Repository navigation
Unicode \u202f breaks the next pages of template #627
Description
Describe the bug
I recently develop auto document generation using gpt-oss:20b model. Unexpectedly, the model gives output like 5\u202fseconds and this results a bug to the template where all next pages disappeared.
Expected behavior
I think we should normalize the problematic unicode before rendering.
Screenshots
Activity
CAOShurong commentedon Aug 13, 2026
I could not reproduce this yet against a wheel built from current
master(1f143fbe86c19ecb28c3205d5f4b1547c7e2d7ad).I created a two-page DOCX template with the value on page 1 and a fixed
SECOND_PAGE_MARKERafter an explicit page break. I tested both forms that the report might mean:"5\u202fseconds"r"5\u202fseconds"After rendering, both output packages reopened successfully with
python-docx; both retained all three paragraphs (value, page-break paragraph, second-page marker), and the marker remained inword/document.xml. The outputs were 36,683 and 36,681 bytes respectively. I did not render them through Microsoft Word, so this does not rule out a layout interaction in the original template.Could you attach a minimal source template plus the broken rendered output, and show
repr(the_value_passed_to_render)? That would distinguish an actual U+202F code point from an escaped string and let us test the specific page/section XML that disappears.OpenAI Codex assisted with the reproduction script and package inspection; I verified the run and results before posting.