Repository navigation
0.4.0
A new EPUB core. Each item names its work item in
dev-docs/plans/2026-10-improve-tepub.md.
⬆️ Upgrading from 0.3
- Run
tepub extractonce on each book you are working on. Segments are now
found by a new rule and get new ids. Extraction carries every finished
translation and synthesised audio clip over to the new units by matching file
and text, after saving a timestamped copy of the state; it reports anything it
could not match, which is translated again. On a real 1,623-segment workspace
every translation carried over.
🐛 Fixed — the EPUB written back
- Translated books are valid EPUBs (WI-4.1). The writer copies the source
book entry for entry and replaces only the chapters it translated. The old
one, through ebooklib, rebuilt every chapter's head, moved every file and
rewrote the navigation; written back unchanged, 15 of 19 real books gained
epubcheck errors. Now none does, translated or not. - Translated chapters keep their title and stylesheet links (WI-4.2).
ebooklib's reading API returned chapters with an empty head. - SVG covers, MathML and
epub:switchsurvive (WI-4.2). Chapters are
parsed as XML; the HTML parser lowercased SVG names and crashed on prefixed
elements. - No text is translated twice or lost (WI-4.3). A paragraph inside a
blockquote used to be extracted twice; text beside block children, and lists
wrapped in<span>, are now covered. Across 10,583 chapters of a 19-book
corpus every piece of text is in exactly one unit, skipped on purpose (SVG,
MathML, code listings), or reported. - Footnotes, links and emphasis survive translation (WI-4.5). Paragraphs
with inline markup are translated as HTML, and each reply must keep its
links, anchors and images or it is retried once and then marked as an error.
Measured with TranslateGemma 12B into Chinese: 30 of 30 paragraphs kept their
markup. - Bilingual output is valid in EPUB 2 too (WI-4.6). Originals and
translations are marked withtepub-originalandtepub-translation
classes;data-langis added only in EPUB 3, and translated copies carry no
duplicate ids. - Both tables of contents are retitled in translated-only output, the EPUB 3
nav and the NCX, and publisher stylesheets are no longer edited.
🐛 Fixed — reading books
- ebooklib is no longer used (WI-4.7). Its navigation parser crashed on a
nav with landmarks only. The spine, contents and metadata are read directly. - A moved or renamed book is still recognised (WI-5.3). Workspaces record
the book's SHA-256 and accept the same content wherever it sits. resume,formatand the debug commands find the right workspace under
--work-dir(WI-5.3).
🐛 Fixed — audiobooks
- Files the table of contents leaves out are no longer dropped (WI-5.2).
- The chapter preview matches the book (WI-5.2).
- Statements are synthesised once per wording and voice (WI-5.2).
- Chinese sentences no longer split inside numbers such as 3.5 (WI-5.2).
🔧 Changed
- The web export keeps classes and language attributes (WI-5.4), so its
original/translation toggles work for EPUB 2 books. assembly.pyis split into modules with one job each (WI-5.1).
🧪 Development
- Fixture books and a translating epubcheck gate run in the default suite;
TEPUB_CORPUS_DIRruns the same gate over a folder of real books.