shabbos-goy

A Hebrew-speaking, speech-to-speech household agent that helps observant Jews on Shabbat and Yom Kippur without breaking them. It never takes direct commands. It only infers intent from indirect remarks, e.g. "הלוואי שהיה קר" ("I wish it was cold") → turn on the AC.

About the name. Shabbos goy is the familiar Yiddish term for the traditional role of a non-Jew who helps a Jewish household on Shabbat, often in response to a hint rather than a request. The name describes that role; it is not meant as a slur.

No rabbinic approval. This project makes no claim of rabbinic approval (hechsher) and does not decide halacha. Whether and how a device like this may be used on Shabbat or Yom Kippur is an open question. Ask your own rav. The permitted actions live in configuration so a community can narrow them to match its posek.

Status

Early scaffold. The repository was provisioned from culture-agent-template and so far contains only the agent-first CLI baseline (identity, learn, explain, overview, doctor). The build brief is issue #1. Everything under How it will work is planned.

How it will work (planned)

The one rule: hints, never commands

• No wake word and no confirmation questions. Both would turn the exchange into a command. When unsure, it does nothing.
• Commands are dropped, never queued, including across restarts.
• It may speak a short, neutral Hebrew remark ("המזגן פועל", "the AC is on") that does not invite a reply.
• The rule is enforced by a tested Hebrew utterance classifier in this repo, not by a prompt. Its headline metric is how rarely it acts on a command.

Pipeline

microphone → lobes (local Hebrew ASR, ears-only) → transcript
          → classifier → Shabbat/Yom Kippur calendar gate (zmanim)
          → whitelisted tool call (AC via sensibo-cli) → optional neutral remark (TTS)

• Speech runs locally on the lobes Hebrew realtime stack (ivrit.ai Whisper, local TTS). Audio stays on the device and is never recorded. Logs keep only the classified intent and the action taken.
• Climate control via tool calling. Actions are declared as tools (function name + JSON-schema arguments). The first tool backend is sensibo-cli for Sensibo smart-AC control. Every tool call, whether it comes from a rule or from a model, passes through the classifier, the calendar gate and the configured whitelist (devices, modes, temperature range) before anything actuates. Sensibo is a cloud service, so AC control needs internet access.
• Calendar-aware. Shabbat/Yom Kippur mode turns on and off automatically from zmanim for a configured location. Everything is configured before Shabbat, so nothing has to be toggled during it.

Deployment: Docker, survives reboots

It will run as a Docker Compose service on the machine that owns the microphone (currently a DGX Spark), with restart: unless-stopped and the Docker daemon enabled at boot. After a power cut it comes back by itself and works out the current mode from the clock, with no one touching it. Secrets (the Sensibo API key, the lobes gateway key) and private config (location, whitelist, device ids) stay outside the repository: a gitignored env file and a read-only mount of $XDG_CONFIG_HOME/shabbos-goy on the host.

Quickstart (what exists today)

uv sync
uv run pytest -n auto                 # run the test suite
uv run shabbos-goy whoami             # identity from culture.yaml
uv run shabbos-goy learn              # self-teaching prompt (add --json)
uv run teken cli doctor . --strict    # the agent-first rubric gate CI runs

CLI

Every command supports --json. Results go to stdout and errors/diagnostics to stderr (never mixed). Exit codes: 0 success, 1 user error, 2 environment error, 3+ reserved. Any verb that actuates is dry-run by default, and --apply actuates.

Agent harnesses

This is an AgentCulture mesh agent. culture.yaml declares backend: claude, so CLAUDE.md is the mesh resident's prompt and the fullest write-up of the repo's design and conventions. Read it first.

Four harnesses can work in this repo interactively, each reading its own file. There is intentionally no shared AGENTS.md:

Which binary you run chooses the interactive harness. culture.yaml's backend chooses only the mesh resident. See docs/harness-selection.md.

License

Apache 2.0. See LICENSE.