skill-usage-governor（技能使用调节器）

What This Skill Does

Governs the lifecycle of skills under the skills/ directory: tracks usage frequency, identifies cold (inactive) skills, provides cleanup recommendations, and performs recoverable soft-archiving.

Quick Start

# 1. Collect usage data and generate rankings
python3 scripts/collect.py && python3 scripts/rank.py

# 2. Enable multi-platform hook injection (Gemini/Claude/Windsurf/Vibe)
# Critical step: this script must be run for the governor to detect skill usage
python3 scripts/enable_governor.py

# 3. Dry-run archive (no actual changes)
python3 scripts/archive.py --dry-run

Activation & Proactive Guidance

This skill supports multi-platform governance: OpenClaw, Gemini CLI, Claude Code, Windsurf, Vibe CLI.

🚨 Agent Awareness Rules

1. First-contact prompt: When the user first installs this skill or asks “how to manage/clean up skills”, you must check the current hook injection status. If not enabled, inform the user in natural language: “The skill is installed, but hook injection is not yet active — I cannot track your skill usage in real time. Would you like me to run the injection script to enable fully automatic governance?”
2. No false activation claims: If check_activation.py returns not_enabled, you must never tell the user that governance is active. Clearly identify the missing component (e.g., OpenClaw entries missing, Gemini settings not wired, etc.).
3. Injection value proposition: Inform the user that once injection is enabled, the governor works like a “silent sensor” — accumulating cleanup recommendations without manual intervention.

Verification Criteria

If the user requests activation, guide them to run scripts/enable_governor.py, then verify:

1. The adapter layer hooks/adapter.py exists
2. The corresponding CLI config directory (e.g., ~/.gemini or ~/.claude) has the hook injected
3. python3 scripts/check_activation.py passes

Only confirm “activated” after all checks pass.

Core Output

Reports contain the following information:

See references/output-samples.md for detailed report samples and field descriptions.

Safety Boundaries

1. Never hard-delete — only recoverable soft-archiving
2. Dry-run before archiving — policy.yaml defaults to dry_run_default: true
3. Archives must be recoverable — via manifest files + restore.py
4. Cold skills are only suggested, never auto-acted upon — no destructive actions without user confirmation
5. Protected skills whitelist — skill-usage-governor, long-running-agent, self-improvement are excluded from archive evaluation

User-Facing Communication

When communicating with users, use natural language. Do not expose variable names, stat field names, or script internals.

Examples:

• Good: “Your most-used skills in the past 7 days are tmux and codex. rag-everything-enhancer hasn’t been used in 30 days.”
• Bad: “activation_count_7d: 12, uses_30d: 0, trailing_percentile: 0.34”

Files in this skill

Verification Baseline

At minimum, verify:

• collect.py + rank.py can generate data/usage_stats.json and data/report-latest.md
• archive.py --dry-run runs normally without modifying any directories
• After a real archive, restore.py --skill <name> can restore the skill
• check_activation.py correctly determines activation status

Exit Criteria

The current task is considered complete when:

• Usage data collection is finished
• Results and reports have been generated
• Dry-run or real archive/restore verification is complete
• A user-readable conclusion has been provided