Signals and outcomes

A signal is a telemetry envelope your app sends so FireWeave can see how a release is going. The spec defines four kinds only: health, error, metric, and outcome. OpenFeature and the wire still say flagKey when a signal names a control point. Capture uses type: "signal". These are not the in-app Log / Alert / Block severities.

Why it exists

Evaluation says what value was chosen. Signals say what happened next — a component is degraded, an error you already classified, a numeric observation, or a user-visible result. Messages are secret-redacted. Attributes pass an allowlist so arbitrary PII does not go on the wire.

When to use it

Kind	Use when
health	A component’s status (`ok`, `degraded`, …)
error	A failure you already classified (`errorKind` + redacted `message`)
metric	A named observation (`value` + optional `unit`)
outcome	A user-visible or release-level result (checkout completed, release completed)

Do not invent extra kinds. Do not map console Block onto recordError.

How it relates

Concept	Relation
Releases	Node `releases.complete` / `fail` also record an outcome signal. That is not a substitute for `recordOutcome` in other languages
Exposures	Different capture type. Exposures prove assignment; signals report aftermath
Capabilities	`signals.recordHealth` / `recordError` / `recordMetric` / `recordOutcome`
Adapters	Remote may batch to `POST /v1/capture`. Some paths stay in-process

Whether signals always reach fw-server is NEEDS VERIFICATION (compatibility known gap: Go/Java sink vs Node/Python in-process record). Check capabilities.get().runtime.features and do not claim universal delivery.

Shared envelope

Required by spec/signal.schema.json: kind + name (1–256 chars). Optional correlation: targetingKey, rolloutId, changeId, stampId, flagKey, variant. Optional status, errorKind, message, value, unit (language-specific), attributes.

Calls are opt-in per invocation — nothing is emitted until you call a record* method.
Empty name is rejected (InvalidContext on Node).
Extension lifecycle gate: degrade before READY / after shutdown; never throw (Go returns error).
Node/Python keep an in-process list (getRecorded() / signals.recorded) useful for tests.

Health

Purpose: “Is this component ok?” Typical fields: name, status (spec examples: ok, degraded).

Node
Python
Go
Java

client.signals.recordHealth({ name: 'checkout-api', status: 'healthy' });

client.signals.record_health(
    "checkout-api",
    "healthy",
    rollout_id="rollout_01HZXEXAMPLE000000000001",
)

err := client.Signals().RecordHealth(ctx, fireweave.HealthSignal{
    Name: "checkout-api", Status: "ok",
})

client.signals().recordHealth("checkout-api", "ok");

Error

Purpose: A classified failure. errorKind uses the SDK’s 15-kind taxonomy (for example Timeout). message is redacted.

Node
Python
Go
Java

client.signals.recordError({
  name: 'checkout-api',
  errorKind: 'Timeout',
  message: 'upstream deadline exceeded',
});

client.signals.record_error(
    "checkout-api",
    error_kind="Timeout",
    message="upstream deadline exceeded",
)

err := client.Signals().RecordError(ctx, fireweave.ErrorSignal{
    Name: "checkout-api", ErrorKind: fireweave.KindTimeout, Message: "...",
})

client.signals().recordError("checkout-api", ErrorKind.Timeout, "...");

Metric

Purpose: A named observation. Spec value is number, boolean, or string. Python’s helper types value as float. Optional unit.

Node
Python
Go
Java

client.signals.recordMetric({ name: 'p99_latency_ms', value: 187, unit: 'ms' });

client.signals.record_metric("p99_latency_ms", 187.0, unit="ms")

err := client.Signals().RecordMetric(ctx, fireweave.MetricSignal{
    Name: "p99_latency_ms", Value: 187,
})

client.signals().recordMetric("p99_latency_ms", JsonValue.of(187));

There is no SDK type named “adoption signal.” Use recordMetric or recordOutcome.

Outcome

Purpose: Record that a result happened. This is a signal kind, not its own product object and not a separate docs page. signals.recordOutcome / record_outcome / RecordOutcome takes a name + status and optional rollout correlation (rolloutId, changeId on Python). Node releases.complete() / fail() also records { kind: 'outcome', name: 'release', status: 'completed' | 'failed' }. Completing a release is still a different call from recording an application outcome (for example checkout completed).

Node
Python
Go
Java

client.signals.recordOutcome({
  name: 'checkout',
  status: 'completed',
  rolloutId: 'rollout_01HZXEXAMPLE000000000001',
});

client.signals.record_outcome(
    "checkout",
    "completed",
    rollout_id="rollout_01HZXEXAMPLE000000000001",
)

err := client.Signals().RecordOutcome(ctx, fireweave.OutcomeSignal{
    Name: "checkout", Status: "completed",
})

client.signals().recordOutcome("checkout", "completed");

Whether console Completed / Rolled back consume these signals is NEEDS VERIFICATION. Node also exposes generic signals.record({ kind, name, ... }). Java has signals().record(Signal).

Attribute allowlist

Unknown attribute keys are dropped, not rejected. String values are secret-redacted. Node default (DEFAULT_SIGNAL_ATTRIBUTE_ALLOWLIST, on unless you override telemetry.attributeAllowlist) and Python _SIGNAL_ATTRIBUTE_ALLOWLIST: name, kind, status, value, unit, rolloutId, changeId, stampId, errorKind, message, flagKey, variant, environment, service Go’s emission allowlist is slightly different: flagKey, variant, value, rolloutId, changeId, stampId, stampIds, status, name, kind, errorKind, message, metricValue. Java signals use a fixed canonical field set; FireweaveConfig.telemetryAttributeAllowlist can filter further. Do not send arbitrary PII as signal attributes and expect it to survive.

Batching and failure

Signals are recorded when you call record*. Remote adapters capture via POST /v1/capture when a sink is attached.
Node calls adapter.recordSignal when present. Python calls deliver_signal and swallows sink exceptions.
Telemetry loss must not break evaluate or login paths.
After shutdown, calls degrade with AlreadyClosed.

Releases — lifecycle vs recordOutcome
Exposures — assignment proof
Errors — the 15 errorKind values
Quickstart — record an outcome in the first-hour path

Get Started

SDKs

Concepts

OpenFeature

Production

Testing

Troubleshooting

Migration

Reference

Signals and outcomes

Why it exists

When to use it

How it relates

Shared envelope

Health

Error

Metric

Outcome

Attribute allowlist

Batching and failure

​Why it exists

​When to use it

​How it relates

​Shared envelope

​Health

​Error

​Metric

​Outcome

​Attribute allowlist

​Batching and failure

​Related

Why it exists

When to use it

How it relates

Shared envelope

Health

Error

Metric

Outcome

Attribute allowlist

Batching and failure

Related