The Agent Skills Directory

[COMMAND_EXECUTION]: The skill installs lifecycle hooks that execute shell commands (powershell.exe, /bin/sh) automatically whenever the AI agent reaches specific lifecycle events like stopping or requesting input.
[COMMAND_EXECUTION]: Modifies agent configuration files in the user home directory (e.g., ~/.claude/settings.json, ~/.gemini/settings.json, ~/.codex/hooks.json) to inject persistent command hooks.
[COMMAND_EXECUTION]: Installer and runner scripts use the -ExecutionPolicy Bypass flag in PowerShell to execute local scripts, which bypasses default script execution restrictions.
[PROMPT_INJECTION]: The skill processes agent-generated output to create user notifications, creating a surface for indirect prompt injection.
Ingestion points: scripts/hook-runner.js reads agent event data from standard input.
Boundary markers: No explicit boundary markers or instructions are used to separate agent-provided content from command logic.
Capability inventory: scripts/hook-runner.js and scripts/ring.js use the ingested data to build arguments for subprocesses that trigger UI notifications.
Sanitization: scripts/ring.js and scripts/ring.ps1 implement character escaping for quotes and length limiting but do not perform full validation of the external content.

agent-doorbell