LumenPDF

Handoff / 接手先讀

1. STATUS.md：現在、下一個動作、卡住／依賴。
2. AGENTS.md：這個專案的規則。
3. pipeline/PIPELINE.md：步驟與狀態。
4. Latest logs/conversations/ record.

A PDF viewer and editor built for three things, in this order: speed, restraint, and craft.

Most PDF software picks one. SumatraPDF is instant but cannot edit and looks like 2009. Acrobat can do everything and takes eight seconds to show you a page. Stirling-PDF has fifty tools and no coherent interface. LumenPDF is an attempt at all three at once.

Status: working editor. Opens, renders, prints, searches, selects, annotates, redacts, fills forms, recognises scanned text, and saves.

What works today

Reading — open, render, scroll, zoom, fit-width, page thumbnails, document outline with collapsible sections, full-text search (245 pages and 1028 matches in 130 ms), match highlighting with a strong marker on the current hit. It reopens each document where you stopped reading, and remembers the last dozen files you opened.

Printing — page ranges, copies, greyscale, fit-to-page or actual size, and printing to a PDF file that keeps the source page size. Pages are rasterised through the same engine that draws them on screen, so annotations, filled form fields and OCR layers all come out — rather than depending on the printer's own PDF interpreter.

Links — internal jumps follow on click. External links never open without showing you the actual target first, because what a PDF prints and where its link points are unrelated strings. Only http, https and mailto are ever handed to the shell; a Launch action naming a program is not offered as a link at all.

Encrypted documents — a password-protected PDF asks for its password instead of refusing to open. The password is used once and never written anywhere: this application has no keystore of its own, and storing someone else's document password without one is not a trade worth making.

Your language — the interface is available in English, 繁體中文, 简体中文 and 日本語, following the system locale by default. Only fully translated languages are offered; a partial translation produces a window in two languages and no error anywhere. Adding one is a single JSON file — see translations/README.md.

Selecting — click-drag and double-click selection that follows the real text flow, including across columns and across pages. Copy to clipboard.

Every script, not just Latin — text extraction is where PDF tools quietly go wrong, and each writing system fails differently: right-to-left order in Arabic and Hebrew, stacked tone marks in Vietnamese, clusters in Devanagari and Thai, scripts written without spaces. Ten writing systems are regression-tested on every change.

Chinese alone had four real bugs, all now fixed and held by tests: extraction returned Kangxi-radical lookalikes so copied text matched nothing when pasted; searching for a single ideograph was blocked by a minimum query length; a one-character selection read as empty; and Latin word rules swallowed whole runs of Han on a double-click.

OCR — scanned pages get an invisible text layer written underneath the image, so the scan looks identical and becomes searchable, selectable and copyable. Backed by Windows' own recogniser: offline, no extra download, and it uses whichever languages your system already has.

Google Drive — open from Drive, edit, and save back to the same file, keeping its ID and sharing. Scoped to drive.file, so it can only touch files you explicitly open or that LumenPDF created. Requires your own OAuth client — see docs/GOOGLE-DRIVE-SETUP.md for why that cannot be shipped for you, and how to get one in a few minutes.

Editing — highlight, underline and strike-through in four colours, applied from a floating bar that appears over the selection. Rotate, reorder and delete pages. Append another PDF, or export any page as its own file. Sign by drawing, stamped as vector paths rather than a bitmap. Everything that modifies the document is undoable. Save, with the annotations written as standards-conformant PDF markup that other PDF software reads.

The signature is an ink stamp, not a cryptographic digital signature. It does not identify who signed and does not break if the document is altered afterwards. If you need a legally recognised signature, this is not it.

Redaction that actually redacts — the selected text is destroyed, not covered. The affected page is rasterised and its objects replaced, so there is nothing left underneath to recover. That costs the page its selectable text, and the confirmation dialog says so rather than hiding it. Verified by extracting text from the saved file, not by looking at the black box.

Text correction — press Ctrl+E, click a run of text, and edit it in place. Be aware of what this is and is not.

PDFium can replace a text object's string but cannot split one, and most PDFs embed fonts as subsets containing only the glyphs already used. Typing a character the font does not have produces silent garbage — the first version of this turned "Latin fixture page 1" into "Latin fixite Luen ure page 1" while reporting success. So every edit is read back and verified, and one that does not survive re-encoding is rolled back with an explanation. Corrupting a document and calling it success is worse than refusing.

What that leaves is a reliable tool for correcting text — typos, numbers, labels — and an honest refusal everywhere else. It is not a word processor, and the editor shows you the exact run it is about to replace rather than letting you assume you are changing one word.

Forms — click a field and type. AcroForm text fields, checkboxes and radio buttons, filled through PDFium's form-fill environment and written back into the document. The cursor changes over a field, which is how anyone discovers a PDF is fillable at all.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