coop-dax-review 0.9.0

Offline, advisory DAX/model standards linter for Power BI semantic models: parses TMDL/.bim, checks measures and model structure against the team's DAX standards, and reports anything that doesn't match. Human report + machine JSON for the company agent. Never edits or blocks.

coop-dax-review

Offline, advisory DAX/model standards linter for our Power BI semantic models. It parses TMDL (and legacy .bim) models, builds a model catalog, checks measures and model structure against docs/standards.md (our DAX standards + Microsoft/Tabular best practices), and surfaces anything that doesn't match. It never edits or blocks — it only reports. Human reports (a sectioned, colorized terminal report, Markdown, or a self-contained branded HTML file) and machine JSON for the company analytics agent.

Sibling tool to coop-sql-review — same architecture and contracts.

Install

pipx install coop-dax-review        # from PyPI

Use pipx, not system pip, so the tool stays isolated from other CLIs (ms-fabric-cli, azure-cli) it might otherwise fight over shared pins. For local development:

python -m venv .venv && .venv/bin/pip install -e ".[dev]"

Usage

coop-dax-review check [MODEL_PATHS...] [--format text|json|markdown|html] [-o FILE]
                      [--html FILE] [--md FILE] [--open/--no-open] [--color/--no-color]
                      [--log-file FILE] [--baseline FILE] [--write-baseline FILE]
                      [--save-ignores] [--min-severity error|warning|info] [--strict]
coop-dax-review rules                 # list every rule (id, severity, tier, agent?)
coop-dax-review upgrade [--check]     # show the command to update (never self-applies; alias: update)
coop-dax-review --version

• MODEL_PATHS point at a PBIP/TMDL model folder (*.SemanticModel/definition/...), any folder of .tmdl files, or a legacy .bim file. Directories are searched recursively; defaults to ..
• Run it with no paths in a terminal and it offers a checkbox picker of the subfolders to check (all selected by default — press ENTER to scan everything).
• Advisory: exit code is always 0. --strict is the opt-in CI gate — exit 2 when any finding remains at/above --min-severity, or when no models were found/checked at all (so a typo'd path can't pass silently). A zero-model run still renders the full report (with models_checked: 0 and a scan_empty diagnostic per searched path).
• --standards <path> overrides the bundled standards (e.g. point it at a canonical company standards file). Its sha256 travels in the JSON so the agent knows which standards a report used.
• A rules.yml (found via --config, else a rules.yml in the current directory, else beside the standards file) can disable rules, override severities, and tune thresholds — all with no rebuild. A broken rules.yml (bad YAML, wrong shape, a non-UTF-8 save) is a friendly one-line error naming the file, and a --config path that doesn't exist is an error too (a typo can't silently drop your overrides). For example, raise what counts as a "non-trivial" measure:

  rules:
    DAX-VAR-RETURN:
      params: { min_functions: 5 }   # also: DAX-COMPLEX-NO-HEADER.min_vars,
                                      # DAX-DISPLAY-FOLDERS.min_measures, DAX-SIMPLE-FUNCTIONS.min_calculates
  

coop-dax-review check ./MyModel.SemanticModel
coop-dax-review check . --format json --strict --min-severity warning
coop-dax-review check . --format html              # writes a report file and opens it in your browser
coop-dax-review check . --format markdown -o report.md

The default --format text is a sectioned terminal report: a banner, one section per model with ERROR/WARN/INFO severity badges, and a SUMMARY panel. It's colorized automatically when you're at a terminal and falls back to plain text when piped or redirected (override with --color/--no-color; NO_COLOR is respected).

--format html produces a self-contained, branded HTML report (inline CSS + embedded logo, no network). It is always written to a file — coop-dax-review-report.html by default, or wherever -o points — and the path is printed and opened in your browser (pass --no-open to skip the open, e.g. in CI). upgrade/update print the exact command to run yourself (pipx upgrade coop-dax-review, etc.) rather than self-applying, since a package manager can't replace the tool while it is running; upgrade --check reports whether an update is available and stops there.

Want a report file and the usual console/JSON output in one run? --html FILE and --md FILE are extra sinks: they write a self-contained HTML and/or Markdown report to the paths you name in addition to whatever --format prints, and never open a browser. Handy for CI — e.g. keep the JSON contract on stdout while dropping a human-readable HTML artifact alongside it:

coop-dax-review check . --format json --html report.html --md report.md

What it checks

Run coop-dax-review rules for the live list. Deterministic rules (reported as findings):

Rule	§	Sev	Flags
DAX-MEASURE-CATEGORY	1	warning	measure not named [Category: Name]
DAX-MEASURE-NOT-PREFIXED	1	warning	Table[X] where X is a measure (measures take no prefix)
DAX-COLUMN-PREFIXED	1	warning	bare [X] where X is a column (columns need Table[X])
DAX-VAR-RETURN	2	info	non-trivial measure with no VAR/RETURN structure
DAX-NO-NESTED-CALCULATE	3	warning	CALCULATE nested directly inside CALCULATE (iterator-mediated per-row nesting is fine)
DAX-FILTER-TABLE-IN-CALCULATE	4	warning	FILTER(<table>, <col> = ...) where a plain column filter suffices
DAX-SNOWFLAKE	6	info	a table with relationships chained through it (snowflake link)
DAX-BIDI-RELATIONSHIP	7	warning	a bidirectional cross-filter relationship
DAX-MARKED-DATE-TABLE	8	warning	time-intelligence used (in a measure or calculated column) but no marked Date table
DAX-MEASURE-IN-ITERATOR	9	info	a measure referenced inside a row iterator (hidden context transition)
DAX-COMPLEX-NO-HEADER	12	info	a complex measure (≥3 VARs) without a /* ... */ header
DAX-DIRECTLAKE-NO-CALC-COL	13	warning	a calculated column in a Direct Lake model
DAX-USE-DIVIDE	14	warning	the / operator where DIVIDE() should be used
DAX-FORMAT-STRING	15	warning	a measure with no explicit formatString
DAX-NO-FLOAT-KEYS	16	info	a relationship key column typed double
DAX-HIDE-FK-COLUMNS	17	info	a visible foreign-key (relationship) column
DAX-KEY-SUMMARIZEBY-NONE	18	info	a numeric key column that auto-aggregates (summarizeBy ≠ none)
DAX-DISPLAY-FOLDERS	19	info	a measure-heavy table with no display folders

