/analyze-variants

name: analyze-variants
description: Use when investigating common or slow variants, repeated work, unusual paths, or a specific Process Mining case.
allowed-tools: Read, AskUserQuestion

Variants, rework and exact cases

Use the [shared contract](../../references/analysis-contract.md) and [filters/units](../../references/filters-and-units.md). Choose the branch answering the question; do not run a variant overview before an exact-case or rework request.

Ordered traces

**Common or slow variants**

1. `get_variants_with_metrics_v2 {S,filterOptions:F,itemsPerPage:5,itemsToSkip:0,metricToSortBy:"CaseCount",sortOrder:"Descending"}`. 2. Only if mean-case slowness still needs comparison, repeat with `metricToSortBy:"CaseDuration"`. If the first complete result already answers it, stop. `CaseDuration` is a mean; variant `Duration` is not its substitute.

**Cases with repeated work**

1. Copy F, appending to `caseMetricConditionFilters`: `{metric:"CaseReworkCount",comparisonOperator:"GreaterThan",values:[0],isInclusive:true}`. 2. `get_cases_with_metrics_v2 {S,filterOptions:Fplus,itemsPerPage:5,itemsToSkip:0,metricToSortBy:"ReworkCount",sortOrder:"Descending"}`. A native count suffices; no formula or saved metric is needed. 3. Only for unresolved transition evidence: `get_edges_with_metrics_v2 {S,filterOptions:Fplus,itemsPerPage:10,itemsToSkip:0}`. No edge sort exists; mark incomplete evidence page-local.

**A specific case**

1. Reuse details to identify the metadata-proven case-key attribute K and its type T. Bind the exact user/returned value X, e.g. synthetic `"Demo-17"`. 2. Append `{attributeName:K,dataType:T,attributeValues:[X],isInclusive:true}` to F's `attributeValueFilters`; call `get_cases_with_metrics_v2 {S,filterOptions:Fcase,itemsPerPage:5,itemsToSkip:0}`. 3. An optional same-cohort bottleneck/edge query can summarize activities/handoffs. If no exposed key expresses the case, stop; do not scan all cases or invent `caseId`.

Guardrails

Continue relevant pages or drilldowns until the question is answered or evidence cannot discriminate; no fixed call/total-row quota or automatic repeated queries. Show at most five headline rows by default. Use `filterOptions`, `itemsPerPage` and `itemsToSkip`, not `filters`, `metrics`, `limit` or `offset`. No `variantId` filter exists. For a user-specified sequence use only discovered attribute values and the advertised sequence-filter schema; do not reconstruct unavailable variant `Values`. No event-timeline or conformance-check tool exists. Rework/variation alone proves neither event order nor a conformance violation. Retain scope, versions, metric units and partial coverage.

Activity occurrence is not full payment, successful dismissal, business closure or compliance. Check the business outcome definition and available attributes/measures; route missing measures to derive-metric, never infer success from an event name. Missing events or terminal-looking traces may be censored, ongoing, legitimately ended or incompletely recorded. Do not choose without closure, recording and observation-coverage evidence. Use the [investigation method](../../references/investigation-method.md): distinguish observed presence/counts from tested outcomes; mark an unresolved outcome **inconclusive** and retain definition provenance, baseline/test filters, evidence references and the next discriminating check.