Transforms

Transforms are pure, deterministic, local, and free. Parameters sit flat on the step alongside id, transform, skip_when, and fail_on — there is no wrapper object. Use the parameter names below exactly. A trailing ? marks an optional parameter; {{item.*}} is valid only inside map expressions and foreach iterations.

{ "id": "days", "transform": "date_diff",
  "from": "{{step.deal.last_activity_date}}", "to": "now", "unit": "days" }

#Collections

reduce operators: sum, count, avg, min, max, concat, or, and. filter conditions default to AND; mode: "any" makes them OR.

map has three expression forms — a single ref (extracts, preserving type), a math expression over {{item.*}} (+ - * / %, min, max, abs, ceil, floor, round), or an object template. output_field attaches the result to each item instead of replacing it:

{ "id": "weighted", "transform": "map", "array": "{{step.deals}}",
  "expression": "{{item.amount}} * {{item.stage_probability}} / 100",
  "output_field": "weighted_amount" }

group_by with aggregate computes per group and returns an array:

{ "id": "by_stage", "transform": "group_by",
  "array": "{{step.deals}}", "field": "stage",
  "aggregate": {
    "count": { "operator": "count" },
    "total": { "field": "amount", "operator": "sum" }
  } }

Without aggregate it returns { key: [items] } — chain to_list then map to iterate groups.

#Objects

#Strings

#Encoding and size

Useful when an operation has a payload limit, or when you need a deterministic id so that re-running is a replay rather than a duplicate.

#Numbers

math operators: add, subtract, multiply, divide, modulo, and the unary abs, ceil, floor.

#Dates

Units: days, hours, minutes, seconds. date_period names: today, yesterday, this_week, last_week, this_month, last_month, this_quarter, last_quarter, this_year, last_year, last_7_days, last_30_days, last_90_days, next_7_days, next_30_days — inclusive UTC day bounds, ISO weeks, computed from the run clock so scheduled runs stay deterministic.

#Timestamps

#Logic

default treats the empty string as missing but preserves 0 and false — use it over coalesce when a templated value may interpolate to "".

#Privacy

hash_replace before an AI step and hash_restore after it keeps chosen fields out of the model's sight; the mapping never leaves the run.

#Display

Checklist items are { label, issue, detail_ok, detail_issue, detail_null } where issue is a condition string — true renders as a problem, false as fine, null as unknown. Table column formats: text, currency, number, date, relative_date, boolean, percent. Checklist items and table columns (type: "action") may carry action descriptors; actions always render as owner-clicked controls that open a normal, still-governed recipe run — never a silent execution. to_csv escapes per RFC 4180; pair it with a file write to produce a spreadsheet-openable artifact.

#Warehouse

enrichment-or-fetch is the read-through helper for precomputed facts: it returns the stored enrichment when one exists, is fresh, and meets the trust floor — and signals your fallback step to fetch fresh otherwise.

Returns { value, source, confidence?, fallback? } where source is enrichment, miss, stale, low_trust, unparseable_ref, or no_runtime.

#Related

• Recipe schema — the step model these run inside.
• Authoring recipes — when to reach for a transform versus an AI step.