技能 数据科学 自定义指标推导与验证

自定义指标推导与验证

v20261008
derive-metric
此技能用于解决缺失指标问题,通过复用已保存的自定义指标或创建临时公式。它指导用户确认指标定义、单位和聚合逻辑后再进行计算。优先使用原生输出和已保存指标,避免不必要的公式创建。确保公式验证并报告非持久化结果,不保存新定义。
获取技能
371 次下载
概览

Resolve a metric gap without saving definitions

Use the shared contract and
filters/units. Follow this order; stop at the first sufficient option.

For a business-derived measure, confirm case/event entity, outcome definition/provenance,
units, time basis and aggregation. Are amounts incremental transactions or cumulative snapshots?
Never assume sum, latest or maximum: nonmonotonic snapshots require authoritative ordering and
refund/correction semantics. Missing definitions mean inconclusive; ask before computing.
Do not infer formula grammar from an example name or business description.

  1. Native output first. Use the existing metric and its correct scope/aggregation.

For cases over one day append
{metric:"CaseDuration",comparisonOperator:"GreaterThan",values:["1.00:00:00"],isInclusive:true}
to F's caseMetricConditionFilters, then
get_cases_with_metrics_v2 {S,filterOptions:Fplus,itemsPerPage:5,itemsToSkip:0}.
No formula is needed for ordinary duration/waiting/counts; use native grouped statistics
instead of summing a truncated cases page.

  1. Enabled saved metric. Reuse fresh details' CustomMetrics ID/name/type/dataType.

Details is not a generic metric catalog and contains no saved formula text.
Match Case/Event/Edge scope and scalar type. For a Boolean Case metric C append
{metric:"CustomMetric",customMetricId:C,comparisonOperator:"EqualTo",values:[true],isInclusive:true}
to caseMetricConditionFilters; query get_process_overall_metrics_v2 {S,filterOptions:Fplus}
for a count. An ID filter needs no formula validation.
A saved-name output expression, e.g. synthetic EscalatedCase(), does require
the validation/column path below; do not guess escaping or missing dependencies.

  1. Exact local scalar. Reuse complete, definition-compatible, same-scope totals with a nonzero

denominator. Label the result derived, not saved. Never average averages or use overlapping
groups/top-N samples as population totals.

  1. Ephemeral formula only if necessary. Load formula execution.

Reuse PV/PMV from current details, not process/view IDs; refresh when stale.
Fetch needed targeted reference pages without a lookup quota; then validate:
validate_custom_metric_formula {processVersionId:PV,processExtendedMetadataVersionId:PMV,formula:Fexpr}.
For a candidate and its repairs, allow three validation attempts total, not endless retries.
Repair only from structured errors/definitions; stop early on an unavailable prerequisite.
Require isValid:true, a compatible dataType and the exact target context in types.
Validation alone neither executes nor saves anything.

  1. Execute the smallest matching filter/query or default page-five column query.

For cases: get_cases_with_metrics_v2 {S,filterOptions:F,itemsPerPage:5,itemsToSkip:0,formulaMetricInput:{name:N,formula:Fexpr}}.
N must not collide with returned built-ins. Never invent computedColumns or metrics.

Continue necessary pages or distinct evidence queries without ordinary-call or total-row quotas.
Reuse completed results; stop when answered or no new evidence is available.
Missing validator forbids unvalidated execution even if the output field is advertised.
Column and formula-filter capabilities are independent; stop after an explicit gate rejection.
No create/save metric tool exists. If a prerequisite cannot be inlined, explain creation/enablement
outside MCP, then refresh metadata and revalidate after the user's action.

Report definition/formula, units/context, validation state, at most five headline rows and ephemeral/not saved.
Retain definition provenance and scoped evidence/status using the investigation method.

信息
分类 数据科学
名称 derive-metric
版本 v20261008
大小 4.01KB
更新时间 2026-10-09
语言