sheetdelta

Compare two Excel workbooks and find out what actually changed.

sheetdelta reads two .xlsx files, shows you the difference cell by cell, and then tells you which of those changes reach other cells. It runs on Linux, in CI, with no copy of Excel anywhere near it.

$ sheetdelta diff budget_v1.xlsx budget_v2.xlsx
budget_v1.xlsx  ->  budget_v2.xlsx

Sheet 'Revenue'  (1 change, 1 breaking)
  !! D12    formula changed
          -  =SUM(D2:D11)
          +  =SUM(D2:D13)
        affects 3 cells downstream: Revenue!E12, Revenue!F12, Summary!B4

Sheet 'Summary'  (1 change)
  !  B7     value changed
          -  1200
          +  1450

1 formula; 1 value across 2 edited sheet(s). 1 breaking change reaching 3 downstream cells.

Why

Two things go wrong when a spreadsheet changes and nobody looks closely.

The first is the change itself. git diff on an .xlsx gives you binary noise, so the usual answer is to open both files side by side and squint.

The second is worse. A formula edit looks harmless on the sheet you are looking at, but the cell feeds a summary three sheets away. The number that matters is now wrong, and nothing tells you.

sheetdelta answers both. Every change is followed through the dependency graph, so a formula edit reports the cells it reaches, and a CI job can fail the build on exactly that.

Install

pip install sheetdelta

No dependencies. The reader is zipfile and xml.etree from the standard library, so it drops into any environment, including a build container that has nothing but Python.

Usage

Compare two workbooks

sheetdelta diff old.xlsx new.xlsx

Exit status is 0 when nothing breaking changed, 1 when something did, 2 on an error. That makes it a CI check without any extra scripting:

- run: pip install sheetdelta
- run: sheetdelta diff before.xlsx after.xlsx --fail-on breaking

--fail-on takes three values:

Machine-readable output

All subcommands take --format:

sheetdelta diff old.xlsx new.xlsx --format json

--json and --summary are kept as aliases for --format json and --format summary. They are mutually exclusive with each other, and with --format; pick one. diff also takes --format markdown, which renders the same report as a Markdown table and list, ready to paste into a pull-request comment or a job summary.

{
  "old": "budget_v1.xlsx",
  "new": "budget_v2.xlsx",
  "has_changes": true,
  "summary": {
    "cell_changes": {"formula": 1, "value": 1},
    "breaking": 1
  },
  "sheets": [
    {
      "kind": "changed",
      "name": "Revenue",
      "old_name": "Revenue",
      "shifts": [
        {"axis": "row", "inserted": true, "count": 1, "at": 3, "label": "1 row inserted at row 3"}
      ],
      "moves": [
        {"old_row": 3, "new_row": 5, "label": "row 3 moved to row 5"}
      ],
      "changes": [
        {
          "cell": "Revenue!D12",
          "a1": "D12",
          "kind": "formula",
          "severity": "breaking",
          "detail": "formula changed",
          "old": "=SUM(D2:D11)",
          "new": "=SUM(D2:D13)",
          "affected": ["Revenue!E12", "Revenue!F12", "Summary!B4"]
        }
      ]
    }
  ]
}

shifts and moves are always lists, empty when the sheet held no insert, remove or reorder, so a consumer never has to branch on their presence.

Summary output

sheetdelta diff old.xlsx new.xlsx --summary
sheetdelta diff old.xlsx new.xlsx --format summary

The full report is what you want on a laptop. In a CI log you often want the shape of the change and nothing else: one line per sheet, with the counts.

$ sheetdelta diff budget_v1.xlsx budget_v2.xlsx --summary
budget_v1.xlsx  ->  budget_v2.xlsx
  Revenue: 1 row inserted at row 3, 2 changes, 1 breaking
  Summary: 1 change
1 formula; 2 added across 2 edited sheet(s). 1 breaking change.

Check one workbook

sheetdelta audit workbook.xlsx

An audit looks for the two defects that make a workbook unsound:

• a formula pointing at a sheet that does not exist, which Excel turns into #REF! when it recalculates; and
• a circular reference, a cell that depends on itself through a chain of other cells, which Excel refuses to calculate at all.

$ sheetdelta audit model.xlsx
model.xlsx
  3 sheet(s), 412 cell(s), 118 formula(s)

1 broken reference(s):
  !! C4     Removed Sheet!B2
        sheet 'Removed Sheet' does not exist
        ='Removed Sheet'!B2*1.2

1 circular reference(s):
  !! Revenue!D12 -> Revenue!E12 -> Summary!B4 -> Revenue!D12

A reference to a cell that is simply empty is not reported. That is normal in a spreadsheet, and flagging it would make the audit useless on real files.

Volatile references

INDIRECT and OFFSET build their target at runtime, from a string or an offset, so there is no way to know which cells they read without evaluating the formula. The scanner does not guess, but it does not stay silent either: the audit lists each call site as a volatile reference, separately from the broken and circular ones.

$ sheetdelta audit model.xlsx
model.xlsx
  2 sheet(s), 96 cell(s), 31 formula(s)

No broken references, no circular references.

1 volatile reference(s):
  ?  D4     INDIRECT(Sheet2!A1)
        =INDIRECT("Sheet2!A1")

A volatile reference does not make the workbook unsound, so it does not change the exit code -- Excel will compute it fine. What it changes is the dependency graph: a change made through one cannot be followed, so an audit that only said "sound" would be overstating what it can see.

For the graph itself, a volatile reference is treated as conservatively reaching its whole sheet, so an edit on that sheet is still reported as affecting the volatile cell. --volatile-scope workbook widens that to the whole workbook on both diff and audit, at the cost of more changes being reported as impacting each other.

