Yesterday I shipped Local Waifu 1.7.0 with more cloud-provider choices, safer cleanup for local processes, and a handful of stability improvements.
Today I shipped 1.7.1.
It contains no new AI feature. No new model. No prettier UI. It is a repair release for a mistake that made the app look broken immediately after an in-app update.
Some Mac users updated successfully, opened the app, and found that image generation no longer worked. The app said the local image engine was not bundled in that build.
In some cases, local chat replies were empty too.
The reason was not the models, the GPU, or a complicated inference bug. The updater had installed an incomplete app.
Here is the embarrassing detail: the normal macOS installer is 245 MB. The broken in-app update was 70 MB.
The update package had been assembled before the local engines were copied into the finished app bundle. So it contained the interface, settings, and most of the app — but not the pieces that actually run local chat and image generation.
That also created a particularly bad dead end in Settings:
the image engine appeared as “not installed”;
the button to download a model was hidden because the engine was missing;
so the user had no obvious path to fix it themselves.
Installing from the .dmg was never affected. Neither were characters, chats, licences, or downloaded models. Windows was unaffected too.
But none of that changes the first experience after a broken update: you open an AI companion app, send a message, and get an empty reply. Or you try image generation and the app tells you that its own engine does not exist.
At that moment, it does not matter how many provider options I added in 1.7.0. The product has already failed its most basic promise.
So 1.7.1 does one thing: it repairs the installation automatically.
More importantly, I changed the packaging flow. The update is now assembled from the finished application bundle, not from an earlier build stage. The build also stops with an error if any required local engine is missing.
I spend a lot of time thinking about features in a local AI product: memory, voices, models, images, privacy, and all the interesting visible things.
This release was a reminder that the feature users notice most is often much less exciting:
does the thing still work after I press Update?