Your Own AI v0.5.0

Your health questions, your choice of model

• Health questions are answered on your device by default, and you now
  choose which of your downloaded models answers them - Settings >
  Routing. Downloading a medical specialist offers the switch once.
• If you prefer an online model for health questions, that is your
  choice too - and the app asks before each question leaves your
  device, every time, unless you tell it to stop asking. Health images
  never go online.

Agent permissions: one simple choice

• Your AI can run ordinary project work without asking. One
  three-level choice: Ask every time, Auto - ordinary work inside the
  project folder runs unasked while risky, irreversible, or
  outside-the-folder actions still ask - or Approve everything.
  Settings > Agent sets the default; the project chip in the header
  switches one project live, and a project's own choice sticks.
• Whatever you choose, every action goes in your records: automatic
  allows show as receipts on the rail, and each turn keeps a ledger of
  everything the agent was allowed to do. Permission answers record
  the decision, its scope, and how it was answered.
• Needs the Your Own AI Build update offered in the folder menu and in
  Settings > Components.

Projects: the working session, reworked front to back

Working in a project with your AI got the deepest pass of this
release. The goal, in one line: you always know what your AI is doing,
and nothing it does makes you wait.

• The rail always shows life. The activity indicator sits at the
  true tip of the turn, so a long think after a spoken paragraph never
  looks dead. A conclusion your AI speaks mid-turn is never buried
  behind silent steps - it becomes the answer. A finished turn's
  summary says when its background task is still running, and the rail
  shows what that task is doing while the agent waits on it.
• Nothing holds up your turn. Remembering something for a project
  happens in the background, project memory opens instantly with
  cached and indexed reads, and reading the project's own memory no
  longer asks permission.
• The project menu is rebuilt: the project's name and live status
  in plain words, permissions as one control with its meaning always
  visible, and memory and close as clear tiles. A stopped project can
  be reopened right there.
• Web, on your terms. The Build agent's web search works through
  the AI's own routing - an offline-only AI stays off the web in
  project sessions, and says so. A silent online call is named for
  what it is: waiting on the provider.
• Opening a project conversation lands on its last message, and its
  history is whole - a conclusion spoken before silent steps shows as
  the answer even for conversations recorded before this release.

Your records, fast and honest

• The conversations drawer opens instantly. Every AI's conversation
  list is kept as an encrypted local copy that shows immediately - even
  right after launch - while live reads refresh it in the background. A
  slow read never blanks what you can see, and deleting a conversation
  disappears from the list at once.
• An AI now always finds and keeps its existing records after launch -
  previously a slow start could leave an AI beginning a fresh set,
  with its earlier conversations present but out of view.
• Opening a conversation lands on its last message, names the
  conversation it is opening, and if your records don't answer in
  time it says so - nothing is lost - instead of showing an empty
  chat.
• Every memory surface says your records are warming up after launch
  instead of claiming "nothing yet" about records that exist.
• Loading spinners no longer slow the app down on machines where
  animations are expensive - pages stay responsive through the whole
  warmup.

Models

• The offline models page opens with "Best for this computer" - one
  pick per activity (coding with you, everyday chat, seeing images,
  health questions), each chosen from what your machine can actually
  run, with download in place. The health pick can become your
  health-questions model in one tap. Downloaded models are now a
  compact list with your disk total, collapsible out of the way.
• Every model card says who made the weights - the maker, who
  packaged the files, and for community builds, what they are based on
  and how they differ.
• Recorded conversations now carry a fingerprint of the exact model
  file that answered, alongside its name - proof of which weights
  produced a response, sealed in your records.
• New in the offline catalog: Ornith 1.5 (the 35B now sees images),
  the Qwen 3.8 Distilled family (2B, 4B, and 9B - strong small models
  for modest hardware), Qwythos v2 (the repetition fixes), and Qwen
  3.8 Uncensored (27B, with the vision add-on).
• New online: GLM-5.3.
• Fixed: the Qwen 3.8 27B download had stopped working after its
  publisher renamed the file.
• Online chat turns send a per-conversation cache key, so consecutive
  turns reuse the provider's cache and bill at the cached-input rate;
  Online Models shows that rate.

App updates

• The app can now tell you when a new version is out - a quiet note on
  the home screen linking to the download page, once per release.
  Nothing about you or your machine is sent; it reads one public file.
  Turn it off in Settings > Help & diagnostics.

Fixes

• The hero "Continue" line shows its state the moment you click - a
  spinner while it opens, and the records-warming message when your
  records are still loading after launch.
• Long command lines in a turn's summary stay inside the window.
• The working rail's thinking indicator no longer occasionally doubles.
• Embedding inputs are capped to the model's context so one long query
  cannot fail a whole recall batch.