Skip to content

Skill Registry

v4.1 Sprint A adds an optional skill registry foundation for MultiModel Dev OS.

This does not replace existing markdown skills. Existing files under .ai/skills/ remain valid.

Files

text
.ai/registries/skills.yaml
.ai/schema/skill.schema.json

The registry provides example metadata for reusable skills such as:

  • Release governance
  • Security audit
  • SEO audit
  • Content brief
  • Document production
  • Business ops
  • Code review

Sprint F Status

The registry remains declarative. Sprint C added read-only CLI inspection, Sprint E allows workflows to reference skill IDs as metadata, and Sprint F adds draft-only business operator template skills:

  • No runtime behavior change.
  • No automatic skill triggering.
  • No permission enforcement.
  • Validation checks required fields, slug-safe IDs, semver-like versions, risk levels, permission classes, safe relative paths, and referenced files.
  • multimodel-dev-os skill-os list skills prints known skill IDs.
  • multimodel-dev-os skill-os show skill <id> prints selected metadata.
  • Workflow skill_os.skills references are validated, but they do not trigger skill execution.
  • Business operator skills use the business-operator category and draft-only permissions.

Future sprints may add richer inspection before any execution integration.

Metadata Shape

Skill entries are expected to support:

  • id
  • name
  • version
  • description
  • category
  • triggers
  • required_context
  • provided_outputs
  • risk_level
  • permissions
  • recommended_models
  • token_budget
  • skill_file
  • checks
  • examples

Compatibility

The registry is an overlay on top of markdown-compatible skills. A workspace can continue using .ai/skills/*.md without a registry.

Registry entries should point to safe relative workspace paths and use the shared tool permission classes:

  • read-only
  • draft-only
  • write-with-confirmation
  • restricted-admin
  1. Keep current markdown skills.
  2. Add metadata only for skills that need discovery or routing.
  3. Use low-risk or draft-only permissions by default.
  4. Add checks for skills that may lead to writes or release operations.
  5. Treat validation as a safety net, not as automation or permission enforcement.

See Skill OS CLI for read-only inspection commands. See Business Operator Layer for the Sprint F operator templates. See Skill OS Migration Guide for converting markdown-only skills into registered metadata.

Released under the MIT License.