Agent-judgment rules — the tool detects the construct but emits to the JSON agent_review list (never an auto-finding), because the call needs intent the linter can't infer:

Rule	§	Judges
DAX-KEEPFILTERS-NEEDED	5	whether a CALCULATE boolean filter needs KEEPFILTERS (checked per top-level filter argument)
DAX-STAR-SCHEMA	6	whether a snowflake chain should be flattened to a star
DAX-CONTEXT-TRANSITION	9	whether an iterator's context transition is intended/correct
DAX-SIMPLE-FUNCTIONS	10	whether a CALCULATE-heavy measure could use simpler functions
DAX-VALIDATION	11	whether the §11 validation checklist was run for a non-trivial measure
DAX-IMPLICIT-MEASURE	20	whether a visible auto-aggregating numeric column should become an explicit measure

See RULES.md for the full taxonomy. docs/standards.md §14–§20 are adopted Microsoft/Tabular best practices (DIVIDE, format strings, key column types, hidden FKs, key summarizeBy, display folders, explicit measures); docs/standards-proposed-additions.md is the original candidate list.

Suppressing findings (adopting on an existing model)

Three deterministic, never-blocking ways to silence findings you've already triaged. All three apply to regular findings and to agent_review items (the judgment-call prompts in the JSON), so a triaged item stays silenced everywhere:

• Inline — drop a comment on a finding's line (or the line directly above it):

  // coop-dax-review:ignore DAX-VAR-RETURN reason: legacy measure, rewrite scheduled
  

  List several rule ids (ignore DAX-A, DAX-B), or use a bare ignore / * to silence every rule on that line. The reason: text is for humans; it's ignored by the parser.

• rules.yml ignore list — a human-readable, fingerprint-matched suppression list that lives in the one config file (like a baseline, but readable and hand-editable). Add an ignore: block:

  ignore:
    - fingerprint: 4ad6aeb79867
      rule: DAX-BIDI-RELATIONSHIP        # rule / where / note are for humans; matching is by fingerprint
      where: Sales/FactSales[ProductId] -> DimCustomer[CustomerId]
      note: intentional many-to-many, reviewed 2026-07
  

  You don't have to hand-copy fingerprints: run check --save-ignores and, at an interactive terminal, you get a checkbox of this run's findings (all unchecked — opt in to the ones you want gone); the picks are appended to rules.yml for you, so the next run silences them. A rules.yml in your current directory is auto-discovered with no --config flag, so the loop is just "run, --save-ignores, re-run". An ignore entry that no longer matches any finding (you fixed it) is reported as a diagnostic so the list self-cleans.

• Baseline (ratchet) — record today's findings and surface only new ones going forward:

  coop-dax-review check . --write-baseline dax-baseline.json   # once, to capture the status quo
  coop-dax-review check . --baseline dax-baseline.json         # thereafter: only new findings appear
  

  Each finding has a stable fingerprint (in the JSON) — independent of line numbers and of file paths, so a baseline written on one machine or from one directory still matches from another — and the baseline is a sorted list of those. A baseline entry that no longer matches any finding (you fixed it) is reported as a diagnostic so the file self-cleans (--write-baseline to prune).

  One-time migration (schema_version 2): fingerprints used to include the cwd-relative file path and changed in this release. Delete and regenerate any baseline files and rules.yml ignore: lists written by earlier versions (re-run --write-baseline / --save-ignores).

Agent JSON contract

{
  "tool": "coop-dax-review", "schema_version": 2, "version": "x.y.z",
  "standards": {"path": "...", "sha256": "..."},
  "models_checked": 2,
  "verdict": {"clean": false, "highest_severity": "warning"},
  "findings": [{"rule_id":"...","severity":"warning","model":"Sales","file":"...","line":12,
                "object":"[Sales: Revenue YTD]","message":"...","standard_ref":"§3","fingerprint":"4ad6aeb79867"}],
  "summary": {"error":0,"warning":2,"info":4},
  "agent_review": [{"rule_id":"...","model":"Sales","file":"...","line":40,"object":"[...]","note":"...","standard_ref":"§5","fingerprint":"..."}],
  "diagnostics": [{"severity":"warning","category":"parse_failed","file":"...","message":"..."}]
}

schema_version lets a consumer pin the shape; verdict/models_checked give a quick machine verdict + coverage signal; each finding's fingerprint is a stable id for tracking across runs.

Project docs

• SPEC.md — architecture, CLI, agent contract, milestones.
• RULES.md — every standard mapped to a concrete check (deterministic vs agent-judgment).
• docs/standards.md — the canonical DAX standards the linter checks against (bundled as package data).
• CLAUDE.md — orientation for Claude Code sessions in this repo.