Flow operation flows

How flow composes operations in @pencroff-lab/kore. For error creation, wrapping, codes, metadata, and aggregation, see error-handling-patterns.md .

Every operation here is a free function over ResultTuple<T>:

type ResultTuple<T> = readonly [T, null] | readonly [null, Err];

The error slot is the discriminant. There is no container, so a result is destructured wherever it is convenient — including in the middle of a chain.

Creating results

Success

const result = ok(42);
const [value, error] = result;
// value === 42, error === null

Failure

fail("User not found", "NOT_FOUND");
fail(existingErr);                     // preserved by reference
fail(Err.from(caught, { code: "READ_FAILED", metadata: { path } }));

fail has one signature: an Err or a message, plus an optional code. Every richer construction goes through Err.from(...) and is passed in as fail(err). fail returns ResultTuple<never>, which threads through later operations with no generic claim.

Void and null success

ok();       // [undefined, null] — ResultTuple<void>
ok(null);   // [null, null]      — ResultTuple<null>

[null, null] is a valid success: the invariant is “the error slot is null”, not “exactly one slot is null”.

Error- and tuple-shaped data

ok(...) never inspects its argument, so it is the one way to carry a value that happens to look like a failure:

ok(Err.from("this is data"));   // ResultTuple<Err>
ok(new Error("this is data"));  // ResultTuple<Error>
ok(ok(1));                      // ResultTuple<ResultTuple<number>>

State checking

Destructuring narrows both branches — no type guard needed:

const [value, error] = parsePort(raw);
if (error) return error;   // error: Err
return value * 2;          // value: number

Filtering a collection maps to tuples first:

const values = results
  .filter((t): t is readonly [number, null] => t[1] === null)
  .map(([v]) => v);

Exception boundaries

attempt() — catch synchronous throws

const parsed = attempt(() => ok(JSON.parse(raw)));

The callback returns a tuple, so it can also decide the channel itself:

attempt(() => (isValid(input) ? ok(input) : fail("invalid", "VALIDATION")));

A caught value becomes Err.from(caught): a native Error keeps its stack and cause chain, a serialized Err is reconstructed, and anything else lands in the UNKNOWN branch with the value kept as metadata.originalValue.

attemptAsync() — catch throws and rejections

await attemptAsync(async () => {
  const res = await fetch(url);
  return res.ok ? ok(await res.json()) : fail(`HTTP ${res.status}`, "HTTP");
});

Transformations

map() — transform a success value

map(ok(5), (n) => n * 2);        // [10, null]
map(fail("nope"), (n) => n * 2); // the same failure, by reference

map is the only operation whose callback returns plain data. Its return is never inspected, so an Err or a tuple returned from map is the success value. Only a throw changes the channel.

flatMap() — chain a step that can fail

flatMap(ok(id), readConfig);              // a tuple-returning callee composes directly
flatMap(ok(2), (n) => ok(n * 2));         // a pure step gains an explicit ok()
flatMap(ok(2), () => fail("no", "NO"));   // an explicit failure

The callback’s tuple is returned by reference. A failure input passes through untouched and the callback never runs.

mapErr() — transform or recover from a failure

// enrich: chain the Err methods inside one callback
mapErr(result, (e) =>
  fail(e.wrap("Failed to load project").withCode("PROJECT_LOAD").withMetadata({ projectId })),
);

// recover with a value — the success type widens
mapErr(result, () => ok("fallback"));

// recover with the error itself as data
mapErr(result, (e) => ok(e));

There are no flow error-enrichment helpers; Err already owns wrap, withCode and withMetadata.

ensure() — validate, optionally refining the type

ensure(ok(port), (n) => n > 0, (n) => Err.from(`invalid port ${n}`, "PORT"));

// a type predicate narrows the success type
ensure(input, (v): v is string => typeof v === "string", () => Err.from("not a string"));

A passing predicate returns the original tuple untouched. The error factory is lazy, and a throw from either callback becomes the failure.

Async forms

mapAsync, flatMapAsync, mapErrAsync and effectAsync mirror their synchronous counterparts and additionally convert rejections. ensure has no async form.

Pipelines

pipe() — a synchronous chain

const loaded = pipe(
  ok(raw),
  onOk(parse),
  onOk(validate),
  onErr((e) => fail(e.wrap("Failed to load config", { code: "CONFIG" }))),
);

Stage factories classify a stage by what its callback receives:

A return-preserving map stage has no curried form — write it explicitly:

pipe(source, (tuple) => map(tuple, formatName));

Mid-chain recovery

Each stage is caught separately, so a later onErr recovers from a failing or throwing stage:

pipe(
  ok(1),
  () => { throw new Error("stage boom"); },
  onErr((e) => ok(e.message)),
);
// ["stage boom", null]

pipeAsync() — mixed and promised

await pipeAsync(
  Promise.resolve(ok(userId)),   // a promised source needs no extra await
  onOkAsync(fetchUser),
  onOk(normalize),               // sync stages are welcome
  onErr((e) => ok(anonymous(e))),
);

A rejected source enters the failure channel before the first stage, so a later onErr can still recover.

