Skills

Skills are Kenzy's abilities — checking the weather, reading the news, controlling your smart home, announcing to other rooms. A good set ships built in and works out of the box; the dashboard's Skills tab shows what's loaded and lets you switch any of them on or off with a click. If you can write a little Python, adding your own is a single small file.

Technically, each skill is an async Python function decorated with @skill. The language model calls skills as tools, deciding when to use each one from its docstring and type signature.

How skills work

Skills load from two places at LLM-service startup:

1. Built-in skills ship inside the package (kenzy/llm/builtin_skills/) and always load first.
2. Your overlay — the directory named by skills.dir in llm.yaml (default skills/ under your config home: ~/.config/kenzy/skills, or the repo root in a dev checkout). Loaded second, so a file here that defines a skill of the same name overrides the built-in one.

Every .py file is imported and all @skill-decorated functions are registered. The decorator introspects each function's signature, type annotations, and docstring to generate the JSON Schema tool definition that LiteLLM passes to the model — no separate configuration or registration step. Disable any skill (built-in or custom) by its function name under skills.disabled in llm.yaml.

kenzy/llm/builtin_skills/   ← bundled: weather, news, home_assistant, stocks, …
~/.config/kenzy/skills/
  my_skill.py               ← your custom skill (overrides a built-in of the same name)

The LLM decides which skills to call and with what arguments based entirely on the docstrings. Well-written docstrings that describe when to use a skill are as important as the skill's implementation.

Two resolution tiers

Kenzy resolves a request in two stages, fastest first:

1. Fast path (deterministic). Before the LLM is consulted, the request is run through any registered fast intents — @fast_intent matchers that recognise common, high-frequency commands locally and answer instantly with no remote model call. Time/date queries and the bulk of Home Assistant control ("turn on the kitchen lights") are handled here.
2. LLM fallback. Anything a fast intent doesn't confidently match falls through to the tool-calling LLM exactly as before. The fast layer is deliberately high-precision: when it isn't sure, it defers.

A skill can expose both — a deterministic @fast_intent front-end for the easy cases and an @skill tool definition for the LLM to fall back on. See Writing Skills for the fast-intent API.

Skill configuration

Per-skill settings live under skills.<skill_name> in llm.yaml:

skills:
  home_assistant:
    url: "http://homeassistant.local:8123"

  news:
    max_headlines: 5

Skills read these values at runtime via get_config(section, key, default):

from kenzy.llm.skills import get_config

url = get_config("home_assistant", "url", "http://homeassistant.local:8123")

Disabling skills

List skill function names under skills.disabled in llm.yaml to disable them without deleting the file:

skills:
  disabled:
    - get_stock_info
    - yes_no_maybe

Secrets

Skills read API keys from environment variables (.env). Never hardcode secrets in skill files.

Included skills

See Built-in Skills for documentation on all skills that ship with Kenzy.

Writing a skill

See Writing Skills for a step-by-step guide.