The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to establish persistence on the host system using cron jobs, systemd timers, and infinite shell loops (while true). This allows the agent to maintain execution capabilities across sessions and reboots.
Evidence: Examples such as */10 * * * * cd /path/to/project && pi run and while true; do pi run "check X"; sleep 30; done demonstrate persistent execution patterns.
[PROMPT_INJECTION]: The skill explicitly directs the agent to use command-line flags to bypass interactive authorization prompts, which disables the user's ability to review or block actions during automated runs.
Evidence: Instructions to use --allowedTools for Claude or sandbox/auto-approve flags for Codex to ensure runs do not 'block on a prompt'.
[PROMPT_INJECTION]: The skill creates an attack surface for indirect prompt injection by instructing agents to process external task lists and pipe outputs from one scheduled job into the next without sanitization.
Evidence: The 'Heartbeat pattern' describes reading a task list to decide actions, and 'Hermes' chaining pipes output into the next session via context_from without explicit boundary markers.

agent-self-scheduling