Somnia#

Overview#

Somnia is the sleep-cycle maintenance layer for OpenClaw skills. It provides a repeatable review workflow that checks installed skills during quiet hours, summarizes health risks, and writes proposal artifacts without silently mutating runtime skills.

Current version: v0.4.3 "Standalone Safety".

Trigger Cues#

Use this skill when the user mentions:

• nightly skill review
• sleep-time maintenance
• skill health report
• skill bug scanning
• replay regression check
• feedback-driven upgrade
• proposal-based update
• Somnia

Default Workflow#

1. Confirm the review scope: managed skills, feedback-related skills, or all installed skills.
2. Run lightweight package validation, feedback summary, and replay-case availability checks for each selected skill.
3. Write JSON and Markdown health reports under the configured learning/report directory.
4. Write proposal artifacts only when feedback or quality gates justify the change.
5. Hand proposal artifacts to Skill Forge or a human maintainer before any install decision.
6. Keep simulated evaluation details hidden from user-facing reports.

Output Contract#

The final answer or artifact should include:

• Review scope and schedule assumption
• Skills checked and health summary
• Issues found, grouped by skill
• Update candidates proposed or blocked
• Replay and hidden-evaluation pass/fail summary
• Next action: no-op, review proposal, approve install, or adjust schedule

Quality Gates#

• Never auto-install skill changes; Somnia writes proposals and reports only.
• Keep hidden evaluation and replay case details out of user-facing Telegram reports.
• Redact feedback-derived content before it becomes a replay case or report item.
• Prefer proposal files and manifests over direct mutation of installed skills.
• Keep Somnia self-contained; do not execute out-of-package Skill Forge code.

Resources#

References:

• references/somnia-architecture.md
• references/schedule-and-policy.md

Scripts:

• scripts/nightly_skill_review.py
• scripts/schedule_nightly_review.py