##A Little Bit of Context:
I am pushing in my company to adopt skills, but this seems like a "must have" for Enterprise clients, every software engineer that has worked with Maven/Npm or so will ask for something like that.
I am posting everything required to adopt this change request, code for the "validator" included, in the repo mentioned above.

Let me know if you like the idea, i was gonna implement this for myself and just felt like sharing it.

Summary

This RFC proposes two optional additions to the Agent Skills specification:

1. requires — Declare dependencies between skills with version validation
2. test — Declare test cases for skill validation

I've seen related discussions in #90 (skill relationships) and #46 (versioning). This proposal takes a complementary approach focused on tooling-side validation rather than agent-side hints, plus adds testing which hasn't been addressed yet.

Relationship to Existing Proposals

The key difference: #90 helps agents reason about skill combinations. This RFC helps tooling catch problems before skills reach agents.

Both can coexist — prerequisite-skills for agent hints, requires for tooling validation.

The Problems

1. Silent breakage

Skill A assumes skill B exists. Someone modifies B. A breaks at runtime with no warning.

2. Version drift

Skill A was developed with B@1.2.0. Environment has B@1.0.0. Subtle bugs emerge.

3. No standard testing

Every organization invents their own approach. Skills can't be validated portably.

Proposed Solution

requires field

---
name: integration-test-runner
description: Runs integration tests on specified environment.

metadata:
  version: "2.1.0"

requires:
  - skill: environment-selector
    version: "1.2.0"
  - skill: logging-standards
---

Validation (at submission time):

• environment-selector must exist
• Its version must be >= 1.2.0
• Circular dependencies are rejected

Agents require no changes. They never see requires — tooling validates before skills become available.

test field

---
name: environment-selector
description: Selects deployment environment.

test:
  cases: test/cases.yaml
---

# test/cases.yaml
cases:
  - name: select_dev
    input: "Select DEV environment"
    assertions:
      output_contains:
        - "DEV"
      output_not_contains:
        - "PROD"
        - "error"

  - name: require_prod_confirmation
    input: "Select PROD environment"
    assertions:
      output_contains:
        - "confirm"
      semantic_match:
        criterion: "Response asks for explicit confirmation before production deployment"

Assertions are simple and agent-agnostic:

• output_contains / output_not_contains — string matching
• output_matches — regex
• semantic_match — LLM judge for fuzzy criteria

Tests are deterministic. No retry logic. If a well-written criterion fails, either the skill is broken or the test needs fixing.

Design Decisions

Reference Implementation

I've built a Python tool that extends skills-ref:

# Validate dependencies and versions
skills-ref validate ./my-skill --skills-root ./skills

# Initialize skill with dependency snapshot
skills-ref init ./new-skill --skills-root ./skills

# Run tests
skills-ref test ./my-skill

# Check for circular dependencies
skills-ref deps --check-circular

# Visualize dependency graph
skills-ref deps --graph

Repository: https://github.com/AndoSan84/scalble_skills

Full RFC Document

Complete specification with schema definitions, validation semantics, security considerations, and migration guide:

📄 RFC-001-dependencies-testing.md

Questions

1. Does this complement Proposal: Skill Relationship Fields - adding prerequisite_skills and related_skills to the SKILL.md spec #90 well, or should we merge the approaches?
2. Is the testing spec useful? Any concerns about semantic_match?
3. PR to skills-ref or separate package?

Thanks for considering!