Skill management

Follow AGENTS.md for scope and authority. Use the available skill-creator guidance for authoring; this card manages existing discovery locations.

Scope and commands

Standard Codex skills live in .agents/skills/<name>/SKILL.md (repository) or ~/.agents/skills/ (user). This OMX installation also exposes .codex/skills/ and the configured Codex home skills directory. Inspect the active catalog and preserve a requested path; do not duplicate names across roots visible to the same host.

• list / scan: inventory names, descriptions, paths, and duplicate entries.
• search <query> / info <name>: find metadata/body matches or read the selected skill.
• add / edit: infer routine metadata from the requested purpose, load the existing file/references, then make the scoped change.
• sync: compare requested roots and copy only the selected skills, preserving unrelated content.
• remove: resolve the exact skill and its references before removal; confirm destructive targets when authorization is missing.
• validate: inspect without writing; report invalid metadata, unresolved local references, duplicate discovery, and incomplete templates.
• setup: inspect current discovery and create only requested missing entries.

Use the supplied command directly. If no operation can be inferred, ask what the user wants to manage.

Authoring and validation

Require YAML name and description; use a concise scope-specific description. Keep optional UI/invocation metadata in agents/openai.yaml and preserve existing policy. OMX-specific metadata is an extension, not a requirement for portable skills.

Each body describes the outcome, necessary constraints, relevant reference branches, and observable completion. Put large conditional procedures in linked references; reuse the project's existing recipe instead of copying it.

Validate YAML, name/description, referenced paths, and any changed scripts. Use the installed skill-creator validator when available; distinguish unsupported extension warnings from invalid YAML. Review consequential routing with realistic tasks. A static check cannot prove runtime behavior.

Report affected paths and validation evidence. Preserve files on cancellation or read-only commands.