The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill provides setup commands for Telegram and Discord notifications that instruct users to pass sensitive bot tokens and webhooks as plain-text command-line arguments. This practice exposes credentials to capture in shell history files, process monitoring tools, and system logs.
[PRIVILEGE_ESCALATION]: The instructions require the modification of the ~/.claude/settings.json file to enable experimental multi-agent features. Modifying hidden configuration files in the user's home directory allows the skill to alter the core security and operational behavior of the Claude Code environment.
[PERSISTENCE]: The skill includes a utility to start an "auto-resume daemon" (omc wait --start). This creates a background process that persists across sessions to automatically restart agent loops, which can result in the agent performing autonomous actions without active user supervision or review.
[EXTERNAL_DOWNLOADS]: The documentation suggests a global npm installation of a package named oh-my-claude-sisyphus. The discrepancy between this package name and the skill's name (omc) or repository name (oh-my-claudecode) is a suspicious pattern often associated with package confusion or supply chain attacks.
[REMOTE_CODE_EXECUTION]: The installation process involves adding a marketplace plugin directly from an external third-party GitHub repository (Yeachan-Heo/oh-my-claudecode). Installing unverified third-party plugins grants external code significant execution privileges within the agent's runtime context.
[INDIRECT_PROMPT_INJECTION]: The skill promotes an autonomous orchestration layer that processes natural language tasks through 32 specialized agents. This architecture provides a significant attack surface for indirect prompt injection, especially given the skill's lack of explicit boundary markers or sanitization for the high-privilege tools (Bash, Write, Edit) it accesses.

omc