AGENTS.md — Auto-Ontology agent entry point

Skills that teach an agent how to set up the current Auto-Ontology implementation and use its ontology layer. They live in skills/ (repo root), one directory per skill, under the durable nvidia-ontology* namespace.

This file is for developers bringing Auto-Ontology up and partners embedding it in an agent harness. Auto-Ontology maintainers (build, test, schema, frontend conventions) stay on CLAUDE.md. Do not fold those conventions into these skills.

Which skill to read

• skills/nvidia-ontology-setup/ — Helm (the official install), local Compose, developer --dev / --ds, MCP against an already-running instance, and troubleshooting. Start here when Auto-Ontology is not up yet or something failed during bring-up.
• skills/nvidia-ontology-management/ — inspect, model, modify, and safely publish through glossary terms, SQL attributes, semantic relationships, model YAML import/export, and governed result definitions. MCP cannot write; this skill covers the approved write paths.
• skills/nvidia-ontology-query/ — how an agent calls Auto-Ontology (MCP first, REST fallback), discovers meaning through the semantic layer, and validates SQL, rows, and answers. Start here when Auto-Ontology is already deployed.

Workflow ownership and handoffs

The composed lifecycle is install/connect/readiness → inspect/model → staged apply/readback → query validation → governed publication → fresh query readback. A step that lacks product support must return an explicit gap or receipt rather than inventing an operation. Authentication, permissions, provenance, and error handling remain shared concerns rather than separate overlapping skills.

Sources of truth (do not duplicate)

Skills encode workflow, order, and pitfalls. They must not become a second API catalog or a second installer.

MCP is read-only by design. Writes go through the Next.js /api/... surface. The live MCP tool list at handshake is the source of truth if it disagrees with anything in a skill.

Catalog mirror

These directories sync to github.com/nvidia/skills (npx skills add nvidia/skills) and build.nvidia.com/skills. That pipeline requires, per skill:

• SKILL.md, skill-card.md, evals/evals.json (maintained here)
• BENCHMARK.md and skill.oms.sig — generated by NVSkills CI. A maintainer comments /nvskills-ci on any pull request that changes skills/; the nv-skills-ci[bot] pushes an "Attach NVSkills validation signatures" commit that must stay in the PR. Never hand-edit or fabricate these files, and rerun NVSkills CI after any later change to a skill in the same PR.
• the components.d/auto-ontology.yml entry in nvidia/skills registering:

name: Auto-Ontology
repo: NVIDIA/auto-ontology
description: Set up Auto-Ontology, model and publish its semantic layer, and run validated text-to-SQL queries through it.
skills:
  - path: skills/nvidia-ontology-setup/
    catalog_dir: nvidia-ontology-setup
  - path: skills/nvidia-ontology-management/
    catalog_dir: nvidia-ontology-management
  - path: skills/nvidia-ontology-query/
    catalog_dir: nvidia-ontology-query

Skills must stay free of staging hostnames, cluster names, Vault paths, and credentials — the catalog is public even when Auto-Ontology itself is access-restricted.