The Agent Skills Directory

[System Persistence]: The skill utilizes shell commands to install and manage nightly cron entries for background processing.
Evidence: SKILL.md contains instructions for scheduling the optimization cycle using cron to automate the 'sleep' cycle.
[Session History Data Access]: The tool reads interaction history from local logs to identify recurring tasks for optimization.
Evidence: SKILL.md (Stage 1: Harvest) describes reading ~/.claude/history.jsonl and project-specific logs.
[Remote Data Processing]: When configured with a model-driven backend, the skill transmits transcript excerpts to external API providers for task mining and reflection.
Evidence: SKILL.md notes that backends like claude or azure_openai send task content to providers and advises user caution regarding sensitive data.
[Indirect Prompt Injection Surface]: The optimization engine derives instructions and skill updates from past user-agent interactions, which could contain instructions intended to influence future behavior.
Ingestion points: Session history files processed in SKILL.md (Stage 1 and 2).
Boundary markers: Employs a 'validation gate' and 'held-out scoring' to verify that proposed changes result in objective improvement.
Capability inventory: Authorizes modifications to CLAUDE.md and SKILL.md and execution of scripts located in ${CLAUDE_PLUGIN_ROOT}.
Sanitization: Recommends manual review and redaction of logs and provides a deterministic mock backend for offline use.

skillopt-sleep