| name | mcp-engine-dax-debugging |
| description | Use when a Power BI measure or query returns wrong values, unexpected blanks, inflated or duplicated totals, a total row that disagrees with its detail rows, a slicer that has no effect, or numbers that disagree with the source system. For slow-but-correct queries, use mcp-engine-dax-performance; for a whole-model assessment, use mcp-engine-model-quality. |
PBI DAX Debugging
Correctness debugging for wrong numbers. The full loop lives in dax-debugging-workflow — follow it step by step; stay read-only until the root cause is proven and reconciles the discrepancy numerically.
Workflow
- State the discrepancy — measure, exact filter context, actual vs. expected, and where the expectation comes from.
- Reproduce minimally —
run_query { "operation": "execute", "query": "<CALCULATETABLE(ROW(...), <filters>) reproduction query>" } that rebuilds the wrong number outside the report.
- Establish ground truth — agree on the business definition and grain, then compute that same aggregation from the most primitive fields while bypassing the suspect measure. Use raw
SUM only for additive metrics; preserve DISTINCTCOUNT, averages, ratios, semi-additive balance logic, and other non-additive semantics. Where the two diverge decides the branch: measure defect, model/data defect, or definition mismatch.
- Decompose — fetch the expression with
list_model { "operation": "list", "spec": { "type": "measures", "include_expression": true } } and binary-search its parts, one ROW probe per sub-expression, recursing into referenced measures by extracting the [Measure Name] references from the expression itself (manage_dependencies used_by lists consumers, not inputs).
- Inspect the filter path —
list_model relationships listing for every hop between filter table and fact: existence, active state, direction, cardinality, key data types; duplicate-key and orphan-key probes from the guide.
- Match the symptom — the guide's symptom table (inflated, blank, total mismatch, dead slicer, source mismatch) names the deciding probe for each cause.
- Rule out security and masking —
run_query { "operation": "test_access", "query": "<reproduction query>", "spec": { "roles": ["<role>"] } } under the affected role; check masking flags before calling data wrong.
- Close numerically, then fix and protect — the explanation must reconcile actual to expected arithmetically; the fix routes to the authoring skill (if it is not installed, apply the fix from the tool's
inputSchema or ask the user to add it), and a manage_tests measure_assertion or referential_integrity test pins the corrected behavior.
Guardrails
- No fixes during diagnosis — a change that "makes it right" without an explained cause hides the bug.
- One variable per probe; keep a written trail of every probe and result.
- Keep probes aggregated; never dump raw rows that may contain sensitive values.
- A total-vs-detail mismatch can be correct by design for non-additive measures — verify intent before declaring a defect.
Report results
After a debugging session, report:
- The discrepancy: measure, context, actual vs. expected.
- The root cause with probe evidence and the reconciling arithmetic.
- Its class: measure defect, model defect, data defect, security behavior, or definition mismatch.
- The fix applied or recommended, with the re-run verification result.
- The regression test added, with its id.
References