skill-eval

Skill Eval — Agent Skill Evaluation Framework

Evaluate Agent Skills across four dimensions: safety (audit), quality (functional), reliability (trigger), and cost efficiency (Pareto classification).

Quick Start

skill-eval audit /path/to/skill          # Is it safe?
skill-eval report /path/to/skill         # Full grade (audit + functional + trigger)
skill-eval functional /path/to/skill     # Quality: with-skill vs without-skill
skill-eval trigger /path/to/skill        # Reliability: activation precision

Decision Tree

• "Is this skill safe?" → skill-eval audit <path>
• "Full evaluation with grade" → skill-eval report <path>
• "Full repo security review" → skill-eval audit <path> --include-all
• "Write eval cases" → skill-eval init <path>, then edit evals/
• "Compare two versions" → skill-eval compare <old> <new>
• "Check for regressions" → skill-eval snapshot <path>, then skill-eval regression <path>
• "Track changes" → skill-eval lifecycle <path> --save --label v1.0

Commands

Command	Purpose
audit	Security & structure scan (secrets, permissions, spec compliance)
functional	Quality eval — runs prompts with and without skill, grades output
trigger	Reliability eval — tests activation precision for relevant/irrelevant queries
report	Unified grade combining audit (40%) + functional (40%) + trigger (20%)
compare	Side-by-side comparison of two skills on the same eval cases
snapshot	Save current audit as regression baseline
regression	Check for score regressions against baseline
lifecycle	Version tracking and change detection
init	Generate eval scaffold from SKILL.md frontmatter

For detailed flags and examples, see references/cli-reference.md.

Eval File Format

Functional evals (evals/evals.json):

[{"id": "case-1", "prompt": "...", "assertions": ["contains 'expected'"], "files": ["files/input.csv"]}]

Trigger queries (evals/eval_queries.json):

[{"query": "relevant question", "should_trigger": true}, {"query": "unrelated question", "should_trigger": false}]

Scoring

Grades: A (90+), B (80-89), C (70-79), D (60-69), F (<60). Findings deduct: CRITICAL −25, WARNING −10, INFO −2.

For the full security check reference and OWASP mapping, see references/security-checks.md.