988hj7tczd-oss/dsh-invoice-tools

dsh-invoice-tools is a native tool plugin built specifically for DSH (DeepSeek Harness) and mounted into the host through a cordis assembly manifest named dsh-invoice-tools, where it registers two tools against the host tools registry. The first tool, invoice_parse, reads VAT electronic-invoice PDFs and images located inside the DSH workspace and structures them into JSON: it first tries to extract the embedded XML under /EmbeddedFiles, falls back to a regex text-layer pass for label-driven fields such as 发票号码：, 金额（大写）…（小写）¥x, 开票日期： and 购买方名称：, runs a cross-check that 价税合计 = 金额 + 税额 within a 0.01 tolerance, and tags each record with a high/medium/low confidence plus a checklist flag.
The second tool, invoice_summary, aggregates parsed invoices, subtotals by category, detects duplicates on 发票号码 + 金额, and writes a 报销单-<日期>.md Markdown report or an xlsx file produced through a minimal in-house OOXML zip packager, defaulting to the previous parse cache when files is omitted. Two installation modes are documented: production via dsh plugin add with the bare package name resolved by Node module loading so that dsh.bundle.patch ships with the package, and a development direct-mount that uses an absolute path in the host cordis.yml and requires a sibling dsh-src checkout followed by pnpm install && pnpm run build. Typical users are finance and admin staff who need to batch-process Chinese VAT invoices entirely on-device through a DSH agent loop.
Dependencies are intentionally minimal: the host must provide the tools and fs services, while the plugin itself relies only on Node's built-in zlib and ships a custom minimal PDF parser that supports FlateDecode streams, Tj/TJ text operators, latin1 and UTF-16BE strings, the /EmbeddedFiles attachment tree, and is robust to old-style xref plus xref streams without depending on the xref table itself. Documented limits include a directory recursion depth of 3, a 16MB per-file read ceiling, no support for indirect-reference streams inside object streams, no JPEG2000 decoding, no page rendering, no Chinese uppercase amount parsing, and an OCR-free first release that returns explicit skip reasons for image-only PDFs. Because the DSH filesystem service only exposes a UTF-8 text channel without a binary write API, the xlsx output is written as 报销单-<日期>.xlsx.b64 base64 and must be decoded with base64 -d before Excel/WPS/openpyxl can open it, while the tool return value also carries an xlsxBase64 field for in-session decoding. The license is MIT, no verification network calls are made and no credentials are read, file I/O goes through ctx.fs under the host sandbox, and the parse cache key is namespaced by <session id>:<absolute path> with a documented process-level global: fallback when agent.id is unavailable, so sensitive invoices should not be cached.