)
  • 一次性任务叙事、未解决的失败

渐进改进(技能是活的):当新证据落在已有进化技能的序列上(尤其是错误恢复经验),评审会把新教训合并进已有技能内容并版本 +1,而不是创建重复技能。

产出格式:类级命名的 kebab-case 技能,标准 SKILL.md 结构(# Title / ## When to Use / ## Procedure / ## Pitfalls / ## Verification)。

组成

双面插件包(host + web client):

文件角色
plugin.mjs宿主半:轨迹观察、信号检测、LLM 评审、技能注册、./generated/ 持久化、/skill-evolution 认证 RPC 通道
client.js浏览器半:设置页「技能进化」分区(技能卡片、信号展示、两步确认删除/清空、实时模式统计)
cordis.patch.ymldsh.bundle.patch 挂载层

仅依赖 node: 内建模块,不引入任何 @deepseek-ai 包,避免与宿主 cordis 双实例冲突。

安装

方式一:官方 CLI(发布到 npm 后)

dsh plugin --profile web add dsh-skill-evolution

包内 dsh.bundle.patch 会自动挂载,无需手改 profile 文件。

方式二:git clone + link

git clone https://github.com/VanadisGithub/dsh-skill-evolution.git ~/Code/dsh-skill-evolution

在 ~/.dsh/profiles/web/package.json 的 dependencies 中加入:

"dsh-skill-evolution": "link:/Users/<you>/Code/dsh-skill-evolution"

然后在 ~/.dsh/profiles/web 执行 pnpm install,硬刷新浏览器。

方式三:file:// 手动挂载(原型调试)

把仓库放到任意目录(如 ~/.dsh/plugins/skill-evolution/),在 ~/.dsh/cordis.patch.yml 追加:

- insert:
  - id: 'skill-evolution'
    name: file:///Users/<you>/.dsh/plugins/skill-evolution/plugin.mjs?v=1
    config:
      minToolCalls: 5
      minPatternOccurrences: 3
      autoRegister: true
      llmProvider: deepseek      # 按你的部署可用的 provider 修改
      llmModel: deepseek-chat

?v=N 用于绕过 Node 模块缓存;改了 plugin.mjs 就 +1。宿主半改动需 loader 重放补丁,浏览器半改动硬刷新即可。

⚠️ 三种方式只能选一种,重复挂载会导致双实例。

配置

键默认说明
minTraceSteps3轨迹最小步数(噪音地板)
minToolCalls5complex 信号阈值
minPatternOccurrences3repeated 信号阈值
minSuccessRate0.7repeated 信号成功率下限
signals全开{complex, recovered, repeated} 分别开关
autoRegistertrue结晶后自动注册进技能目录
llmProvider / llmModeldeepseek / deepseek-chat评审用的模型(必须是部署中存在的 provider)
maxEvolvedSkills20进化技能容量上限
improvementEvery3每 N 次同序列重跑触发一次改进评审
maxImprovementsPerProcess5每进程每技能改进次数上限
maxEvidenceSteps30送评审的最大步骤数

管理面板

设置 → 技能进化:

  • 配置与信号开关一览、实时模式统计(各序列出现次数/成功率/是否结晶)
  • 技能卡片:结晶信号、工具序列、版本与改进次数;展开查看完整 SKILL.md
  • 单个删除 / 清空全部(均为两步确认);删除会同时反注册并删除磁盘文件

截图 / Screenshots

设置页「技能进化」总览——信号配置、实时模式统计与已结晶技能卡片:

Settings section overview

展开技能卡片——结晶信号、工具序列与完整 SKILL.md 内容:

Expanded skill card with crystallization signals

宿主侧另暴露认证 RPC 通道 /skill-evolution:list / get / remove / clear。

数据与隐私

  • 进化技能持久化在插件目录 generated/<name>.json(原子写入),服务重启自动恢复注册
  • 轨迹统计仅在本进程内存,重启重新累计
  • 评审调用走 DSH 自身的 LLM 通道;证据只包含工具名、截断后的参数/结果摘要(默认 ≤200 字符)与首条用户消息

English

Hermes-style skill self-evolution plugin for DeepSeek Harness (DSH). It watches agent execution traces and, at the end of every successful turn, fires an LLM review when any signal trips — a complex turn (≥5 tool calls), a recovered turn (failed steps but a working path found), or a repeated tool sequence. The reviewer either crystallizes a class-level reusable skill (registered into the skill catalog and persisted under generated/) or answers NOTHING_TO_SAVE, following Hermes' do-not-capture rules (environment failures, negative tool claims, transient errors, one-off narratives, unresolved failures). Existing evolved skills are progressively improved with fresh evidence instead of duplicated. Ships a Settings-section management UI (view / expand / delete / clear) backed by an authenticated /skill-evolution RPC channel. Host half uses node: builtins only.

Install: dsh plugin --profile web add dsh-skill-evolution (npm), or link: / file:// mounts — see the Chinese section above for details.

License

MIT