create-metric
Create an FME metric (fme_metric): resolves traffic type and event type IDs first, drafts the payload for confirmation, then creates. Doesn't attach metrics to experiments or add the tracking call - see instrument-metric. Trigger phrases: create/define a metric, new metric.
Harness · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add harness/harness-skills --skill create-metric技能说明
name: create-metric description: >- Create an FME metric (fme_metric): resolves traffic type and event type IDs first, drafts the payload for confirmation, then creates. Doesn't attach metrics to experiments or add the tracking call - see instrument-metric. Trigger phrases: create/define a metric, new metric. metadata: author: Harness version: 1.2.0 mcp-server: harness-mcp-v2 license: Apache-2.0 compatibility: >- Requires Harness MCP v2 server (harness-mcp-v2) with fme_metric (create/update/delete), fme_traffic_type, and fme_event_type. fme_metric.delete is a hard delete with no archive/restore.
Create Metric
Create an FME metric definition, resolving every reference (traffic type, event type, owners) against real data first instead of guessing IDs.
For every step below marked Stop condition, read
references/stop-conditions.md before improvising - each covers a case
where a plausible guess is wrong often enough that presenting real options
and waiting for a pick is required, not optional.
Prerequisites
Establish org_id + project_id. If /choose-metric was already run and
found no suitable metric, reuse its context (hypothesis, traffic type)
instead of re-asking.
Instructions
Step 1: Check for an existing metric first
Call MCP tool: harness_list
Parameters:
resource_type: "fme_metric"
org_id: "<org_id>"
project_id: "<project_id>"
filters: { name: "<proposed name, substring>", limit: 20 }
Filter fields go under filters, not top-level - passed top-level they're
silently ignored and every metric comes back unfiltered. limit belongs in
filters too; harness_list's top-level size maps to the same thing, so
don't send both. (Same for
traffic_type in Step 3). Keep name as specific as the request allows:
these are full metric definitions, so a broad substring like purchase can
return dozens of them.
The backend returns 409 for a duplicate name and for a duplicate definition (same attribute combination) under a different name - catching this early saves a round trip. If something close already exists, confirm a new metric is actually needed rather than reusing/updating the existing one.
Stop condition if the request is vague (e.g. "track checkouts" with no
aggregation specified) - see references/stop-conditions.md.
Step 2: Resolve trafficType
Call MCP tool: harness_list
Parameters:
resource_type: "fme_traffic_type"
org_id: "<org_id>"
project_id: "<project_id>"
Use the traffic type name (immutable after creation, like name
itself). Don't ask the user for an ID - resolve it from this list.
Step 3: Resolve baseEventTypes[].eventTypeId
Call MCP tool: harness_list
Parameters:
resource_type: "fme_event_type"
org_id: "<org_id>"
project_id: "<project_id>"
filters: { traffic_type: "<traffic type name or id>", name: "<event name, if the user named one>" }
Add name whenever the user named the event; drop it only when you need
the full list to show what's available. Both are substring/ID matches.
Stop condition if the event isn't in this list, or if several events
match and none is an exact match for what the user said - a substring search
on a word like purchase routinely returns a dozen variants, so present the
real ones and ask instead of picking the shortest or cleanest-looking name.
Don't invent an ID and
don't silently substitute the closest-looking real event either; see
references/stop-conditions.md.
Step 4: Decide aggregation, format, spread, isPositive
aggregation:COUNT(count of events),TOTAL(sum of event values),AVERAGE(average event value),RATE(unique units that did the event).TOTAL/AVERAGEsum thetrack()value, or the property named inbaseEventTypes[].propertyForValue.spread: not part of the create body - every new metric isPER, so don't put it in the draft.PER= computed per unit, then compared across treatments (RATE+PER= percent of users who converted,COUNT+PER= events per user).ACROSS(one aggregate over the whole treatment) is a deprecated legacy value with no create path and no significance test in experiment results - don't offer it as a choice. ExistingACROSSmetrics can still be read and patched viafme_metric.update.format:NUMBER,DOLLAR,PERCENTAGE,SECONDS,MILLISECONDS,BYTES- display only.NUMBERis the right default when the request implies no unit.RATE+PERalways reads back asPERCENTAGE;PERCENTAGEon anything else reads back asNUMBER.isPositive:trueif an increase is a good outcome. Get this right - it drives significance-direction interpretation downstream (see/review-experiment-results).
Step 5: Name and owners
-
name: must match^[A-Za-z][-_A-Za-z0-9]*$, max 100 chars, immutable. No spaces - if the user gives a human title like "Checkout Conversion", derive a validname(e.g.checkout_conversion) and confirm it rather than sending the title verbatim and hitting an unhelpful 400. -
owners: required for now - the backend rejects an empty/missing list with 400"Owners cannot be empty"until the planned owners deprecation ships. Each entry is{type: "USER", email: "..."}or{type: "GROUP", identifier: "..."}. AGROUPidentifieris matched against the ids of the groups in the project - theidan existingGROUPowner shows on read, not its display name. An unmatched identifier fails withgroup owner not found.Stop condition if no owner was specified, or one turns out to be invalid - see
references/stop-conditions.md.
Step 6: Optional fields
filterEventType-{eventTypeId, filterAggregation, propertyFilters}, resolved the same way as Step 3. Two uses, selected byfilterAggregation:"RATE"- HAS_DONE filter: only count units that did this event."COUNT"with `aggregation: COUN