The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed entirely around executing external CLI tools (claude, codex, opencode, agy) that perform autonomous file system and shell operations.
[PRIVILEGE_ESCALATION]: Several configurations explicitly encourage bypassing security prompts and permission checks:
Recommends using --dangerously-skip-permissions with Claude Code to auto-approve file writes, shell commands, and network access.
Suggests using --sandbox danger-full-access with Codex to bypass bubblewrap/sandboxing when environment errors occur.
Recommends using --yolo mode for Codex which provides no sandbox and no manual approvals for faster execution.
[REMOTE_CODE_EXECUTION]: The skill facilitates remote code execution by design, as the delegated agents (Claude, Codex, OpenCode) autonomously fetch and execute logic from their respective model providers based on the task prompt.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for installing external agent software from npm and official taps (e.g., @anthropic-ai/claude-code, @openai/codex, opencode-ai). These are recognized well-known services, and while the installation itself is standard, the subsequent autonomous behavior of the tools is the primary risk factor.
[OBFUSCATION]: In references/opencode-cli-legacy.md, the skill instructs sending a raw hex escape sequence \x03 (Ctrl+C) to the terminal to exit sessions. While used for a legitimate purpose (process termination), sending raw hex codes to a shell can be a medium-risk pattern for hiding commands.

autonomous-coding-agents