Agent skills

One skill per module

Run:

home skill install

home writes one focused skill for every module. The skill describes when to use that module, how to configure it, the available command tree, the JSON contract, and the safety requirements for mutations.

Skills are generated from the same ModuleManifest objects that build the runtime CLI. A command cannot land in one surface while silently missing from the other.

Why the split matters

A single giant homelab skill would force every interaction to carry irrelevant context. Focused skills let an agent load only the service boundary it needs:

• schedule questions load gcal
• network and client questions load unifi
• playback requests load spotify and sonos
• operational briefings load beszel and uptime-kuma

Machine-readable output

Every command accepts --json. Success and error payloads use stable shapes and codes, so an agent can distinguish a bad argument, an unreachable service, and an unconfigured module.

home gcal agenda --days 2 --json
home github prs list --json
home beszel alerts list --json

Effect classification

Every command declares one effect:

The live E2E harness can automatically exercise reads, runs writes only through snapshot/mutate/assert/restore scenarios, and refuses destructive commands at the process-spawn boundary.

Credentials

Agent prompts never need to contain service credentials. The CLI resolves secrets from the local backend and passes only the normalized result back to the caller.

See Configuration and secrets for storage and migration details.