CanGeng/yolo-mode

CanGeng★ 0JavaScript最后同步: 2026-08-18

在 GitHub 打开

Unattended full-access (yolo) window for the DeepSeek Harness: human-only /yolo switch, lazy expiry with auto-revert, catastrophic-command tripwire guard, desktop/webhook/email notifications

README 摘要

yolo-mode English 中文 An unattended full-access window plugin for the DeepSeek Harness ( dsh ): when you walk away and want the agent to grind through a long task, one human-only command — /yolo on — arms danger-full-access sandbox + never approval for that session, with an automatic way back, a tripwire against catastrophic commands, and notifications when the work is done or something needs attention. Read this first. In yolo mode the model executes with your uid and full filesystem reach — every operation is equivalent to your own. The bundled guard is a best-effort tripwire against accidents, not a security boundary (a regex blocklist is trivially bypassed by deliberate obfuscation). Arm it only for sessions whose task you would be comfortable running yourself. See Residual risks. Why Per-call approvals and the workspace sandbox turn every wall into a stall when nobody is at the keyboard. dsh already ships a danger-full-access permission preset; what the unattended case additionally needs is: - a fast, unambiguous, human-only switch (the model never gets a tool to grant itself access), - an automatic way back (time-boxed expiry, circuit breaker), - a guard for the operations no …

在 GitHub 查看完整 README →
Agent/智能体agentdsh-plugindsh-plugin-marketdsh-plugins

分类