@zaowu-labs/doc

Document workflow package for ZaoWu.

This package extracts and summarizes text from supported document inputs, including text-like files, HTML, PDF, and DOCX in the foundation version. HTML extraction preserves heading levels and lightweight table row structure when producing text for outline, search, summary, or conversion commands.

PDF Support Note (Foundation Workaround)

PDF text extraction uses pdf-parse. That library has a transitive/optional dependency on @napi-rs/canvas. Even for pure text extraction, in many Node environments it will emit a warning or assume certain DOM globals exist.

We install minimal shims for DOMMatrix, ImageData and Path2D on globalThis (only if they are missing) and temporarily suppress the exact canvas warning during the dynamic import. This is a contained, documented workaround.

Long-term we would prefer a pure-JS text-only PDF extractor that does not require these shims. See the source comments in src/pdf.ts (functions ensurePdfTextRuntime and importPdfParse) for the full explanation and the class names used for the shims.

Safety:

• Document conversion previews by default when an output path is supplied.
• Confirmed writes refuse unsafe overwrites.
• Format support should stay explicit in command help and docs.

Extraction contract and limits

DOCX uses the existing focused JSZip and xmldom libraries directly, removing Mammoth and its unused CLI dependency chain. It preserves main-body paragraphs, heading styles (including inherited/strict namespaces), list structure, tables, and HTTP(S)/mailto hyperlinks. It never follows external relationships, extracts archives to disk, or runs macros. Literal Word text is escaped for Markdown. Footnotes, images/OCR, complex layout, embedded objects, and tracked deletions are outside this text extraction contract. Re-export complex documents as text when exact visual fidelity matters.

Input files are limited to 20 MiB; requested DOCX XML parts to 8 MiB each and 2,000 archive entries; extracted text to two million characters. DTD/entity XML and excessive nesting are rejected. XML budgets apply to actual inflated bytes. PDF extraction with no text layer explains how to use OCR rather than returning a misleading successful empty summary. Concurrent PDF extraction serializes temporary runtime shims and restores the original globals.

summary explicitly reports method: "extractive": it takes the first three prose lines, without calling a model. Frontmatter, fenced code, and headings do not become summary prose; fenced examples do not create false outline headings. See the report-to-model recipe.

Confirmation execution also checks the freshly prepared content against the already verified plan immediately before the write/request. Changes between reads fail with OPERATION_PLAN_MISMATCH and produce no write or network call.