SeverusZh/dsh-yolo-mode
30Last commit Aug 14, 2026
dsh-yolo-mode DSH plugin
dsh-yolo-mode intercepts sandbox escalation requests (e.g., escalate sandbox to) in writable workspace sessions and uses a large language model to decide whether to allow, deny, or delegate to a human. It supports multiple presets (off, strict, balanced, permissive, yolo, custom) and custom permission levels, with a fail-closed design that refuses or delegates on any error or uncertainty.
How to install the dsh-yolo-mode DSH plugin
dsh plugin --profile web add dsh-yolo-modeCopying does not run this command. Review the repository and version before installing the dsh-yolo-mode DSH plugin.
dsh-yolo-mode DSH plugin data source
dsh-yolo-mode DSH plugin snapshot date: Aug 16, 2026
discovered
What the dsh-yolo-mode DSH plugin can do
- LLM auto-approval: Uses a configurable LLM to judge escalation requests based on preset or custom policies.
- Multiple presets: Six built-in presets (off, strict, balanced, permissive, yolo, custom) covering the full spectrum from all-delegate to fully-automatic.
- Custom permission levels: Define per-target-mode and per-tool strategies (allow, judge, delegate, deny) with fallback for error and unsure outcomes.
- Graphical configuration: An online settings panel to change preset, model, and levels, with live preview and statistics.
- Audit logging: Each decision is logged as JSONL with full context (tool, target mode, reason, decision, outcome).
- Fail-closed: Any exception (timeout, invalid output, model unavailability, concurrency overflow) results in deny or delegate, never allow.
Where the dsh-yolo-mode DSH plugin fits
- Automate approval of low-risk workspace-write escalations in development environments.
- Strictly control dangerous-full-access escalations, requiring human review or LLM judgement only when conditions are safe.
- Enable rapid iteration with permissive presets for trusted sandbox sessions, while logging all decisions for audit.
- Customize per-tool escalation policies (e.g., always allow write, always delegate pwsh).
Who the dsh-yolo-mode DSH plugin is for
- DSH users who want to reduce manual interruption for sandbox approval requests.
- Teams that need a configurable, auditable, and safe auto-approval mechanism for sandbox privilege escalation.
dsh-yolo-mode DSH plugin limitations
- Requires an LLM provider and model to be configured in DSH and selected in the plugin config; without them, LLM judgement is disabled and defaults to delegate.
- Only intercepts escalation requests in sessions with sandbox mode matching the configured modes (default: workspace-write).
- The plugin is still in early development; the README does not guarantee stability or production readiness.
- The default preset (balanced) delegates on uncertainty, which may still require human intervention.
dsh-yolo-mode DSH plugin: from the repository README
Quoted from the SeverusZh/dsh-yolo-mode README, the upstream source of the dsh-yolo-mode DSH plugin. Copyright remains with the original authors.
> 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**）。 [](https://www.npmjs.com/package/dsh-yolo-mode) [](LICENSE) --- ## 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。 - **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 ```bash # 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. 追
Read the full READMERepository license: MIT
dsh-yolo-mode DSH plugin questions
How do I install dsh-yolo-mode?
Run `dsh plugin --profile web add dsh-yolo-mode` from the npm registry. Then add two insert entries to your `cordis.patch.yml` as described in the README. After restarting DSH and refreshing the browser, the plugin will be active.
What presets does dsh-yolo-mode offer?
It offers six presets: off, strict, balanced, permissive, yolo, and custom. Balanced is the default. Off disables the plugin entirely (all requests go to human). Strict is very conservative, only allowing workspace-write with strong justification. Yolo automatically allows all escalations without LLM calls. Custom lets you define your own levels table.
Can I use my own LLM model for judging?
Yes, you can set any provider and model that are available in your DSH Harness configuration. Use the `judge.provider` and `judge.model` fields in the plugin config. The plugin also provides a dropdown in the settings panel to select from configured models.
Is it safe to use the permissive or yolo preset?
The permissive preset treats uncertain outcomes as allowed-once, which can be risky. The yolo preset bypasses LLM entirely and allows all escalations. Both are intended for trusted environments or testing. The default balanced preset is recommended for production because it delegates to human on uncertainty.
What happens if the LLM model fails or times out?
The plugin is fail-closed. If the LLM call times out (default 20s), returns invalid output, or the model is unavailable, the plugin treats the outcome as an error and falls back to the configured error strategy (default: delegate to human). It will never auto-allow on error.