Skip to content

Utsuwa 0.21.3

Choose a tag to compare

@github-actions github-actions released this 03 Nov 16:18
0069f22

A bug-hunting release: we audited chat, voice, memory, providers and the UI and fixed what turned up.

Fixed

  • Voice: streaming speech (OmniVoice) repeated the last sentence or two of a reply when the model paused while writing it (#320, closes #319).
  • Voice: sending a message while she was still talking could play two sentences over each other, and lip sync stopped for the rest of the reply (#324).
  • Voice: a lone < or { in a reply, as in "if x < 5" or ":-{", silenced everything after it. Decimals, version numbers and web addresses like "3.5" or "utsuwa.ai" were spoken as separate pieces (#324).
  • Voice: text written just before a reply was cut off mid-sentence went unspoken; Stop didn't stop streaming speech once the text had finished; a quick tap of push-to-talk could leave the microphone recording; and browser voice input always listened for English. It now follows your browser's language (#324).
  • Chat: when a provider reported an error in the middle of a reply (Anthropic's "overloaded", OpenRouter error chunks), she answered "Hmm... thinking", and that stand-in was saved, spoken and counted toward your stats. Temporary errors are now retried quietly, other errors return your message to the composer, and an empty reply shows an error (#326).
  • Chat: thinking models on Ollama and OpenRouter in the desktop app were cut off with "Reply timed out" while they were still thinking (#326).
  • Chat: bulleted lists and multiplication vanished from replies, so "6 * 7 = 42" became "6 5". Chat bubbles also lost their line breaks, so lists ran together on one line (#326, #327).
  • Chat: a voice message or typed message sent while she was still replying, or while chat was turned off, was thrown away. It now goes back into the message box (#326).
  • Chat: pressing Stop after the reply arrived left the raw reply text in the bubble, and a "Run tool?" prompt from a stopped turn could still run the tool later. The "retrying" notice now clears as soon as the reply arrives (#326, #327).
  • Memory: an updated fact ("I'm 26 now", "I moved to Osaka") was treated as a duplicate of the old one and dropped, and a memory added by hand in the memory inspector could disappear. The newer version now replaces the old one (#322).
  • Memory: half-sentences like "User said: I'm not" were saved as memories when the model had nothing to remember (#322).
  • Events: the "Return After Absence" and "Missed You" moments could never happen, and a moment you closed without choosing reopened after every message (#322).
  • Relationship: affection and trust kept fading during time away in Companion Mode, where they can't be earned back, so switching back to Dating Sim could drop your stage (#322).
  • Sending several photos in one message kept only one of them on the photo board (#322).
  • Merging a backup into an existing companion could permanently block one-time story moments (#322).
  • A day streak froze at 1 after one visit with the device clock set far ahead (#322).
  • Providers: OpenAI-Compatible endpoints whose address already includes a version path (Z.ai .../api/paas/v4, GitHub Models .../inference, Gemini .../v1beta/openai) got /v1 added and failed. An Ollama address ending in /v1 showed no models (#321).
  • Providers: the OpenAI model list hid o1, o3 and o4-mini but showed image, speech and transcription models that can't chat. Refreshing a custom provider's list replaced the model name you'd typed, and a model list from an unreachable server loaded forever with an empty error (#321).
  • Settings: text you were typing in Settings > Character was wiped when a reply finished in the background (#323).
  • Photo mode: the camera limits never switched on, so you could orbit under the floor (#323).
  • Switching avatars quickly could save one avatar's thumbnail under another, and the desktop overlay leaked memory each time a new custom avatar was picked (#323).

Changed

  • OpenAI-Compatible base URLs: Utsuwa now adds /v1 only when you enter just the host (like http://192.168.1.5:8000). An address that already has a path is used as is. If you entered a partial path that relied on the old behavior, such as https://api.groq.com/openai, add the /v1 yourself (#321).
  • Security updates: the desktop updater moves to 2.14.1, which clears a critical advisory in its zip library, along with rustls, time and other dependency fixes. 14 of 32 open advisories are cleared. The rest have no compatible fix yet; the pull request lists each one and why the risk is low (#325).