dsh-gamepad-approval

When an AI Agent executes high-risk commands like rm -rf, git push --force, or disk formatting on your behalf, the approval step is often the weakest link in the security chain. A software-level dialog can be spoofed by same-privilege malicious code, keyboard confirmation can be captured by a keylogger, and remote desktop sessions can simulate clicks. An attacker does not even need your password; injecting a fake event at the software layer is enough.\n\ndsh-gamepad-approval moves that approval step from the software layer to the physical layer. When the Agent triggers a high-risk command, your Xbox-compatible gamepad buzzes twice: press A to approve, B to deny. The signal travels over the XInput protocol read directly via PowerShell, bypassing the browser, the network, and the keyboard entirely. Eighteen built-in permission rules cover file deletion, system shutdown, disk operations, database destruction, Git destructive actions, and key access, graded from critical to medium. The design follows a fail-closed principle: no controller, timeout, or disconnection means automatic rejection. After approval the command still passes through every built-in dsh security policy, so no other defenses are bypassed.\n\nThis plugin is ideal for developers running Agents inside DeepSeek Harness with system-level command privileges, or anyone who wants the approval action performed by a human physically present with a controller in hand. It takes effect immediately after installation; when no controller is connected, only high-risk commands are blocked while everyday operations continue as usual. Any XInput-compatible PC gamepad works.