Skill Radar

Meta-skill that acts as a routing layer for multi-skill agents. When installed, it intercepts user queries BEFORE other skills are loaded, scores all available skills using declarative routing rules, and instructs the agent to load ONLY the top-matching skills into context. This dramatically reduces system prompt bloat and improves skill dispatch accuracy. Zero ML infrastructure required.

Skill Radar — Declarative Skill Routing Engine

What this skill does

When you (the agent) receive a user query, do NOT load all available skills at once. Instead, follow the execution protocol below to route the query to only the most relevant skill(s).

---

Execution Protocol (D-mode: self-routing)

You MUST follow these steps for every user query when this skill is active:

Step 0: Check if routing is needed

Skip routing if:

• The user explicitly names a skill (e.g., "use the pdf skill")
• The query is a simple greeting or meta-question about you
• There are fewer than 4 skills installed

Otherwise, proceed to Step 1.

Step 1: Run the router

Execute the routing script with the user's query:

bash

python3 {SKILL_DIR}/skill_radar/cli.py route "{USER_QUERY}" --skills-dir {SKILLS_DIR} --format json

Where:

• {SKILL_DIR} = the directory where this skill (skill-radar) is installed
• {SKILLS_DIR} = the directory containing all other skills (each with a routing.yaml)
• {USER_QUERY} = the user's raw input text

Step 2: Read the output

The script returns JSON:

json

{
  "query": "review this contract for risks",
  "results": [
    {"skill": "contract-review", "score": 0.41, ...},
    {"skill": "code-review", "score": 0.12, ...}
  ],
  "excluded": [
    {"skill": "some-skill", "reason": "anti_pattern: draft a contract"}
  ]
}

Step 3: Load only matched skills

• If results is non-empty: load ONLY the skills listed in results (by name). Do NOT load other skills.
• If results is empty: fall back to your default skill loading behavior (load all, or use your own heuristic).
• If the script fails to run: fall back to default behavior. Never block the user.

Step 4: Proceed normally

With the filtered skill set loaded, respond to the user's query as usual.

---

Integration Modes

This skill supports 4 integration modes to work with any agent framework:

Mode D: Self-routing (recommended for ClawHub/OpenClaw/WorkBuddy)

The agent itself runs the routing script as described above. No framework changes needed — just install this skill and it provides routing instructions that the agent follows.

Mode A: CLI (for any framework with shell access)

bash

# Install
pip install skill-radar   # or: pip install -e /path/to/skill-radar

# Route a query
skill-radar route "review this contract" --skills-dir ./skills/ --format json

The framework calls this command before assembling the system prompt, and only includes the returned skills.

Mode B: Python SDK (for Python-based agents)

python

from skill_radar import load_skills

router = load_skills("~/.workbuddy/skills/")
results = router.route("review this contract")
# results = [ScoringResult(skill_name="contract-review", score=0.41, ...)]

# Only load these skills into your prompt:
skills_to_load = [r.skill_name for r in results]

Mode C: HTTP microservice (for cloud-based agents)

bash

skill-radar serve --skills-dir ./skills/ --port 8900

Then from your agent framework:

text

POST http://localhost:8900/route
Body: {"query": "review this contract", "context": {"file_types": [".docx"]}}

---

Setup: Adding routing declarations to your skills

Each skill needs a routing.yaml file declaring when it should trigger:

yaml

name: contract-review
description: "Legal contract review and risk analysis"

routing:
  keywords:
    - "contract review"
    - "review contract"
    - "NDA"
    - "agreement audit"
  patterns:
    - "(review|check|audit).{0,8}(contract|agreement|NDA|terms)"
    - "(contract|agreement).{0,6}(review|check|risk)"
  anti_patterns:
    - "draft a contract"
    - "contract template"
  priority: 80
  context:
    file_types: [".docx", ".pdf"]

Auto-generate routing.yaml

For skills that don't have routing declarations yet:

bash

skill-radar init --skills-dir ./skills/

This scans each skill's SKILL.md and auto-generates a basic routing.yaml from its metadata (name, description, trigger keywords).

---

Scoring Formula

text

Score(q, skill) = 0.30 × keyword_hit_ratio
               + 0.25 × pattern_matched
               + 0.15 × intent_match
               + 0.15 × context_bonus
               + 0.15 × (priority / 100)
               - anti_pattern_penalty

Anti-pattern hit = immediate exclusion (score forced to 0).

Threshold strategy (default: gap-based):

• If top-1 score leads top-2 by > 0.15 → only load top-1
• Otherwise → load all skills scoring above 0.30

---

File Structure

text

skill-radar/
├── SKILL.md                          ← This file (meta-skill instructions)
├── pyproject.toml                    ← Python package definition
├── skill_radar/                      ← Python package
│   ├── __init__.py                   ← SDK entry point
│   ├── core.py                       ← Framework-agnostic routing engine
│   ├── loader.py                     ← File system skill loading
│   ├── cli.py                        ← CLI entry point (route/init/serve)
│   ├── init_routing.py               ← Auto-generate routing.yaml
│   └── server.py                     ← HTTP server
├── references/
│   ├── scoring-theory.md             ← Mathematical foundations
│   └── routing-schema.md             ← Full YAML schema spec
├── assets/
│   └── skill-routing-config-template.yaml
└── examples/                         ← 6 cross-domain example skills

---

Important Notes

• This skill should be loaded with HIGH priority (it gates other skill loading)
• If routing script execution fails, ALWAYS fall back to default behavior — never block the user
• The routing decision is transparent: the JSON output includes matched keywords/patterns for auditability
• Skills without routing.yaml are invisible to the router — they will only load via fallback
• Routing adds ~50ms latency per query (regex matching, no network calls)