GitHub Actions

Both subcommands can emit GitHub Actions annotations:

sheetdelta diff old.xlsx new.xlsx --format github
sheetdelta audit workbook.xlsx --format github

Each line is a workflow command, so the change shows as a marker on the pull request. The severity carries over: a change that reaches another cell is ::error, an unread change is ::warning, and additions or sheet changes are ::notice. When nothing changed, diff emits a single notice so the step does not look skipped.

A reusable composite action wraps all of this. A workflow that compares the workbook on the branch against the one on the base revision:

- uses: actions/checkout@v4
  with:
    fetch-depth: 0
- run: git show "origin/${{ github.base_ref }}:models/forecast.xlsx" > /tmp/before.xlsx
- uses: sqmyou/sheetdelta/.github/actions/sheetdelta-diff@main
  with:
    old: /tmp/before.xlsx
    new: models/forecast.xlsx
    fail-on: breaking

The action installs sheetdelta, runs the selected mode (diff or audit) in the github format, writes the human report to the job summary, and exits with the diff's code. See .github/workflows/workbook-diff.yml for a full example.

What it reports

Formatting alone is not a change. Switching a cell between a date format and a plain number shows a different string but leaves the stored value identical, so it is not reported. Likewise 1 and 1.0 compare equal. Reports still show the formatted text.

A row or column insert is reported as one event, not as every cell below it. Excel has no "insert row" in the file format: it rewrites each cell below to a new address. Compared naively, inserting one row near the top makes the whole rest of the sheet look changed and buries the real edit. sheetdelta lines the contents up instead, so it can say 1 row inserted at row 3 and leave only the genuinely new cells as changes. A sheet can hold more than one: each insert or remove is found in turn and the remaining difference re-checked, so two separate blocks inserted far apart are both reported, and an insert and a remove in the same sheet are too. An insert is only claimed when the block below it really does line up, so two unrelated sheets are never aligned by force. A real edit that moved with the insert is still reported. Shift rows are given in the new sheet's numbering, the number a reader sees on screen.

A row move is reported as one event too. Reordering rows -- dragging one up, or sorting a block while a header stays put -- rewrites the moved rows' addresses the same way an insert does, so the same problem arises. When the rows present on each side are the same and unchanged, only in a different order, sheetdelta re-keys the moved rows and reports row 4 moved to 2 rather than a screenful of cell changes. It only reads a difference this way when the rows really are a permutation: add, remove or edit even one row and the move no longer holds, and the change is reported as ordinary cell changes. A row whose contents appear twice is left alone, because which copy moved cannot be known.

A structured reference into an Excel table is understood, not treated as text. A formula like SUM(Sales[Amount]) is resolved against the table's definition, so the dependency graph knows which cells it reads and an edit to any of them is reported as reaching the formula. The current-row form, [@Amount], is resolved to the single cell on the formula's own row rather than the whole column, so a formula that reads its own row does not look like it depends on every row.

The stale case is worth explaining. Excel stores both a formula and the last value it calculated for it. When a file is edited by something that does not recalculate, the two disagree. Excel will happily show you the new formula next to the old number, which is a quiet way to ship a wrong total. sheetdelta flags it.

A rename is detected by matching the contents of a removed sheet against an added one. Without that, renaming a sheet would look like every cell in it changed and bury the real diff. The contents need not match exactly: a sheet renamed in the same commit as a light edit is still recognised, as long as at least 90% of its cells are unchanged.

What it does not do

Being clear about this saves you time.

• It does not evaluate formulas. There is no calculation engine, and there will not be one. Excel already stored the value it last computed, and that is what gets compared. This is why the tool needs no Excel and no dependencies, and why it is fast on large files.
• It does not write or edit workbooks. Read-only, by design.
• It only reads .xlsx and .xlsm. The old .xls and binary .xlsb formats are rejected with a clear message rather than half-parsed.
• It does not read VBA macros inside .xlsm files.
• It aligns an insert, and a clean reorder. Inserting or removing rows or columns is recognised as one event, and so is a pure permutation of the rows: when the same rows are all present and unchanged, only in a different order, the report says which row moved where instead of listing every cell of every affected row as changed. The moment a reorder is mixed with an add, a remove or an edit, it stops being a permutation and is reported as ordinary cell changes -- the tool never claims a move it cannot prove.

Using it as a library

from sheetdelta import diff_workbooks, read_workbook

result = diff_workbooks(read_workbook("old.xlsx"), read_workbook("new.xlsx"))

for change in result.breaking:
    print(change.ref, "->", change.affected)
    print("  was", change.old)
    print("  now", change.new)

Every type is exported from the package root, and the whole thing is typed.

How it works

An .xlsx file is a zip containing XML. sheetdelta reads the sheet list, the shared string table, the number formats and each sheet's cells, then stops.

Each formula is scanned for the cells and ranges it mentions. That scan is more careful than it looks: LOG10(...) looks exactly like a reference to column LOG row 10, and "see A1" looks like a reference to A1. Both are handled, because a dependency graph built on false references is worse than no graph.

References are kept as ranges rather than expanded. SUM(A:A) covers a million cells, and expanding it would cost more than the entire diff; the graph asks whether a range contains a cell instead.

Excel tables are read from the table parts, and a structured reference such as Sales[Amount] is resolved to the column of cells it names, so the graph sees through the table syntax to the addresses underneath.

A call to INDIRECT or OFFSET is the one case the scan cannot resolve: its target is built at runtime. It is recorded as a volatile reference rather than dropped, so the audit can report it and the graph can treat it as reaching the whole sheet. That keeps a change behind such a formula from being reported as having no effect.

Requirements

Python 3.10 or newer. No runtime dependencies.

License

MIT.