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.

Yolo Mode — DSH Plugin for DeepSeek Harness
← Plugins

dsh-yolo-mode

Yolo Mode

dsh-yolo-mode — DeepSeek Harness dual-package plugin: when a session is in writable sandbox mode and the approval policy is ask, it uses a large language model to automatically adjudicate sandbox privilege escalation requests. Supports built-in presets and custom permission levels, and provides host

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

npx -y @deepseek-ai/dsh plugin --profile web add dsh-yolo-mode@0.5.0
READMECompatibilityVersions

Description

dsh-yolo-mode — DeepSeek Harness dual-package plugin: when a session is in writable sandbox mode and the approval policy is ask, it uses a large language model to automatically adjudicate sandbox privilege escalation requests. Supports built-in presets and custom permission levels, and provides host settings, a self-published settings bridge (/yolo-mode), and a web client UI.

Compatibility and provenance

Yolo Mode is published as dsh-yolo-mode and currently resolves to version 0.5.0. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
web
Release source
npm
Registry updated
9/20/2026

Versions

0.5.0stable
8/28/2026
0.4.1stable
8/20/2026
Show 1 more versionCollapse versions
0.5.0-beta.0beta
9/2/2026

Related plugins

Loading related plugins…

Latest
0.5.0
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
web
License
MIT
Source
npm
GitHub
★ 4
Weekly downloads
68
Last push
9/10/2026
View source ↗Project homepage ↗
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
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in

Related plugins

More verified plugins in security-access.

Doctor@linxin666/dsh-doctorTransactional rescue mode for DSH profiles with a supervised launcher, isolated recovery capsule, deterministic repairs, health monitoring, and a local Web recovery consolePocketdsh-pocketPut DeepSeek Harness in your pocket: one package, one settings page, and scan a QR code on your phone to access DSH on your computer in sync (LAN + public network, real-time screen mirroring).DSCODE@toddzheng024/dscode-bundleA complete DeepSeek coding agent with persistent shell, Ultra collaboration and automatic permission review.Auto Reviewdsh-auto-reviewSecond-model AI auto-review for DeepSeek Harness approval requests: a read-only reviewer subagent decides allow/deny on the approval answerer chain, with fail-closed fallback and full session-log audit.

README

dsh-yolo-mode

DeepSeek Harness 插件:用大模型自动裁决沙箱升权申请(LLM-powered auto-approval)。

当会话处于可写沙箱模式、审批策略为 ask 时,dsh-yolo-mode 拦截 escalate sandbox to ... 升权申请,按你选择的预设或自定义权限层级由大模型裁决「放行 / 拒绝 / 转人工」。任何不确定或失败路径都不会放行(fail-closed)。

兼容性:v0.5.0 支持 DSH 0.1.2-alpha.4+(真实-Cordis 探针 125/125 通过;settings 已迁移至 alpha.4 的 SettingsProvider 类服务)。旧版 DSH(0.1.0-rc.6 / rc.8)请使用最后兼容的 npm 版本 0.4.1。


功能特性

  • LLM 自动裁决:在 workspace-write 会话中,对 danger-full-access / workspace-write 升权申请调用大模型分级(allow / deny / unsure),确定时免去人工弹窗。
  • 内置预设:off / strict / balanced(默认)/ permissive / yolo / custom,覆盖从「全部转人工」到「确定性全放行」的完整光谱。
  • 自定义层级:levels 支持逐目标模式、逐工具(tools.<toolName>)的 allow / judge / delegate / deny 策略,及 error / unsure 回退。
  • 每预设默认提示词:judge.systemPrompt 留空时按预设自动选用对应裁判提示词(strict 最保守、permissive 宽松、custom 按层级表)。
  • 图形化配置:设置面板「YOLO 审批」页在线修改预设 / 生效模式 / 裁判模型(下拉选择)/ 层级表,保存即生效;输入栏状态 chip + 决策统计面板。
  • 完整审计:每次裁决落 JSONL 日志,含工具、目标模式、当前模式、理由、决策与结果;状态弹窗一键「打开日志」用系统默认应用查看。
  • fail-closed:超时、非法输出、模型不可用、并发溢出等一切异常路径都回退为「拒绝」或「转人工」。

安装

1. 安装插件包

# 从 npm
dsh plugin --profile web add dsh-yolo-mode

# 或从本地路径
dsh plugin --profile web add <项目绝对路径>

2.(可选)覆盖插件配置

dsh plugin add 会通过插件包自带的 cordis.patch.yml 自动挂载 yolo-mode 与 yolo-mode-bridge 两个入口,不需要再手动 insert。需要调整配置时,在 $DSH_HOME/profiles/web/cordis.patch.yml 里按 id 覆盖主条目即可:

- id: yolo-mode
  name: dsh-yolo-mode
  config:
    preset: balanced
    judge:
      provider: <provider>
      model: <model>

