skill-net

Analyze OpenClaw skill ecosystem — dependencies, orphan detection, ecosystem health score, impact analysis, and skill relationships. Use when the user asks about skill relationships, "what depends on X", "if I delete Y what breaks", ecosystem health, or wants to find skills without trigger conditions (orhpans).

OpenClaw Skill Net

Analyze, map, and diagnose the OpenClaw skill ecosystem — not a skill creator, a diagnostic lens.

---

Core Positioning

This skill answers: how does my skill ecosystem actually work?

It scans every SKILL.md, detects dependency relationships, scores ecosystem health, and finds orhpans.

---

Modes

Mode 1: Full Analyze (default)

Run the complete ecosystem scan and produce a full diagnostic report.

Trigger: "analyze ecosystem", "full scan", "ecosystem health", "skill health", "技能生态", "生态报告"

Language options (CLI flags):

python3 scripts/analyze_deps.py              # default: ZH then EN
python3 scripts/analyze_deps.py --lang=ZH   # Chinese only
python3 scripts/analyze_deps.py --lang=EN   # English only
python3 scripts/analyze_deps.py --lang=BOTH  # ZH then EN (default)

Output sections:

1. 🌡️ Ecosystem Health Score (0–100) — 生态健康分
2. 📊 Health Breakdown — 健康度明细 (trigger coverage, metadata, cross-references)
3. 🔵 Core Hubs — 核心枢纽 (referenced by 3+ skills)
4. 🟡 Bridge Connectors — 桥接技能
5. 🟢 Leaf Skills — 叶节点技能
6. ⚪ Isolated Skills — 孤立技能
7. ⚠️ Orphan Skills — 孤儿技能 (have SKILL.md but no trigger conditions)
8. 🗑️ Impact Analysis — 删除影响分析 (who breaks what)
9. 📋 ASCII Ecosystem Map — 技能生态地图

All sections rendered in the requested language (ZH/EN) with full bilingual labels.

Mode 2: Query

Answer specific questions from cached or fresh data.

Trigger: "what depends on X", "if I delete Y", "who references Z", "core skills", "most connected skill"

Execution: Answer from data/ecosystem.json or run fresh scan.

Mode 3: Orphan Scan

Find all skills with SKILL.md but missing trigger conditions.

Trigger: "find orphans", "skills without triggers", "dead skills", "missing triggers"

Output: List of orphan skills with line count and frontmatter name.

Mode 4: Compare

Compare two skills side-by-side.

Trigger: "compare X and Y", "X vs Y dependencies", "skill X relationship to Y"

Output: Shared mentions, relationship type, overlap analysis.

---

Key Findings From Real Data

The ecosystem reveals structural patterns invisible from casual observation:

Finding	Evidence
True core hub: review	53 skills reference /review — by far the most connected
qa is a secondary hub	9 skills reference /qa
/summarize, /weather	Referenced by 2+ skills each — utility anchors
100/123 skills lack triggers	Many use /protocol style instead of "use when"
Ecosystem Health: 22.6/100	Most skills missing metadata and trigger conditions
review and qa are invisible hubs	They don't use skill- prefix — protocol commands

---

Execution Steps (Full Analyze)

1. Scan — walk ~/.openclaw/skills/ and ~/.openclaw/workspace/skills/, read every SKILL.md
2. Extract — for each skill:
   • Frontmatter fields (name, version, license, metadata)
   • Trigger presence (use when / trigger / /protocol)
   • ALL cross-skill name mentions (full scan, not just known slugs)
   • Metadata blocks
3. Build graph — mentions (outgoing) + referenced_by (incoming) for each skill
4. Classify — Core (≥3 incoming) → Bridge → Leaf → Isolated
5. Detect orhpans — has SKILL.md but no trigger phrase detected
6. Score health — weighted formula across 4 dimensions
7. Render — bilingual ASCII report (ZH/EN) + save data/ecosystem.json + data/report.md

---

Ecosystem Health Formula

Health Score = (
  trigger_coverage   × 30%  +
  metadata_complete  × 20%  +
  cross_reference     × 20%  +
  ecosystem_cohesion  × 30%
)

Your ecosystem: 22.6/100 — healthy room for improvement.

---

What the Real Data Reveals

Surprising insight: The most-connected nodes are protocol commands (/review, /qa), not skill-* named skills. These protocol skills are referenced by code patterns like:

# Many skills open with this:
# /review — Structured Code Review Protocol
# /qa — Quality Assurance Execution Protocol

This means traditional dependency detection (looking for skill-X mentions) severely underestimates real relationships.

True dependency types:

1. Named skill mentions — skill-factory, gupiao, bazi
2. Protocol command references — /review, /qa, /careful, /cso
3. CLI tool references — clawhub, mmx, summarize, weather

---

Do not

• Do not modify any skill based on the analysis without explicit user request
• Do not publish the ecosystem map without context — it's a diagnostic tool
• Do not call skills "broken" just because they lack trigger phrases — many use protocol-style activation
• Do not include .git/, .venv/, __pycache__/ in scans

---

Quality Bar

The output must:

• Scan both ~/.openclaw/skills/ and ~/.openclaw/workspace/skills/
• Correctly identify incoming + outgoing references per skill
• Detect orhpans (has SKILL.md, no trigger phrase)
• Compute ecosystem health score (0–100)
• Complete full scan in < 15 seconds for 120+ skills
• Save structured data to data/ecosystem.json + data/report.md
• Support bilingual output (ZH/EN) via --lang flag

---

Good vs Bad Examples

Good:

"🔵 review (Core Hub, 53 incoming): /review is referenced by 53 skills. If deleted, these skills lose their review protocol: gupiao, proactive-agent, skill-vetter..."

Bad:

"Here are all skills listed alphabetically"

Good Orphan Report:

"⚠️ Found 100 orhpans — most are protocol-style skills (review, qa, careful, cso) that use /command activation instead of 'use when' phrases. These are not broken, just designed differently."

Good Query:

"DELETE review → Breaks 53 skills including: gupiao, marketing-, engineering-, testing-, project-. This is the most critical skill in the ecosystem."