Authoring a built-in Lagune sub-skill (development workspace)

The end-user command @spec/commands/lagune.specialize.md writes into a scaffolded target project: .lagune/skills/<name>.md and .lagune/skills.json. Neither path exists here. This repo is the Lagune source, where built-in sub-skills live in spec/skills/<name>.md and the catalog is a hardcoded TypeScript array, not a runtime JSON file. There is also no "user" registry: a built-in is registered in code, compiled, scaffolded, and tested.

So you reuse the command's authoring discipline (terrain, defense-only gate, template, tags, reconcile-never-append) but override every output path to the native layout, and you update the registration points the end-user command never has to.

Step 1: Run the command's authoring steps, with redirected paths

Read and follow @spec/commands/lagune.specialize.md in full. Its authoring steps hold unchanged. The one thing that does not is where files are read and written, since the command targets a scaffolded project and you are in the source. Remap every path it names:

Command path (end-user target)	Native path (use this instead)
.lagune/skills/{regex,javascript,browser}.md (worked examples)	every *.md already in spec/skills/ (list the directory, mirror the closest one)
.lagune/templates/specialize-template.md	spec/templates/specialize-template.md
@.lagune/specializations.md (list what exists)	read spec/skills/ and SKILLS_CATALOG in src/hooks/skills/catalog.ts
.lagune/skills/<name>.md (the sub-skill file)	spec/skills/<name>.md
.lagune/skills.json (catalog entry)	a BuiltinSkillEntry in src/hooks/skills/catalog.ts

One path it names stays put: a hook reference inside the .md body keeps the ./.lagune/hooks/<x>.mjs form, because that body ships verbatim to the end user and must point at their scaffolded path, not the source.

Step 2: Decide the category (group)

Every built-in belongs to a category (group), which is how npx lagune add --skills <category> bundles it into a target project. Categories live in SKILL_GROUPS in src/hooks/skills/groups.ts.

• A language sub-skill goes under its language key (python under python, and so on).
• A sub-skill about an application-security risk OWASP tracks goes under owasp.
• Match the sub-skill to the category that actually describes it. If none fits, that is a signal a new category may be needed, adding a SkillGroup to SKILL_GROUPS (key, label, description). Ask the user before inventing one, do not force an unrelated module into owasp to avoid the question.

The end-user skills.json has no groups field by default, but the built-in catalog entry does. Set it.

Step 3: Write and register, then verify

A new built-in is not done when the .md exists. Touch each of these:

Required (a sub-skill is broken without these)

1. Write the authored module to spec/skills/<name>.md (Step 1). src/core/assets.ts reads spec/skills/ by directory listing, so the file loads automatically once it is on disk, with no code change.
2. Add the BuiltinSkillEntry { name, tags, groups } to src/hooks/skills/catalog.ts. Without this row the file scaffolds, but .lagune/specializations.md never lists it and add --skills cannot bundle it.
3. Add a SkillGroup to src/hooks/skills/groups.ts, but only when you introduced a new category.

Tests (the suite asserts the catalog and the docs stay in sync)

4. test/integration/skills/skills.test.ts iterates SKILLS_CATALOG (e.g. "never repeats the name inside its own tags"), so a new entry is exercised automatically. Check that no count-based or hardcoded-list assertion now fails.
5. test/e2e/init/skills-hook.test.ts and test/e2e/init/manage-cli.test.ts assert scaffold and listing behavior, sometimes by sub-skill name. Update any hardcoded expectation that should now include the new one.
6. test/integration/cli/manifest.test.ts and test/integration/cli/scaffold-groups.test.ts reference specific skill paths like .lagune/skills/regex.md. Update them only when your change alters what a given category scaffolds.
7. Run the full suite: npm test. Its pretest hook runs scripts/build.sh first, so the compiled lib/hooks/*.mjs the e2e tests exercise are rebuilt automatically, no separate build step needed. All green before you call it done.

Documentation (the built-in catalog is enumerated by hand in the website)

8. Add a row to the "set that ships with Lagune" table in website/docs/commands/skills.mdx (around line 48), giving the built-in its category and focus. This table is maintained by hand and drifts silently when a built-in is missing.
9. The example listing output in website/docs/hooks/skills.mdx (# => regex: ...) is illustrative. Update it only when it should showcase the new one.
10. Update README.md only when it enumerates built-ins.

Not touched

• src/core/skills-catalog.ts (the empty user catalog factory), src/hooks/skills/discover.ts (reads the user's .lagune/skills.json), and .lagune/skills.json itself: these are the end-user registry. A built-in never writes there.

Quick checklist

[ ] spec/skills/<name>.md            authored from spec/templates/specialize-template.md (reconciled on a refine)
[ ] src/hooks/skills/catalog.ts      BuiltinSkillEntry added { name, tags, groups }
[ ] src/hooks/skills/groups.ts       only if a new category was introduced
[ ] tests updated + npm test green
[ ] website/docs/commands/skills.mdx table row added

Summary to the user

As Step 8 of the command: state the name, tags, category, what it covers, and that the phases now load it on demand once shipped (end users import it with @.lagune/skills/<name>.md). Note it shadows any same-named entry on a refine. Suggest a commit such as feat: add specialization in `<name>` defense, matching the wording the existing built-ins use, as a suggestion, not a mandate.