dsh-excel-chat

简体中文 | English

A DeepSeek Harness plugin for working on Excel in plain language. Say "add a margin formula in column D, bold the header, freeze the first row, add a filter" and the agent calls excel_operate to do it. Every edit is followed by an automatic formula check, and you can also ask it to "find what's wrong with this sheet" and have it repaired deterministically. It all happens in the conversation — no Excel UI to learn.

Two things that set it apart

1. It checks the formulas after every edit, instead of declaring victory. The annoying part of Excel is not writing a formula — it is a formula being silently broken: a fill handle that shortens a range, a row insert that shifts references, a failed lookup returning #N/A. Excel reports nothing; the numbers just start being wrong. This plugin runs a formula check after every excel_operate / excel_task (column pattern consistency, hardcoded cells, empty gaps, circular references, error values), can repair what it finds with a deterministic fixer, and shows you the before/after diff. excel_undo rolls an entire edit back from the audit log.

2. The core capabilities do not need Excel installed. Formula checking and repair, reading and writing, styling, subtotals, merges, mail merge and image embedding are implemented at the XML layer, so they run on macOS, Linux and Windows alike. Only chart creation and editing, native pivot tables, chart PNG export, visual review and PDF export need Windows with a local Excel install (via COM).

Install

Requires DeepSeek Harness (the dsh CLI or the desktop app).

dsh plugin --profile demo add dsh-excel-chat
dsh web --profile demo

Installing from source or a local directory also works (prepare builds it for you):

dsh plugin --profile demo add ./bundle

Slash commands

If the tools never show up, run the self-check first — it separates "the plugin was never loaded" from "it loaded but registration failed":

dsh-excel-chat-doctor                                    # when installed globally / via npx
~/.dsh/profiles/demo/node_modules/.bin/dsh-excel-chat-doctor

Switching language

Output is Chinese by default. Every message a person reads follows this switch: health reports, data insights, the capability menu, repair summaries, operation warnings, formula explanations, verification results, planner salvage/validation notes, failure attribution, doctor output, and the three slash-command descriptions. Add an override to your profile's cordis.patch.yml to switch it to English:

- id: vera
  config:
    language: en        # zh (default) | en

Data inside the workbook is never translated. Subtotal labels (总计), generated sheet names (-汇总) and preset names are artifacts written into the file, and code keys off them — patterns.ts identifies and skips summary rows by matching 总计/小计 — so translating them would silently change behaviour. Your own sheet names, column names and cell contents are of course left exactly as they are.

Things you can say

• "Turn report.xlsx into a report: margin in column D, totals in column E, bold the header, freeze the first row, add a filter"
• "Check whether the formulas in sales.xlsx are wrong and fix them"
• "Build a pivot table by region and a bar chart"
• "What's wrong with this sheet?" — runs data insight plus the formula check first, then suggests next steps
• "Show me this file first" — column types, missing values, uniques, samples and a suggested read range

Reliability

A self-built evaluation corpus of 100 workplace tasks (ExcelBench lite: editing / analysis / formulas / multi-step workflows). In goal mode with glm-5.3-flash the measured task success rate is 86% (DeepSeek baseline 52%). The per-run results table is generated from raw run output and anyone can add a row; the failure attribution, the re-test method and the known limitations are all written up in the evaluation document. 423 unit tests pass.

Tools (25)

Understanding a file

• excel_read — exact cell state (value / formula / type / format / merges / data validation)
• excel_profile — quick look at a big sheet: headers, per-column type / missing / uniques / range / samples, plus a suggested read range
• excel_semantic_profile — semantic profile: column roles (time / dimension / measure / identifier), granularity, cross-sheet join keys
• excel_menu — can't describe what you want? Get a menu: a one-line summary plus options for cleaning, reports, pivots, charts and more
• excel_insight — data insight: summary plus checks for missing values, duplicates, outliers, negatives, stray whitespace, and next-step suggestions
• excel_preview — table preview: a Markdown table inline in the conversation, plus an HTML preview file

Formulas

• excel_validate_formulas — silent formula errors (column pattern, hardcode, empty gaps, circular references, error values)
• excel_compile_formula — Formula IR → a deterministic Excel formula
• excel_explain_formula — plain-language explanation of a formula (functions, referenced ranges, cross-sheet references)
• excel_trace — dependency chain: precedents (what it reads) and dependents (what reads it), with a configurable depth
• excel_find_errors — every error-value cell, with the formula that produced it and a count per error code
• excel_repair_formulas — deterministic repair plus optional LLM repair, writing .repaired.xlsx and re-validating (with autoTable to detect the header row)
• excel_autofix — one-call self-healing: check → repair → re-check → plain-language report, with a hidden health-report sheet
• excel_health_report — writes the health check into the workbook itself (hidden sheet: score plus anomaly list)

Operations and tasks

• excel_operate — workplace-grade Excel operations: set values (with type inference), fill / series, insert and delete rows and columns, copy / move / transpose, format painter, formula-to-value, sorting (multi-key, by colour, custom list), the report one-shot template, subtotals, dynamic pivot-style summaries, crosstabs, two-table keyed backfill, advanced filters, styling, data validation, conditional formatting, autofilter, structured tables, image embedding, CSV import/export (export guards against formula injection by default, import reverses it), page setup, headers and footers, defined names, frozen panes, hidden rows and columns, row and column grouping, hyperlinks, comments, sparklines, find and replace, sheet protection, mail merge, data cleaning (dedupe / fill blanks / split columns / case / fullwidth-to-halfwidth) and fuzzy matching; formulas are re-validated afterwards and an audit log is written
• excel_task — multi-step orchestration (each step is checked and repaired) plus a goal loop (LLM plans → executes → verifies → replans when the goal is not met)
• excel_undo — rolls back an excel_operate edit from the audit log (restoring cell types as well as values)
• excel_diff_workbook — cell-level diff between two workbooks

Charts and export

• excel_validate_charts — chart structure validation
• excel_validate_charts_visual — Excel PNG export plus a visual LLM review (Windows)
• excel_create_chart / excel_modify_chart — create and modify charts with local Excel (type, title, legend, axes; Windows)
• excel_create_pivot — native pivot table (multiple row/column fields, filters and value fields; Windows)
• excel_export_charts — export charts to PNG with local Excel (Windows)
• excel_export_pdf — export a workbook or a single sheet to PDF with local Excel (Windows)

In action

Formula check + auto-repair: the diff at a glance	Data insight: problems found for you

Both screenshots come from a real model calling the tools inside DeepSeek Harness Web and the result being rendered inline in the message stream. More screenshots, a demo animation and the architecture notes are on the GitHub repository.

Documentation

• Full usage guide (scenarios, parameters, platform limits, FAQ)
• Evaluation method and results
• Changelog

Development

npm run build      # runs automatically via prepare when installing from source
node --test tests/*.test.ts

License: MIT