fp

The fp worker moves bulk data worker→worker. fp::pipe runs a short pipeline in one call: each step triggers a function and its result lands in the next step's payload at into (default /value); the caller receives per-step sizes and a preview — never the value itself. The ten fp::* transforms mirror their lodash namesakes, take input at value, and run inline as pipe steps, threading the transformed value itself onward — their { value } response wrapper appears only on standalone direct calls (fine for small values).

When to Use

• Fetch a document and persist it without reading it: scrapling::fetch → fp::get {path: "/content"} → fp::take {n: 20000} → state::set, all inside one fp::pipe.
• Move any big function result into another function's arguments — never re-type a large value into a call by hand.
• Reshape a list worker-side: fp::filter {matches} → fp::map {path} → fp::uniq as pipe steps.
• Top-N worker-side: fp::sortBy {path} → fp::reverse → fp::take {n} as pipe steps.
• Probe a bulk result's shape cheaply: a fp::get step with a wrong path fails naming the keys that were available.
• Slice or subset a small value directly: fp::take, fp::pick, fp::omit, fp::split, fp::join.

Boundaries

• Pipe steps run with the fp worker's authority, not the calling agent's per-step dispatch policy, so fp::pipe is not agent-callable without approval by default; the pure transforms are (see iii-permissions.yaml).
• Refused as steps: shell::*/coder::*, trigger control (engine::register_trigger/engine::unregister_trigger), nested pipes, and the agent-policy hard-denied classes (session::*/approval::*, credentials via configuration::*/oauth::*, model spend via router::*/provider::*, turn control via harness::*/run::*, bus internals) — call those directly; state::* stays allowed on purpose.
• The first pipe step receives no threaded value — start with a producer (a fetch, state::get) or seed a leading transform via payload.value.
• 1–12 steps, 120 s per bus step; the whole pipe must fit the caller's dispatch timeout. Transforms error on type mismatches instead of silently threading {}; map errors when a path matches no element (pointers pluck stored fields, not computed properties like /length).

Functions

• fp::pipe — { through: [{function, payload?, into?}], preview_chars? } → { steps: [{function, chars}], value_preview }.
• fp::get / fp::pick / fp::omit — pointer extract / key subset / key drop.
• fp::take / fp::drop — first-n / skip-n on strings and arrays.
• fp::map / fp::filter / fp::uniq — pluck / partial-object match / dedupe on arrays.
• fp::split / fp::join — string ↔ array.
• fp::size / fp::nth — count a collection / element at index (negative counts from the end).
• fp::getOr — pointer extract with a default on a miss.
• fp::compact / fp::flatten — drop null elements / unnest one level.
• fp::sortBy / fp::reverse — stable ascending sort by pointer ("" = the element itself) / reverse.
• fp::sum / fp::mean / fp::min / fp::max — fold an array of numbers to one number; path plucks the addend from each element (fp::sum {path: "/amount"} over rows). Integer inputs fold to an integer, a non-numeric element errors rather than being skipped, and an empty mean/min/max errors (an empty sum is 0). min/max return the NUMBER, not lodash's …By element, so a fp::when guard can compare it directly.
• fp::groupBy / fp::countBy — bucket or count array elements by a plucked key ("" = the element itself): { key: [elements] } / { key: count }. Per-key totals are fp::groupBy then fp::sum over a bucket. A null or container key errors rather than collapsing distinct groups into one bucket.