Typed stages

Both composers are typed through ten stages. Past ten, the trailing overload widens the result to ResultTuple<unknown>. A reusable pipeline is an ordinary typed function:

const transform = (source: ResultTuple<Input>) =>
  pipe(source, onOk(step1), onOk(step2), onErr(recover));

When to use map, flatMap or pipe

Side effects

effect(result, ([value, error]) => {
  if (error) logger.error(`Failed: ${error.message}`);
  else logger.info(`Success: ${value}`);
});

effect runs on either channel and returns the tuple it handed the callback — an output copy, not your input. Its callback’s normal return value is ignored; a throw becomes the failure.

Combinators

all() — succeed when every input succeeded

all([ok(1), ok(2), ok(3)]);             // [[1, 2, 3], null]
all([ok(1), ok("two")] as const);       // ResultTuple<[number, string]>
all([]);                                 // [[], null]
all([fail("a", "A"), fail("b", "B")]);  // one aggregate Err holding both

all is ordered and non-short-circuiting: it inspects every already-evaluated result and collects every error through addAll. Because all([readA(), readB()]) evaluates both calls before all runs, it is for independent work — it does not replace a sequential guard that must stop at the first failure.

any() — succeed on the first success

any([fail("primary down"), ok("replica")]);  // the winning tuple, by reference
any([]);                                      // failure with code EMPTY_INPUT

The winner is returned by reference, so it is the tuple the caller owns. If every input failed, the errors are aggregated.

Terminal operations

defaultTo() — a value, or a handler’s result

defaultTo(result, () => 0);                  // constant fallback, as a thunk
defaultTo(result, (e) => recover(e));        // error-aware recovery
defaultTo(result, () => fallbackFunction);   // a function-valued fallback

One handler-only signature. The handler is lazy — it never runs on the success path — and its exceptions propagate: defaultTo adds no catch boundary. A bare value as the second argument is a compile error, which is the point: the class-tier value form silently mis-resolved when T was itself a function type.

either() — fold both channels

either(
  processOrder(orderId),
  (order) => ({ status: 200, body: order }),
  (e) => ({ status: e.hasCode("NOT_FOUND") ? 404 : 500, body: { error: e.message } }),
);

Handler exceptions propagate here too.

Serialization

JSON.stringify(ok({ id: 1 }));       // [{"id":1},null]
JSON.stringify(fail("nope"));         // [null, {...Err.toJSON()}]

const [value, error] = fromJSON<Config>(JSON.parse(payload));

fromJSON validates the tuple envelope and the error slot only. T is a caller assertion, not a decode, and a success payload is never interpreted — so error-shaped data survives a round trip unchanged. It never throws:

Every rejection attaches the offending payload as originalValue metadata. ok() serializes to [null, null] and comes back as ok(null) — void does not round-trip.

Input protection

flow protects the tuple container, never the payload:

• pipe / pipeAsync copy the source once on entry, so a stage cannot mutate the caller’s tuple. Built-in operators never expose an input, so every intermediate is private to the run.
• effect hands its callback the output copy; onTuple hands it an input copy.
• map, flatMap, mapErr and their async forms receive a payload only.
• An unchanged passthrough channel is returned by reference — a failure can cross every later stage as one object.
• all always constructs its output; any returns the winning input by reference, so a caller that mutates that tuple sees the change in the result.
• Success values, Err instances, and metadata are never cloned. Use the exported copy when a custom Op needs an isolated tuple.

Real-world patterns

API handler with validation

const handle = (raw: string) =>
  either(
    pipe(
      ok(raw),
      onOk((body: string) => attempt(() => ok(JSON.parse(body) as Draft))),
      onOk((draft: Draft) =>
        ensure(ok(draft), (d) => d.title.length > 0, () =>
          Err.from("title is required", "VALIDATION:TITLE"),
        ),
      ),
      onOk(save),
    ),
    (saved) => ({ status: 201, body: saved }),
    (e) => ({ status: e.hasCodePrefix("VALIDATION") ? 400 : 500, body: { error: e.message } }),
  );

Parallel work with error collection

const [config, err] = all([readSecrets(), readFeatureFlags(), readLimits()] as const);
if (err) logger.error(err.toString());   // one aggregate holding every failure

Fallback chain

const [value] = any([fromCache(key), fromReplica(key), fromOrigin(key)]);

Async pipeline with recovery

const profile = await pipeAsync(
  ok(userId),
  onOkAsync((id: string) => attemptAsync(async () => ok(await fetchUser(id)))),
  onOkAsync(loadPreferences),
  onErr((e) => ok(defaultProfile(e))),
);

See also

• flow examples — every pattern above, executable
• Migration from throwing — converting try/catch code
• Migrating from Outcome — the class-tier operation map
• API reference