The Agent Skills Directory

[PERSISTENCE_MECHANISMS]: The skill provides commands to install a nightly cron entry via the 'sleep.sh schedule' command. This establishes a persistence mechanism that allows the skill's optimization and replay logic to execute automatically in the background without active user supervision.
[DATA_EXPOSURE]: The skill accesses sensitive application data stored in '~/.claude/history.jsonl' and project-specific session logs. These files contain full transcripts of user interactions, which frequently include sensitive information, proprietary code, or credentials. This data is harvested and can be transmitted to external model backends like Claude, Codex, or Copilot during the replay phase.
[INDIRECT_PROMPT_INJECTION]: The skill implements an automated feedback loop that modifies core agent configuration files 'CLAUDE.md' and 'SKILL.md' based on historical session data. This creates a significant attack surface where malicious or adversarial content from past sessions could be permanently consolidated into the agent's long-term instructions. (1) Ingestion points: '~~/.claude/history.jsonl' and session logs in '~~/.claude/projects/'. (2) Boundary markers: No explicit security boundaries or markers are used to delimit untrusted historical data. (3) Capability inventory: The skill has the capability to write to core project configuration files, execute shell scripts, and manage cron jobs. (4) Sanitization: The skill performs validation for performance metrics but lacks security-focused sanitization or filtering to prevent the consolidation of malicious instructions.

skillopt-sleep