hassanzohdy/agent-kit/monorepos

Monorepos

Claude Code (and every other agent) discovers project skills only from the session-root skills directory — it does not recurse into subdirectories. So in a full-stack repo opened at the root, an inner project's backend/.claude/skills/ (or backend/skills/) is invisible. agent-kit closes that gap by aggregating each project's skills up into the root.

repo/                          ← you open your agent here; one .claude/skills/
├── package.json               ← agentKit.monorepo.projects: ["backend", "frontend"]
├── AGENTS.md
├── backend/
│   ├── package.json           ← backend's own agentKit.pick/omit
│   ├── skills/code-standards/SKILL.md
│   └── node_modules/@warlock.js/core/skills/…
└── frontend/
    ├── package.json
    ├── skills/forms/SKILL.md
    └── node_modules/@warlock.js/core/skills/…

npx agent-kit sync --projects backend,frontend

.claude/skills/
  backend-code-standards/        # backend's authored skill — prefixed with the project dir
  frontend-forms/                # frontend's authored skill
  warlock-js-core-use-cache/     # shared dependency — ONE copy, by package slug

The mental model

Each declared project is scanned as if you ran `agent-kit sync` inside it — its node_modules/ dependency skills (filtered by that project's own agentKit config) plus its authored skills/ folder — and the results are merged up into the root's skill directories.

Declaring projects

Two equivalent ways. Pin it in config so contributors don't retype it:

{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": { "monorepo": { "projects": ["backend", "frontend"] } }
}

…or pass it ad-hoc on the CLI (overrides the config):

npx agent-kit sync --projects backend,frontend
npx agent-kit sync --projects "apps/*"        # one-level glob: every dir under apps/

Entries are paths relative to the repo root (or absolute). A one-level glob (apps/*, packages/*, or bare *) expands to each immediate subdirectory; hidden dirs and node_modules are skipped. Multi-level / mid-path globs are intentionally unsupported.

--projects vs --path — they are different

Both pull skills from outside the root's own node_modules/, but they treat the directory differently:

--path backend would scan backend's children as packages (and miss backend/skills/ and backend/node_modules/). --projects backend is what you want for an actual project. They compose — you can pass both.

Authored skills are prefixed by the project directory

A project's own skills/ are exported with the project directory name as the slug prefix:

• backend/skills/code-standards → backend-code-standards
• frontend/skills/code-standards → frontend-code-standards

So two projects can ship a same-named skill without colliding. (Dependency skills keep their package-name slug, e.g. warlock-js-core-use-cache, and dedupe across projects.)

Per-project config is honored

Each project's node_modules/ skills are filtered by that project's own agentKit.pick/omit from its package.json. So backend can omit a noisy SDK that frontend doesn't even depend on:

// backend/package.json
{ "agentKit": { "omit": { "@some/noisy-sdk": true } } }

A project's own authored skills/ are always included — pick (a dependency allowlist) never drops them.

Precedence rules

• Shared dependency → one copy (union). If backend and frontend both depend on @warlock.js/core, its skills are exported once; the kept set is the union across projects (a skill survives if any project kept it). One copy keeps Claude's skill list clean and avoids near-duplicate routing noise.
• Root `omit` is the global veto. After aggregation, the root package.json's omit applies across everything — the root can ban a dependency repo-wide regardless of what a project kept.
• A project's `targets`/`monorepo` are ignored. Only its pick/omit matter, and aggregation never recurses into a project's own monorepo.projects.

Per-workspace AGENTS.md

Aggregation handles skills. For per-project instruction files (a CLAUDE.md tuned to each workspace), run the derive step inside each workspace — each gets its own AGENTS.md → CLAUDE.md:

npx agent-kit sync --cwd backend
npx agent-kit sync --cwd frontend

Editors open the closest *.md to the file being edited, so contributors in backend/ get backend's conventions automatically.

Watch mode across projects

npx agent-kit watch --projects backend,frontend

Watch resolves the real skill-source directories up front and watches them: AGENTS.md, the root skills/, each project's skills/, and each project's package.json (so a pick/omit edit re-syncs too). Dependency skills under node_modules/ are not watched — they only change on (re)install, which fires postinstall → sync.

How big a monorepo?

This is built for full-stack repos — a handful of sibling projects (backend, frontend, mobile, admin), not hundreds. Each project's full node_modules/ is scanned (in parallel), so a few projects is cheap; pointing it at dozens of workspaces will multiply the per-sync scan cost.