Skip to content

0.4.5

Choose a tag to compare

View all tags
@fukuyori fukuyori released this 12 Jul 10:02
· 17 commits to main since this release
0.4.5
9df88c0

[0.4.5] - 2026-07-12

Added

  • Bare-text chapters are now translatable. Some EPUBs (often converted
    ones, e.g. Aozora/Gutenberg conversions) carry chapter prose as bare text
    directly inside <div>/<body>, separated by <br/><br/> instead of
    <p> elements. Such text had no block structure, so translation,
    highlighting, search jumps, and Kindle-note matching only saw the headings.
    Chapters are now normalized on load: each run of bare inline text becomes a
    synthetic paragraph (split at 2+ consecutive <br>, spacing preserved),
    consistently across the reading view, sidecar caches, and translated-EPUB
    export.
  • Wrong-language translations are caught. When the model answers in a
    language whose script clearly doesn't match the target (e.g. Chinese
    returned for a Japanese target), the request is retried once with a
    stronger instruction; if it still doesn't comply the paragraph shows an
    error instead of caching the wrong text.

Fixed

  • Chinese ⇄ Japanese translation reliability. The translation prompt
    told the model to return text "already in the target language" unchanged,
    which made kanji-heavy Japanese pass through untranslated with a Chinese
    target (and vice versa) — and the untranslated result was then cached.
    The prompt now states the task in the user turn in a form that
    translation-tuned models (e.g. TranslateGemma) also follow, and no longer
    offers the pass-through escape hatch.
  • Reasoning models' chain of thought leaked into translations. Content
    ending with a </think> marker (e.g. Qwen3 with thinking nominally
    disabled) now has everything before the marker stripped.
Assets 5
Loading