Skill Garden

Automatically improves installed skills through passive usage observation and periodic batch analysis. Activates after any skill is used, or when you say gro...

Before installing or enabling Skill Garden, consider the following: - This skill needs read/write access to your ~/.openclaw/workspace/skills tree and will read logs and edit other skills' SKILL.md files. That is necessary for its purpose but is a powerful privilege — back up your skills directory first. - SKILL.md promises 'user-in-control' but the decision rules auto-apply edits at high and mid confidence (≥90% and 70–89%). If you want manual approval for all edits, change the min-confidence / auto-apply thresholds or run batch_analyze.py in --dry-run mode only. - The bundled batch_analyze.py listing appears truncated/buggy; review the full scripts in a safe environment before letting them run. Look for file-write locations, how edits are applied, and any notification logic. - Prefer a conservative deployment: run the scripts manually (dry-run) first, inspect proposals in references/improvement_proposals.md, and only allow auto-apply after you validate behavior. - If you must test, limit the skill to a copy of your workspace or exclude high-value/production skills from its scope until you’re comfortable. If you want, I can: (1) point out exact file lines that perform writes, (2) suggest a minimal safe configuration (disable auto-apply, change cron to manual), or (3) produce step-by-step instructions to sandbox and test the skill safely.

Type: OpenClaw Skill Name: skill-garden Version: 1.0.0 The 'skill-garden' bundle is a meta-skill designed to monitor, analyze, and automatically modify other skills within the OpenClaw workspace. It implements a 'Skill Evolution Engine' that uses Python scripts (batch_analyze.py, log_insight.py) to track usage logs across all installed skills and generate improvement proposals. Most significantly, SKILL.md instructs the agent to automatically use the 'edit' tool to modify other skills' instructions (SKILL.md files) when confidence thresholds are met. While the stated intent is optimization, the bundle's core functionality—cross-skill modification, persistence via cron jobs, and continuous observation of all agent activities—represents a high-risk capability that could be used to silently alter agent behavior or compromise the integrity of the skill library.