@narrativetrace/skills

The published skills carrier: narrativetrace-doctor and add-narrative-tracing as resources — one SKILL.md per platform (agents/, claude/) plus a versionless catalogue.json. No code — what an installer reads out of this package is text. The typed catalogue that renders it (the schema, the lints, the eval harness) lives in the private @narrativetrace/skills-catalogue; pnpm run skills-render there writes every file under this package as one more render target, and pnpm run skills-check (wired into pnpm run check) fails the build the moment a page here drifts from the typed source.

agents/
├── narrativetrace-doctor/SKILL.md
└── add-narrative-tracing/SKILL.md
claude/
├── narrativetrace-doctor/SKILL.md
└── add-narrative-tracing/SKILL.md
catalogue.json

agents/ is Codex's own discovered .agents/skills/ layout; claude/ is Claude's plugin layout — the two directories hold the same two skills with different frontmatter, never a rendering choice made by whatever reads this package. catalogue.json names every skill once ({name, description, agents, claude}) with no version literal: the stamp is this package's own package.json version, read by whatever consumes the carrier, so a copy can never go stale independently of the package it shipped in.

@narrativetrace/cli bundles a byte-identical copy of this package's contents under skills/ so npx @narrativetrace/cli init works offline from the one package a consumer already fetched — the same content, two homes, checked byte-for-byte against each other.

Licensed Apache 2.0 — like @narrativetrace/cli, and unlike the Business Source License 1.1 runtime packages published from this repository (see LICENSE).