SpeakoFlow Files

Test build — not a real release. Unsigned, built from the branch test/vision-fallback-build for manual testing only. The app still reports its version as 1.4.0. Windows and Linux x86_64 only; there is no macOS or ARM build here on purpose.

What to test

The assistant no longer dies when the model can't see images. Point the assistant at a model with no vision support (a text-only local .gguf, or a text-only cloud model), then send it a screenshot or an attached picture.

Before: the request was rejected, you got a red banner and no reply, and because screen arming is sticky every following message hit the same wall — the conversation was finished.

Now: the turn retries once without the image, and you get a normal reply — spoken too, if TTS is on — that opens by telling you the model can't see images and points at Settings → Assistant. Screen vision switches itself off so the next message doesn't repeat the failure, and the conversation keeps working.

Worth checking:

• A vision-capable model still sees screenshots exactly as before.
• The reply doesn't invent what was in the image.
• Agent-decides screen mode: the model calls capture_screen, the frame is rejected, and you still get an answer instead of a generic provider error.
• The panel's screen toggle visibly turns off after the failure.

This branch also carries in-progress cloud speech-to-text and hands-free voice conversation work from the working tree, so those are worth a look while you're here — but they are unfinished, not the point of this build.

Install

• Windows — SpeakoFlow_1.4.0-test.1_x64-setup.exe (or the .msi). SmartScreen will warn about an unknown publisher: More info → Run anyway.
• Debian / Ubuntu 24.04+ — sudo apt install ./SpeakoFlow_1.4.0-test.1_amd64.deb
• Any other distro — chmod +x the AppImage and run it.

Installing this over an existing SpeakoFlow replaces it, and it shares the same settings store, since the version string is unchanged.