Review formula changes between two workbooks.

What is the direct answer to this workbook question?

Running git diff on two Tableau .twb or .twbx files produces thousands of noisy XML lines from window coordinates and cache tags. To review real changes with an AI agent, extract only data source columns, calculated field formulas, and shelf encodings, or drop both workbooks into the xmetric inspector to compare their structure.

Why is a raw git diff difficult to read for .twb and .twbx?

Two problems make reviewing workbook changes in pull requests painful:

1. .twbx is a compressed ZIP archive: Git treats .twbx as a binary blob (Binary files a/sales.twbx and b/sales.twbx differ).
2. .twb XML changes every time you click Save: Even if an analyst only changes a single filter from >= 2025 to >= 2026, saving the workbook in Tableau Desktop rewrites thumbnail caches, window pixel dimensions, active tab focus, and metadata object IDs across thousands of lines.

When you feed that 15,000-line raw XML diff to an AI coding agent, it burns tokens on UI noise and can miss a silent change inside a { FIXED } Level of Detail calculation.

How can a Python script compare workbook definitions?

Before asking Claude Code, Cursor, or a CI review bot to review a workbook update, normalize both .twb or .twbx files into a clean JSON summary containing only semantic logic (data sources, calculated fields, and worksheet shelves):

import json, zipfile, sys
import xml.etree.ElementTree as ET

def extract_workbook_spec(path):
    if path.endswith(".twbx"):
        with zipfile.ZipFile(path) as zf:
            twb_name = [n for n in zf.namelist() if n.endswith(".twb")][0]
            root = ET.fromstring(zf.read(twb_name))
    else:
        root = ET.parse(path).getroot()

    calcs = {}
    name_map = {}
    for col in root.findall(".//datasource/column"):
        raw_name = col.attrib.get("name", "")
        caption = col.attrib.get("caption", raw_name)
        name_map[raw_name] = caption
        calc = col.find("calculation")
        if calc is not None and "formula" in calc.attrib:
            calcs[caption] = calc.attrib["formula"]

    sheets = {}
    for ws in root.findall(".//worksheets/worksheet"):
        ws_name = ws.attrib.get("name")
        rows = [t.text for t in ws.findall(".//table/rows") if t.text]
        cols = [t.text for t in ws.findall(".//table/cols") if t.text]
        sheets[ws_name] = {"rows": rows, "cols": cols}

    return {"calculations": calcs, "worksheets": sheets}

if __name__ == "__main__":
    print(json.dumps(extract_workbook_spec(sys.argv[1]), indent=2, sort_keys=True))

Now run a clean diff between the old and new versions:

diff -u <(python3 spec_extract.py main_report.twbx) <(python3 spec_extract.py pr_report.twbx)

Which semantic changes should I check in a workbook review?

Once the XML noise is stripped, ask your agent to audit these four risk areas:

• Modified calculated field formulas: Did a denominator change from COUNTD([Customer ID]) to COUNT([Customer ID]), or did a FIXED LOD gain or lose a partitioning dimension?
• Changed filter types on worksheets: Did a dimension filter get added to or removed from Context (context='true'), which silently changes every FIXED LOD result on that sheet?
• Swapped pills on Rows, Columns, or Marks: Did a measure aggregation change from SUM to AVG, or did a table calculation's addressing fields shift?
• Data source join or Custom SQL edits: Were join keys altered or INNER JOIN clauses switched to LEFT JOIN?

How can I inspect the before and after workbooks?

A formula diff tells you what changed in the calculation syntax. To see the dependency structure of both files, drop before.twbx and after.twbx into the xmetric workbook inspector together; it accepts several files at once to compare them:

• No desktop install needed. Reviewers can inspect both workbooks' data sources, sheets, dashboards, dependencies, fields nothing uses, and calculations that refer to a missing field, free in the browser, without installing Tableau Desktop (which offers a free local edition for one person without publishing or sharing, free Tableau Reader, or Tableau Creator at $75 per user/month, billed annually as of October 2026).
• Everything stays local. Save the data files packaged inside a .twbx, copy the inspection as JSON for diffing, or test modified formulas in the formula playground without uploading company data. To see the changed sheets as charts, open the new version in the xmetric app.
• Agent-ready architecture. The public xmetric MCP server is live at https://getxmetric.com/mcp with no authentication or API key. Anonymous initialize, tool discovery and widget-resource checks passed on 2026-10-29. ChatGPT and Claude host rendering and tool selection remain unverified; their directory listings remain coming. Workbook diffs remain a browser-inspector or script workflow; the public MCP catalog does not provide a workbook-diff tool.

What is live, and what remains planned?

The public xmetric MCP server is live at https://getxmetric.com/mcp with no authentication or API key. Anonymous initialize, tool discovery and widget-resource checks passed on 2026-10-29. ChatGPT and Claude host rendering and tool selection remain unverified; their directory listings remain coming. See the dated public protocol receipt and setup and evidence limits.

The public server is available; host operation and directory approval are separate checks.

The file board parses supported .pbix, .twbx/.twb, CSV, Excel and Parquet files in your browser. Choosing a file shares supported schema, source names and calculation definitions with xmetric and the chat host; requested result cells, filters and SQL or DAX evidence are shared when answering. The original file is not uploaded by this flow. Small aggregates and labels can still be sensitive. Privacy details.

What data does this worked example use?

The initial view uses authored synthetic data, not customer data or a captured agent session.

The example record lists its file, calculation or SQL and expected values. The live browser result is computed by xmetric’s existing engine or workbook inspector. Samples are excluded from own-data activation reporting.