Skill Router

Skill Router

Meta system that analyzes natural language input to auto-select appropriate skill(s), determine order, and chain execution.

🚀 v2 Architecture: Low-level Call Protocol

Execution Flow

1. Scan only skills/*/SKILL.md frontmatter (trigger matching)
   - Quick match with description + trigger fields
   - No full body reading → 83% token savings
   
2. Check run field of matched skill for script path
   - run: "./run.sh" → skills/{name}/run.sh
   - run: "./run.js" → skills/{name}/run.js
   
3. Direct script execution with exec
   WORKSPACE=$HOME/.openclaw/workspace \
   EVENTS_DIR=$WORKSPACE/events \
   MEMORY_DIR=$WORKSPACE/memory \
   bash $WORKSPACE/skills/{name}/run.sh [args]
   
4. Agent processes stdout result
   - Parse if JSON
   - Pass through if text
   - Check stderr on error
   
5. Generate events based on events_out
   - Create events/{type}-{date}.json file
   - Subsequent skills consume via events_in
   
6. Check hooks → trigger subsequent skills
   - post: ["skill-a", "skill-b"] → auto-execute
   - on_error: ["notification-hub"] → notify on error

Skill Metadata Scan

# Extract only frontmatter from all skills
for skill in skills/*/SKILL.md; do
  yq eval '.name, .description, .trigger, .run' "$skill"
done

Execution Example

# User: "daily report"
# → trigger match: daily-report
# → Execute:
cd $HOME/.openclaw/workspace
WORKSPACE=$PWD \
EVENTS_DIR=$PWD/events \
MEMORY_DIR=$PWD/memory \
bash skills/daily-report/run.sh today

# Agent formats stdout result and delivers to user

Token Savings Effect

• Before: SKILL.md 3000 chars × 40 = 120KB (~30K tokens)
• v2: SKILL.md 500 chars × 40 = 20KB (~5K tokens)
• Savings: 83% token reduction

Core Concept

OpenClaw already selects 1 skill via description matching, but this skill:

1. Detect complex intent: "Analyze competitors and make card news" → competitor-watch + copywriting + cardnews + insta-post
2. Context-based auto-hooks: Auto-determine subsequent skills when a skill executes
3. Skill chain templates: Pre-define frequently used combinations

Intent Classification Matrix

Single Skill Mapping (1:1)

• "commit/push/git" → git-auto
• "DM/instagram message" → auto-reply
• "cost/tokens/how much" → tokenmeter
• "translate/to English" → translate
• "invoice/quote" → invoice-gen
• "code review/PR" → code-review
• "system status/health" → health-monitor
• "trends" → trend-radar
• "performance/reactions/likes" → performance-tracker
• "daily report" → daily-report
• "SEO audit" → seo-audit
• "brand tone" → brand-voice

Complex Skill Chains (1:N) — Core Pipelines

Context-based Auto-chain Rules

Skill A execution complete → analyze results → auto-determine next skill:

Auto-chain Rules (if → then)

• IF competitor-watch detects important change → THEN notification-hub(urgent) + include in daily-report
• IF tokenmeter exceeds $500/month → THEN notification-hub(urgent)
• IF code-review detects HIGH severity → THEN block commit + notification-hub
• IF think-tank conclusion has "immediate execution" action → THEN auto-record in decision-log
• IF cardnews generation complete → THEN confirm "post with insta-post?" (approval required)
• IF self-eval detects repeated mistake → THEN trigger learning-engine
• IF performance-tracker finds successful content → THEN suggest content-recycler
• IF trend-radar detects hot trend → THEN auto-suggest seo-content-planner
• IF mail detects important email → THEN notification-hub(important)
• IF health-monitor detects anomaly → THEN attempt auto-recovery + notification-hub(urgent)

Execution Engine Protocol

1. Receive user input
2. Classify intent (single vs complex)
3. If single → execute skill immediately
4. If complex → compose skill chain
   a. Skills without dependencies execute in parallel (sessions_spawn)
   b. Skills with dependencies execute sequentially (pass previous results via events/)
5. Check auto-chain rules on each skill completion
6. Auto-trigger additional skills if needed (or request approval)
7. Synthesize final results and respond

Auto-hook Registration

When skill-router activates, for all skills:

• pre-hook: Input validation + security check
• post-hook: Generate events/ event + check chain rules
• on-error: Error log + notification-hub

Skill Dependency Graph

[User Input]
    ↓
[skill-router] ← Intent classification
    ↓
┌─────────────────────────────────────────┐
│  TIER 1: Data Collection                │
│  competitor-watch, data-scraper,        │
│  trend-radar, tokenmeter, yt-digest     │
└─────────────┬───────────────────────────┘
              ↓ events/
┌─────────────────────────────────────────┐
│  TIER 2: Analysis/Thinking              │
│  think-tank, self-eval, seo-audit,      │
│  code-review, performance-tracker       │
└─────────────┬───────────────────────────┘
              ↓ events/
┌─────────────────────────────────────────┐
│  TIER 3: Production                     │
│  copywriting, cardnews, content-recycler,│
│  translate, invoice-gen                  │
└─────────────┬───────────────────────────┘
              ↓ events/
┌─────────────────────────────────────────┐
│  TIER 4: Deployment/Execution           │
│  insta-post, mail, git-auto,            │
│  release-discipline                     │
└─────────────┬───────────────────────────┘
              ↓ events/
┌─────────────────────────────────────────┐
│  TIER 5: Tracking/Learning              │
│  daily-report, decision-log,            │
│  learning-engine, notification-hub      │
└─────────────────────────────────────────┘

Safety Mechanisms

• Always require approval before external actions (email send, SNS post, payment)
• Prevent infinite loops: Stop after same skill chain repeats 3 times
• Cost limit: Max 5 subagents per chain
• Graceful stop on error + save partial results

🐧 Built by 무펭이 — Mupengism ecosystem skill