skill-guard 0.6.0

The quality gate for Agent Skills — validate, secure, conflict-detect, and test skills across their full lifecycle

skill-guard

The quality gate for Agent Skills.

skill-guard is a CLI tool that validates, secures, and governs Agent Skills across their full lifecycle — from contribution to production monitoring.

The Problem

Agent Skills are powerful. They're also ungoverned. As soon as more than one person contributes skills to a shared agent, things break in hard-to-diagnose ways:

• A new skill's description overlaps with an existing one → agent picks the wrong skill half the time
• Skills with dangerous scripts get merged because nobody reviewed the scripts/ directory
• Nobody knows what skills are installed, who owns them, or whether they still work
• A skill passes every test in isolation but fails when the real agent uses it with 25 other skills loaded

skill-guard is the quality gate that catches these problems before they reach production.

What It Does

ONBOARDING (pre-merge, in CI):
  skill-guard validate   → format compliance + quality scoring
  skill-guard secure     → scan for dangerous patterns  
  skill-guard conflict   → detect trigger overlap with existing skills
  skill-guard test       → runs evals against an OpenAI-compatible endpoint. Use pre_test_hook/post_test_hook for your own deploy/teardown flow.
  skill-guard check      → runs validate + secure + conflict as a single gate. Agent evals run if --endpoint is configured.

ONGOING (post-merge, scheduled):
  skill-guard monitor    → re-run evals, detect drift, manage lifecycle. Run via cron or CI for continuous drift detection. No built-in scheduler.
  skill-guard catalog    → searchable registry of registered skills (approval workflow planned for v0.7)

Quick Start

pip install skill-guard

# Initialize in your skills repo
skill-guard init

# Validate a skill
skill-guard validate ./skills/my-skill/

# Check for security issues
skill-guard secure ./skills/my-skill/

# Check for conflicts with existing skills
skill-guard conflict ./skills/my-skill/ --against ./skills/

# Run the full gate (validate + secure + conflict; test runs if --endpoint is configured)
skill-guard check ./skills/my-skill/ --against ./skills/

Example Output

$ skill-guard validate ./skills/my-skill/

 skill-guard validate — my-skill
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Check                     ┃ Result                                           ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ skill_md_exists           │ ✅ SKILL.md found                                │
│ valid_yaml_frontmatter    │ ✅ Valid YAML frontmatter                        │
│ name_field_present        │ ✅ name: my-skill                                │
│ description_field_present │ ✅ description field present                     │
│ directory_name_matches    │ ✅ Directory name matches skill name             │
│ description_trigger_hint  │ ✅ Description contains trigger hint ('Use when')│
│ no_broken_body_paths      │ ✅ No broken relative paths in SKILL.md body     │
│ evals_directory_exists    │ ⚠️ No evals/ directory found                     │
│                           │ → Create evals/config.yaml with test cases       │
│ metadata_has_author       │ ✅ author: my-team                               │
│ metadata_has_version      │ ✅ version: 1.0                                  │
└───────────────────────────┴──────────────────────────────────────────────────┘
Score: 97/100 | Grade: A | Blockers: 0 | Warnings: 1

Installation

Prerequisites

Requirement	Version	Notes
Python	3.11+	Required. 3.12 and 3.13 tested.
pip	any recent	Bundled with Python
typer	≥0.13.0	Installed automatically
Agent endpoint	—	Required only for skill-guard test (OpenAI-compatible API)

Note: skill-guard validate, secure, conflict, init, catalog, and check work fully offline — no agent or API key needed.

Installation

# Core (static analysis — no agent required)
pip install skill-guard

# Optional embeddings support
pip install skill-guard[embeddings]

# Optional LLM-based conflict detection
pip install skill-guard[llm]

Conflict Detection Modes

# TF-IDF (default)
skill-guard conflict ./skills/my-skill/ --against ./skills/ --method tfidf

# Embeddings-based overlap detection
skill-guard conflict ./skills/my-skill/ --against ./skills/ --method embeddings

# LLM-based overlap detection
export OPENAI_API_KEY=...
skill-guard conflict ./skills/my-skill/ --against ./skills/ --method llm

embeddings uses the all-MiniLM-L6-v2 sentence-transformers model and caches vectors under .skill-guard-cache/embeddings/. llm uses the OpenAI Chat API with gpt-4o-mini by default.

Documentation

• Getting Started
• End-to-End Integration Guide ← start here for real agent setup
• Writing Evals
• Hook Scripts
• CI/CD Integration
• Configuration Reference

Anthropic Spec Validation

skill-guard validate includes Anthropic AgentSkills spec compliance checks by default. Set validate.anthropic_spec: false in skill-guard.yaml if you need to disable those additional findings.

Exit Codes

• 0: success
• 1: validation/security failures
• 2: warnings only (when fail_on_warning is false)
• 3: config error
• 4: parse error
• 5: hook script failure
• 6: health check timeout

Pre-commit

Use pre-commit to enforce checks before skill changes land:

repos:
  - repo: https://github.com/vaibhavtupe/skill-guard
    rev: v0.6.0
    hooks:
      - id: skill-guard-validate
      - id: skill-guard-secure
      - id: skill-guard-check

These hooks run against changed SKILL.md files, deduplicate by skill root, and then execute the corresponding skill-guard command for each affected skill.

Templates

Use skill-guard init --template base to scaffold a new skill, or skill-guard init --list-templates to see the available scaffolds. Generated templates include SKILL.md, evals/, references/, scripts/, and assets/ so they validate immediately.

GitHub Actions

- uses: vaibhavtupe/skill-guard-action@v1
  with:
    path: ./skills/my-skill

See vaibhavtupe/skill-guard-action for the full action repo.

GitHub Actions

Use the separate action repo vaibhavtupe/skill-guard-action@v1 in workflows:

- uses: vaibhavtupe/skill-guard-action@v1
  with:
    command: check
    path: ./skills/my-skill
    against: ./skills/

What skill-guard Does NOT Do

• Does not replace Anthropic's skill-creator for writing skills
• Does not host or serve skills — skills live in your repo
• Does not modify skills — it reports issues, authors fix them
• Does not require a database or server — the catalog is a YAML file in your repo

Contributing

See CONTRIBUTING.md. We welcome contributions of all kinds.

License

Apache 2.0. See LICENSE.