The runtime source of truth is each skills/<name>/SKILL.md. Every skill keeps its references inside its own directory so it can be installed independently.
| Client | Packaging | Explicit-only enforcement |
|---|---|---|
| Codex | .codex-plugin/plugin.json, agents/openai.yaml, or individual skill install |
policy.allow_implicit_invocation: false |
| Claude Code | .claude-plugin/plugin.json and marketplace |
disable-model-invocation: true |
| Cursor | portable root plugin.json or individual skill install |
disable-model-invocation: true where supported |
| GitHub Copilot | portable root plugin.json or individual skill install |
host-dependent; frontmatter supplies the boundary |
| OpenCode | individual skill install | soft guard in description and first workflow rule |
The frontmatter description is the only routing contract shared by every target. Audit and critique therefore state that they require explicit invocation even on clients that ignore client-specific policy fields.
Claude Code extends the Agent Skills frontmatter with
disable-model-invocation. Audit and critique intentionally use that field so
Claude hides them until the user invokes them. The strict core-only
skills-ref validate <skill> command therefore reports that one field as an unknown
extension for those two skills; plan and build pass it unchanged. Repository CI
validates the extension explicitly, and native Codex and Claude smoke tests
cover loading behavior. This is a known format-boundary tradeoff, not a claim
that the field belongs to the portable core specification.
allowed-tools is intentionally absent because its cross-client semantics remain experimental. Browser tools and subagents are optional capabilities, never hard dependencies.
Use the commands in the root README. A selected skill is copied recursively, but root documentation and shared maintainer files are not. This is why runtime links outside a skill directory fail validation.
Manual invocation syntax is client-specific:
- Codex and the portable eval dataset use
$completeness-auditand$consistency-critique. - Direct Claude skill installs use
/completeness-auditand/consistency-critique. - Claude plugin installs use
/web-app-completeness:completeness-auditand/web-app-completeness:consistency-critique.
The portable plugin and provider adapter are included. Live Cursor routing is advisory until a local cursor-agent installation has calibrated the canary trace. Codex and Claude are the live release gates for v0.1.