Troubleshooting

// every warning documented honestly, every fix spelled out

Windows: "Windows protected your PC"

The public train requires Authenticode publisher and timestamp verification, but SmartScreen can still flag a certificate while it builds reputation. Click More info, verify the publisher matches the release notes, and proceed only then. Stop if the publisher is unknown.

macOS: "cannot be opened" / "is damaged"

A public release is required to carry a Developer ID signature and stapled Apple notarization ticket. This warning is therefore a trust failure, not an expected first-run step. Stop and report the release tag, asset name, Mac architecture, macOS version, and exact warning. Do not clear quarantine or use an unsigned-app override for a purported public release.

Updates

The built-in updater (System → Updates) supports Windows and both Mac architectures and verifies every update against a key baked into the app. This is independent of Authenticode, Developer ID signing, and notarization. The build contract is not installed-update evidence; report a failed update and use the matching current release installer when necessary.

Uninstalling

• Windows: Settings → Apps → Installed apps.
• macOS: quit StarNet, drag it from Applications to the Trash.

Behind a VPN or corporate proxy

The app window honors your system proxy, but the engine's own outbound connections — AI providers and messaging channels like Telegram or Discord — go out directly and do not route through a proxy. On a network that forces one, the UI and browser sign-ins can look perfectly healthy while runs and channel connects fail with a plain network error. There is no proxy setting yet; you need direct access to the provider and platform APIs. COPY DIAGNOSTICS (under any failed turn) shows whether a proxy is configured on your machine.

Live Voice is slow to start the first time

Expected, once. Hands-free Live Voice runs on a speech engine inside the app rather than a cloud service, so the first session downloads the open Whisper and Kokoro model weights and caches them in your user profile — the panel shows that download's progress instead of pretending to listen. Once cached, local listening needs no network. Speaking stays local while Kokoro is available; otherwise StarNet may use the keyless Microsoft Edge Read Aloud fallback for answer text. If the panel refuses before opening your microphone, neither the local packages nor a supported offline operating-system dictation fallback is available.

Still stuck?

Open an issue on GitHub or email androo.agi@gmail.com. We'd rather hear about a broken thing than have you assume it's supposed to be that way.