home-companion 0.2.0-rc1

First release candidate of 0.2. It runs on the author's device; a few voice features are still being tried before 0.2.0.

New

• The house from the screen and by voice: a command that went through gets a card with Undo (lights, covers and whole rooms are put back as they were), a bow from the character and a sound. Covers ask first ("Close the blinds?") with Yes/No.
• Do not disturb from the control center or by voice ("do not disturb for an hour", "good night"): no card sounds or announcements, a bell asleep at the character's feet; timers and alarms still ring.
• Named timers ("a 5 minute timer for the pasta"), "what timers do I have", and a timer that says its name when it rings. The last 10 seconds tick.
• "Repeat" says the last answer again. Reflex answers are spoken in Home Assistant's voice (cached, also offline).
• Home Assistant page: choose the AI, speech-to-text, voice and timbre on the screen; "Check and repair" sets up the rest. With OpenAI, the voice and the transcription are told the language through their own options.
• AI use: today's and this month's cost with a 30-day chart, precise with the OpenAI with Token & Cost Tracking integration, an estimate otherwise.
• Cards from Home Assistant, calendars and lists (Bring! or any todo list) chosen on the screen, a weather page, "What to say", "show me …" opens a Home Assistant view.
• Microphone switch on top of the device is read and shown.
• Wi-Fi signal next to the clock, one pop-up style, the character on a third of the screen.

Fixed

• Large Home Assistant installs: the setup failed with "websocket: frame of … bytes" when the entity registry was over 4 MiB (#1). The limit is now 32 MiB and only the enabled entities are kept.
• On an image without the Home Assistant view (WEB=0), the HA button opens the device's Home Assistant page instead of doing nothing (#1).
• Stop words no longer fire when the character itself says a word that sounds like one.
• A made-up entity id from the AI finds the entity by its name.
• The written answer shows before its voice arrives.

Install

On a ThinkSmart View running thinksmart-view-debian:

curl -LO https://github.com/m0s4ik/home-companion/releases/download/v0.2.0-rc1/home-companion-0.2.0-rc1-arm64.tar.gz
tar xzf home-companion-0.2.0-rc1-arm64.tar.gz
cd home-companion-0.2.0-rc1-arm64 && ./install.sh

An installed 0.1 is updated in place; its settings stay.

Alfred: source package (v1)

The source material of Alfred, the butler character, for anyone who wants to redraw, re-animate or improve him (even just the drawings). Everything is dedicated to the public domain (CC0 1.0).

Not a release of the program: the device keeps using the sheets in characters/alfred/.

Inside alfred-sources-v1.tar.gz (94 MB):

• clips/: the final frames of every animation (440x760, 16 fps), exactly what the device plays: idle, listening, thinking, speaking, error, the wake-up (bow tie and glove), and the night clips (falling asleep, asleep, waking up).
• raw/: the video model's original 832x832 output, before cropping.
• comfyui/: the ComfyUI workflow of each clip.
• poses/: the still pictures each clip starts and ends on.
• scripts/: generation through ComfyUI, cropping and cleaning, joining clips.
• POSES.md: the gpt-image-2 prompts of the poses, and the rules a new drawing must keep.
• RECIPE.md: models and licences (Wan 2.2 I2V A14B with the lightx2v 4-step LoRA, all Apache-2.0), parameters, the prompt of every clip, how the loops close.
• README.md: how the clips map to the files in the repository, and how to bring a change into the device.

Check the download with sha256sum -c alfred-sources-v1.tar.gz.sha256.

home-companion 0.1.0

What it does (details and screenshots in the README):

• "Okay Alfred" heard on the device; a Wyoming satellite for Home Assistant's own voice pipeline and AI agent.
• Local reflexes (time, timers, alarms, volume, silence) that work without Home Assistant or the network.
• Timers, alarms and reminders kept on the device, with full-screen notices.
• An animated character, Alfred, with a day and a night pose.
• Home Assistant's own frontend on the device, opened from the home or with a swipe from the edge.
• AirPlay: music with cover and controls, and screen mirroring from an iPhone or iPad (UxPlay).
• First start from the screen: language → Home Assistant found on the network → QR code → automatic setup
  (agent, pipeline, MCP tools), checked and repaired from the device.

Needs a Lenovo ThinkSmart View with thinksmart-view-debian
v0.3-rc3 or later (WEB=1 for the Home Assistant view, AIRPLAY=1 for AirPlay).

Install, on the device:

curl -LO https://github.com/m0s4ik/home-companion/releases/download/v0.1.0/home-companion-0.1.0-arm64.tar.gz
tar xzf home-companion-0.1.0-arm64.tar.gz
cd home-companion-0.1.0-arm64 && ./install.sh

Check the download with SHA256SUMS. Options and manual steps: packaging/README.md.

Known limits: AirPlay pause is heard 2-3 s late; Wi-Fi is set in the system image (or with nmcli), not
from the screen yet; the "basta"/"stop" stop-word models were trained with Piper voices: see their notes in
characters/alfred/stopword/.