New "Paragraph Synthesis" TTS option (off by default). When enabled, replies are synthesized paragraph by paragraph (split at line breaks) instead of sentence by sentence — each paragraph starts playing as soon as it is complete, giving more natural prosody on both built-in and cloud TTS engines. Over-long paragraphs automatically fall back to sentence-level synthesis.
Improvements
Clearer voice-output settings. The "Dialogue Only" read-aloud mode is renamed to "Quoted Speech Only" in all 31 languages — characters whose replies contain no quotation marks no longer fall silent without explanation. The descriptions for Read Aloud, Bubble Display, Skip Action Descriptions, and Paragraph Synthesis are now always visible below each option instead of hidden in hover tooltips.
Fixes
Local AI chat no longer fails with a cryptic "HTTP 400" error when the conversation outgrows the model's context. The app now automatically trims older conversation history and retries once; if it still does not fit, a clear message in your language suggests shortening the character card or increasing the context size, instead of showing a raw technical error.
Cloud model errors now surface the real cause instead of a misleading quota message. When a web-search request is rejected and the plain retry uncovers a different underlying problem (such as an unsupported region or a network block), that underlying error is what you see; the original quota error remains available in the log.
Workshop pets no longer land in the wrong tab or refuse to load while offline. Subscribed Spine/VRM/MMD and other models previously appeared under the Live2D tab — and could not be switched to — whenever Steam was not running. Cached items now keep their real type and are filed under the correct tab, package manifests self-correct misclassified types, and switch failures record the exact reason in the log.