The Agent Skills Directory

Persistence via Scheduled Tasks: The skill provides functionality to schedule its 'sleep cycle' using cron jobs via the schedule command. This establishes a persistent background process that runs automatically at specified intervals (e.g., nightly).
Sensitive Data Access: The optimization engine reads from ~/.claude/history.jsonl and project-specific .jsonl session logs. These files contain a comprehensive record of previous interactions, which may include sensitive code, command outputs, or accidentally disclosed credentials.
Dynamic Command Re-execution: A central feature of this skill is the 'Replay' stage, which re-runs tasks extracted from past sessions. This involves executing shell commands or code that was previously recorded in log files, representing a dynamic execution surface based on historical data.
Indirect Prompt Injection Surface: The skill ingests untrusted data from past session transcripts. If an attacker can influence the content of a previous session (e.g., via a malicious pull request or external input that was logged), they could potentially influence the 'Consolidate' phase to propose malicious changes to CLAUDE.md or SKILL.md during the next sleep cycle. The skill mitigates this through a 'held-out gate' and mandatory user adoption of proposed changes.

skillopt-sleep