cos: Skill for Claude Code

Add an add-on (the framework-native way)

An add-on is an optional, self-contained vertical layered over the core board. The framework already exists — your job is to parametrise it with one manifest, not to invent a parallel one. Copy Nutrition & Chef (board/lib/addons.ts, mcp/nutrition-server/, board/app/api/nutrition/, tests/api-nutrition-gate.mjs) beat for beat. Read docs/architecture/addons.md first; it is the spec, this is the recipe.

WHEN this applies — and the tell that you're doing it wrong

Use this skill when you are adding a whole feature area (its own pages + routes + agent tools) that should ship disabled by default and not bloat the core.

The tell — STOP and make it an add-on if you catch yourself about to:

• create a new board/lib/<x>-store.ts or a new board/data/<x>.json → the data belongs on cases.json via DBShape, written through the existing mutate().
• hardcode a nav entry into the sidebar's core list → nav comes from the manifest's navItems, revealed only when the flag is on.
• add an /api/<x>/* route prefix that nothing in ADDON_REGISTRY claims → register it.
• mint a second persistence chokepoint, a second SSE channel, a second backup path → there is exactly one of each, and riding cases.json gives you all three for free.

Does NOT apply to: a tweak to a core surface (cases/events/reminders/priorities — those are always-on, never gated), or a pure agent-skill with no new data/nav/routes.

The mental model (memorise this)

One four-layer slice, gated by ONE flag. Writes close; reads stay open. A component is a STATE MACHINE — it never calls an LLM; the agent is the intelligence.

An add-on is a deterministic state machine: it persists state and exposes it via an API + an MCP server, and that is the whole contract. It does NOT call an LLM / the Anthropic API. Any generative step (a training plan, a summary, a drafted reply) is produced by the external agent (your operator skill, Claude Cowork) and written back through an MCP write tool / a POST; the add-on only validates, versions, attributes, stores, and serves it. Deterministic server-side compute (stats, projections, aggregations) is fine — only generative LLM inference is delegated. The sole exception in the whole repo is the vault MCP (mcp/vault-server), which embeds the Claude Agent SDK and is itself an agent. If you catch yourself reaching for @anthropic-ai/sdk (or any LLM client) inside an add-on route or MCP server, stop — that generative step belongs in the agent's context, not the component.