huashu-excel: A Traceable Excel Analysis Workflow for DeepSeek Harness

What Does It Do?

huashu-excel is a DSH plugin for DeepSeek Harness that turns a messy Excel workbook into a traceable, defensible analysis report. It solves the core problem that AI data analysis fails silently: a wrong number looks exactly like a right one, and nobody can answer "how did you calculate this?" The plugin runs an eight-stage pipeline covering inspection, cleaning, alignment, analysis, reconciliation, delivery, chart verification, and quality control, so every figure traces back to the original cells. It is maintained by alchaincyf under the MIT license and depends only on openpyxl, with CSV paths and report generation needing nothing beyond the standard library.

Core Features

• Dirty-sheet inspection: reads raw cells before touching any dataframe and lists every issue, including header offset, two-level headers, subtotal rows mixed into the data area, duplicate rows, merged cells, numbers stored as text, and inconsistent missing-value markers.
• In-sheet reconciliation: treats the workbook's own total and subtotal rows as free checksums and cross-validates them against cleaned detail rows, catching even a hand-typed subtotal that is off by ten yuan.
• Eight-stage workflow: inspection, cleaning, alignment, analysis, reconciliation, delivery, chart verification, and quality control, each step traceable and replayable.
• Pre-delivery gate: a master check blocks any number from being delivered until it passes, and a separate agent that did not build the analysis recomputes everything from the raw data to challenge the result.
• Standalone verification scripts: table profiling, number verification, chart verification, and Word document verification all run outside the agent, with a non-zero exit code signaling a problem.
• Environment-aware execution: probes runtime capabilities at start, running parallel analysis when possible, rotating perspectives when serial, and falling back to step-by-step Excel instructions when scripts cannot run.

How to Use This Plugin?

After enabling it in DSH, just describe what you need in plain language and the plugin picks the matching path: "analyze this sales sheet" runs the full eight stages, "why do these two numbers not match" runs inspection plus reconciliation, "how many dirty rows are in this table" runs inspection only, "how did you calculate this" performs reconciliation and definition tracing, and "make me a report" produces charts, insights, and output in html, xlsx, or docx. It probes the environment first and decides between parallel and serial execution, and nothing is delivered until the master check passes. The workflow is agent-agnostic and can also be used without installation by reading the methodology document directly.

How to Troubleshoot This Plugin?

Verification scripts report problems through exit codes: a non-zero exit from the number verification script means some figures do not reconcile, a non-zero exit from the chart verification script means a chart was drawn incorrectly, and a non-zero exit from the Word document verification script means the file will shift when opened on another machine. When these errors appear, fix the corresponding stage according to the official instructions and keep the affected numbers out of any deliverable until the issue is resolved.