dsh-invoice-tools

Every month-end close means staring at a dozen fully-electronic VAT invoice PDFs, transcribing each field by hand, verifying the arithmetic, and stitching together a reconciliation sheet. dsh-invoice-tools collapses that workflow into two steps: point it at the files, and it returns structured JSON plus a ready-to-submit expense report. Parsing first reads the embedded XML attachment inside each shudian or quandian invoice, extracting invoice code, number, pre-tax amount, tax, tax-inclusive total, buyer, seller, and issue date through a bilingual tag whitelist. When no XML is present it falls back to a text-layer regex path, then cross-checks that tax-inclusive total equals amount plus tax (±0.01 tolerance) and tags every record with high, medium, or low confidence. The summary step aggregates by category, flags duplicate invoice numbers with matching amounts, and emits a Markdown or xlsx expense sheet—all with zero third-party runtime dependencies and strictly local execution.

The plugin reads no credentials, makes no network requests to any verification service, and scopes its parse cache per session. It is built for finance and admin staff who reconcile digital VAT invoices every month, and for teams already running DeepSeek Harness workloads who want the ticket-entry step handled automatically by the agent. Pure image scans are skipped with an explicit reason rather than silently mis-parsed, and a community vision plugin can be wired in later for OCR support.