dsh-excel-kit

The moment a workbook balloons past a hundred megabytes, the default move—load the whole xlsx into memory, then describe, filter, or pivot—quickly exhausts the context budget of any inference pipeline. dsh-excel-kit swaps that full-load pattern for a pure streaming engine: lazy zip-entry decompression, row-by-row SAX parsing of worksheet XML, and shared-string tables that flush to disk in 8 MB chunks with a small LRU cache. Memory stays bounded end-to-end; a 133 MiB file peaked at just 153 MiB RSS delta for describe and as low as 36 MiB for pivot.

The plugin exposes three read-only tools. excel_describe profiles the workbook and each column's type distribution, non-empty count, empty rate, numeric min/max/mean, and sample rows. excel_filter supports nine operators (eq, ne, gt, gte, lt, lte, contains, in, between) for row selection, defaulting to a 100-row limit with a hard cap of 500. excel_pivot groups by user-specified row columns and aggregates values with count, sum, mean, min, or max, limited to 50 groups. Date cells are detected via numFmtId and emitted as ISO strings; scientific notation, booleans, inline strings, and empty cells are all handled. Results exceeding the 32 KB spill threshold are persisted through ctx.spillStore so a single response never floods the model context.

If you are building on dsh (DeepSeek Harness) and need the model to inspect a large Excel file, run a quick filter, or produce a small pivot table—without pulling in pandas, without loading the whole workbook, and without any write or format side-effects—dsh-excel-kit fills that gap. It is not a BI replacement; it is a lightweight, read-only, low-memory analytics layer designed for inference-time table access.