注意:不要再用 - insert: 添加 yolo-mode / yolo-mode-bridge,否则启动会报 duplicate loader entry id: yolo-mode。

重启 DSH 并刷新浏览器后生效。

安装注意事项

  • 推荐直接 dsh plugin --profile web add dsh-yolo-mode(npm 发布版已包含构建产物, peer 依赖由 DSH 的 $DSH_HOME/profiles/node_modules fallback 提供,开箱即用)。
  • 本地 link: 开发时,checkout 需位于 $DSH_HOME/profiles/ 下(或仓库自带 node_modules),否则 @deepseek-ai/* peer 依赖会因 Node 按真实路径解析而报 ERR_MODULE_NOT_FOUND。

配置

插件行 config 全字段可选,未填按默认值:

字段类型默认值说明
presetoff|strict|balanced|permissive|yolo|custombalanced使用的预设;custom 时以 levels 为准
modesstring[]['workspace-write']会话有效沙箱模式 ∈ 此列表时才介入(read-only / workspace-write / danger-full-access)
levelsobject{}权限层级表;levels.tools.<toolName> 对任意预设逐工具覆盖
judge.providerstring''裁判模型 provider;空 = 未配置(judge 决策按错误回退)
judge.modelstring''裁判模型;与 provider 同非空才启用裁判
judge.systemPromptstring''裁判 system prompt;空 = 按预设取默认
judge.timeoutMsnumber20000单次裁判超时(毫秒)
judge.maxTokensnumber256裁判输出最大 token 数
judge.concurrencynumber2并发裁判上限,溢出按错误回退
includeSubagentsbooleantrue子代理会话是否同样裁决
auditFilestring''审计日志路径;空 = %TEMP%/dsh-yolo/judge.log

权限层级(levels)

levels:
  workspace-write: judge        # 目标模式 → 策略
  danger-full-access: judge
  error: delegate               # 裁判错误回退
  unsure: delegate              # 裁判不确定回退
  tools:
    pwsh: delegate              # 逐工具覆盖,优先级最高
    write: allow

策略取值 allow | judge | delegate | deny;优先级:levels.tools[toolName] → 基础行(custom 时为 levels[targetMode],其余预设为内置表)。

预设

预设workspace-writedanger-full-access失败回退不确定回退说明
offdelegatedelegatedelegatedelegate不介入,全部转人工
strict 严格judgedelegaterejecteddelegate仅裁决 workspace-write;danger-full-access 恒转人工
balanced 均衡(默认)judgejudgedelegatedelegate裁决全部升级目标;失败 / 不确定转人工
permissive 宽松judgejudgedelegateallowed-once裁决全部目标;不确定视为允许(慎用)
yoloallowallow——确定性全放行,零 LLM 调用
custom依 levels依 levels依 levels.error依 levels.unsure全字段开放

每预设默认裁判提示词(judge.systemPrompt 留空时自动选用):

预设默认裁判立场
off / yolo不调用裁判
strict最保守:danger-full-access 一律拒绝;仅最小范围 workspace-write 且理由极充分才允许
balanced通用审计:只依据事实、防回环、存疑即 deny/unsure
permissive宽松:理由合理且范围可接受即倾向允许,仅明显破坏性 / 供应链风险拒绝
custom按 levels 层级表裁决,存疑按 levels.error / levels.unsure 回退

界面

  • 输入栏 chip:显示 YOLO <preset>,点击弹出统计面板(总审批 / 放行 / 拒绝 / 转人工 + 最近决策表,每页 5 条,带上一页 / 下一页翻页)。
  • 打开日志:弹窗右上「打开日志」按钮用系统默认应用打开审计 JSONL(auditFile 配置或默认 %TEMP%/dsh-yolo/judge.log);文件尚不存在时提示暂无记录。
  • 设置页:「YOLO 审批」页在线修改预设、生效沙箱模式、裁判模型(provider / model 下拉,取自 Harness 模型配置)、层级表(JSON);切换预设时自动预填充该预设的默认提示词与层级表。

安全

  • fail-closed:只有明确得到 allow 才返回一次性 allowed-once;其余一切路径拒绝或转人工。
  • 防回环:裁判 prompt 与 agent 上下文隔离,防止模型借 Web 审批回环自批准 danger-full-access。
  • 不改写策略:仅在 ask 策略下作为应答者,不改变 DSH 的沙箱 / 审批词汇。
  • 默认保守:默认预设 balanced(不确定转人工),不默认启用 permissive / yolo。
  • 审计:每次裁决落一行 JSONL,含 {time, sessionId, origin, toolName, callId?, targetMode, currentMode, justification, decision, outcome, reason?}。

开发

npm test        # node --test 全量测试
npm run build   # 构建客户端 bundle(rolldown)

许可

MIT