Errors and validation

Failures fall into caller mistakes, unsupported requests, and bad provider output. Core reports each with a distinct error so you can handle them differently.

Error	Thrown when
TypeError or RangeError	The request is invalid: non-JSON state, malformed questions, empty names, a threshold outside zero to one, a request over a declared limit, mixed questions on a model that does not support them, or an invalid concurrency.
DecisionCapabilityError	The model declares a question type unsupported.
DecisionProviderOutputError	The provider answered, but the answer violates the question contract or the usage is invalid.
AbortError	The abortSignal fired.
Provider error	The provider or network failed. Propagated as thrown, after any retries.

1. Capability errors

import { DecisionCapabilityError } from '@anvia/core/decision'

try {
  await decide({ model, state, questions })
} catch (error) {
  if (error instanceof DecisionCapabilityError) {
    console.error(error.provider, error.modelId, error.questionType)
  }
}

The error carries provider, modelId, and questionType. It is thrown before any network request.

2. Provider output errors

import { DecisionProviderOutputError } from '@anvia/core/decision'

if (error instanceof DecisionProviderOutputError) {
  console.error(error.provider, error.modelId, error.questionName, error.cause)
}

The error carries provider, modelId, the offending questionName when known, and an optional cause. See Answers and results for the checks. Log it with the request identifier but return a generic message to users, and consider routing the item to a person.

3. Batch failures

decideBatch() does not throw for per-item errors. They appear as { status: 'failed', error }, so classify them with the same instanceof checks. