LifeOS v0.3.0

The release where setup stops being the hard part. First release with an installer asset: download LifeOS-Setup-0.3.0.exe, walk a six-step wizard, no terminal, proven end to end on clean Windows 11 (the wizard installs WebView2 itself where it is missing).

New

• Windows installer + setup wizard: model discovery and testing, optional voice model download (pinned hashes, resumable), microphone test with live device re-detection, theme, phone access.
• Phone access via Tailscale: one switch gives your own devices a private HTTPS address; scan a QR and the full app (recording included) runs in the phone browser. Loopback-only otherwise; nothing is opened to the LAN or published to the internet.
• Two measured model profiles (93-case battery x3, records in the repo): Gemma 4 12B QAT — fits a single 12 GB GPU, 87/93, recommended; Qwen3.8 27B Q8 — 90/93, best quality. Pinned launch lines in docs/setup.md.
• Extraction presence guard: a required value the transcript never said cannot auto-execute, on any model — closed in validation, not in the prompt.
• Speech hardware panel: CPU/GPU choice gated by a probe that proves the GPU can actually compute (not merely open a model), one-click pinned CUDA runtime acquisition, idle-unload presets, bring-your-own whisper model.

Fixed

• Deterministic weekly-recurrence resolution (date math left the model).
• First boots on slow machines: patient server start, resuming retries, honest "still starting" copy, and a virgin-machine config/migration defect that four clean-VM rounds shook out.
• The setup can be left on purpose and only on purpose; taskbar icon identity; a dialog painted under the mobile tab bar; queue copy that blamed the wrong component.

Checksum (SHA-256)
b884922c65abef657f0e58eeae1e8527a1f45853340558bff3397687db434cc7

LifeOS v0.2.2

Fixed

• Clarification cards now keep the original question when you answer with voice or text, so a short answer such as "75" continues the pending note instead of becoming a separate note.
• Weight goals can now be referenced by their target, such as "the one for 85kg", without repeating the full goal name.

LifeOS v0.2.1

Fixed

• Category chips in expense and budget confirmation cards now select the category that was tapped.
• Record a new note now opens the voice recorder from Today and from inside the record sheet.

Changed

• Version advanced to 0.2.1.

v0.2.0: Barcode food logging

Everything since v0.1.0. Details in CHANGELOG.md.

New

• Barcode scanning: point the camera at a packaged food and log the meal from the label
• Open Food Facts lookup for unknown barcodes: server-side, carries the barcode and nothing else, off switch in Settings and a veto in config
• A second scan of the same product resolves from your own presets with no network request
• "Unspent this month" on the Money tab: logged income minus logged spend and set-asides
• A speech model picker in Settings that offers only models on the server's disk and never downloads
• The licence and a source link in Settings (AGPL section 13)
• The attach question shows your saved figure and date beside the scanned label's

Fixed

• The scan sheet could be covered by the phone's tab bar; it now draws above everything
• Changing the speech model on a running server was silently ignored until a restart
• A speech model name written through the API could trigger a download; models not on disk are refused
• The Settings page could name two different speech models at once
• Preset rows crushed their name when the delete confirmation appeared

Changed

• The scan sheet is redesigned: found products render as a card with an energy and serving row, a viewfinder that names which barcode reader is running, and every failure states itself in a sentence
• Empty states space their action buttons apart