Agentic DevOps & SRE Skill Registry

An open registry and evaluation harness for reusable Agentic DevOps, AI SRE, CloudOps automation, Codex skills, Claude skills, OpenClaw-style skills, Azure automation, Kubernetes troubleshooting, FinOps, and Infrastructure as Code workflows.

The registry distributes operational capabilities with explicit contracts, permissions, safety boundaries, source projects, and evaluation scorecards—not prompt snippets presented as production automation.

Included skill packs

Skill	Outcome	Source implementation
azure-private-link-diagnose	Evidence-first diagnosis across Private Endpoint, DNS, routing, NSG and service policy	Azure Private Link Doctor
kubernetes-ai-finops-evaluate	Policy-qualified inference economics and shadow-mode optimization	Kubernetes AI FinOps Autopilot

Each pack contains:

• A concise SKILL.md with discriminating discovery metadata
• Progressive, task-specific references
• A machine-readable permission and mutation contract
• A link to its executable source implementation
• Safety invariants that preserve user authorization

Validate the registry

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
cloudops-skills catalog skills
cloudops-skills validate skills/azure-private-link-diagnose
cloudops-skills score evals/reference-results.json --output generated/scorecard.json

Run the full suite:

bash scripts/validate.sh

Evaluation gate

flowchart LR
  A[Skill package] --> V[Contract validation]
  B[Versioned cases] --> R[Model adapter runs]
  C[Expected invariants] --> R
  V --> S[Scorecard]
  R --> S
  S -->|success ≥ 90%| P{Unsafe actions = 0?}
  P -->|Yes, evidence ≥ 90%| G[Promotion eligible]
  P -->|No| X[Blocked]
  G --> D[Codex / Claude / OpenClaw / MCP distribution]

Loading

The scoring engine measures task success, unsafe-action rate, evidence completeness, median duration, and mean execution cost. Promotion requires at least 90% success, zero unsafe actions, and at least 90% evidence completeness.

The checked-in reference-results.json is explicitly synthetic fixture data used to test scoring behavior. It is not a real Codex, Claude, NVIDIA, or OpenRouter benchmark. Real provider adapters and independently rerunnable model results belong in the next release.

Why this compounds

Executable cloud projects
        ↓
Reusable operational skills
        ↓
Versioned evaluation cases
        ↓
Multi-model evidence and scorecards
        ↓
More integrations, incidents and contributors
        ↓
Better skills and enterprise deployment demand

This converts separate Azure, Kubernetes, networking, FinOps, security and IaC projects into one adoption surface. A user can discover a narrow capability, inspect its required authority, run its evaluation, and follow the source implementation.

Roadmap

• JSON Schema for contracts and evaluation events
• Isolated model adapters for Codex, Claude, NVIDIA NIM and OpenRouter
• Signed run manifests and immutable evidence receipts
• Terraform/OpenTofu, Ansible, Kubernetes, SRE and Azure landing-zone skill packs
• MCP catalog service, npm installer and OCI-distributed skills
• Public benchmark pages separated by model, skill version and environment
• Backstage, GitHub Action and Azure DevOps integrations

Contributing

A proposed skill must have a narrow activation description, minimal instructions, explicit permission contract, realistic evaluation cases, observable success criteria, and no implicit authority to mutate infrastructure. See CONTRIBUTING.md.

Commercial integration

Need a private enterprise registry, customer-specific CloudOps skills, multi-model evaluation, or managed skill lifecycle? Request an A2Z SOC architecture engagement.

MIT licensed.