yolo-mode
Unattended full-access (yolo) mode for the DeepSeek Harness: a human-only /yolo switch arms danger-full-access + never-approval for long autonomous runs, with a best-effort catastrophic-command tripwire guard, optional time-boxed expiry, and idle/denial notifications (desktop / webhook / email)
AI 分析
核心用途是允许 Agent 在无需人工确认的情况下自主运行长任务。适合需要高强度自主执行、同时需要熔断保护的安全敏感型任务。
インストール
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:CanGeng/yolo-modeドキュメント
README 全文を読む ↗Usage
| Command | Effect |
|---|---|
/yolo on | Arm (no expiry by default, until /yolo off) |
/yolo on 4h · /yolo on 90m · /yolo on 1h30m · /yolo on 2d · /yolo on 45 (bare number = minutes) | Arm with an expiry; auto-reverts when it passes |
/yolo off | Disarm now; reverts to the pre-arm permission snapshot |
/yolo status | Armed state, remaining time, revert target, guard strike count, notify channels |
Arming writes the canonical sandbox/mode and approval/policy session events through their canonical setters, so the model-visible runtime context updates immediately, the state survives restarts via log replay, and the Web UI's permission selector shows danger-full-access.
Configuration
Every knob lives in the yolo-mode row of your profile's cordis.patch.yml — see cordis.patch.yml for the annotated default configuration (guard.*, notify.*, context.*).