Your Own AI v0.2.0

Changed

• Projects are dramatically faster. Model metadata is now cached
  instead of re-read on every step - agent steps that took close to a
  minute on modest hardware now land in seconds, and opening a project
  is near-instant.
• GPT-5.6 Sol drives project work by default, and your own model
  picks from Settings now apply to project sessions too (they were
  silently ignored before).
• Model choices respect your hardware honestly. Recommendations
  count the memory that's actually free, integrated graphics is sized
  as what it is (with a truthful "Integrated graphics" card badge), and
  automatic picks prefer a model that runs comfortably over a smarter
  one that barely loads.
• Privacy-first routing means what it says. With the online lean
  set to privacy-first, hard questions stay on your device and only
  genuine live-web needs go online.
• Gemma 4 downloads use Google's official builds carrying their July
  refresh (better tool use, wider vision); the larger variants now
  show their true 256K context. Existing Gemma downloads keep working -
  re-download to get the refresh.
• The app opens at a roomier default size, the Your AIs cards say
  "Edit" where before there was only an icon, and every
  project-readiness notice says exactly what to look for offline: a
  model marked "Agentic" on the Offline Models page.

Added

• See your plan usage in Settings - spend against your monthly
  allowance, with an optional live ticker in the header (off by
  default).
• System Information with "Copy for support" - everything that
  determines model fit, copyable in one click, with nothing personal
  included.
• Project sessions warn before they can't work: opening a folder
  with an offline-only AI now tells you up front if no downloaded
  model can drive project work, or if the capable one won't fit
  comfortably on your hardware.
• Only one copy of the app runs at a time - launching it again
  focuses the open window instead of starting a second instance that
  would fight the first over your models and graphics memory.
• Older log files are kept when the log rotates, so a problem's
  history survives an app restart.

Fixed

• Editing your selected AI now applies immediately. Before, an
  open chat kept the AI's previous settings until you switched away -
  including its online/offline mode, which could route a message
  online after you had chosen Offline Only. Fixed, and this class of
  routing promise is now verified automatically across nearly two
  thousand setting combinations before each release.
• A model that crashed the app while loading is never automatically
  loaded again on the next start - no more crash loops that required
  deleting files by hand.
• A model too slow to load says so clearly and isn't retried that
  session, instead of failing with a raw error code.
• Models whose chat format can't support project work are no longer
  offered for it, whatever their name suggests.
• The "permission needed" button scrolls to the permission card
  instead of past it.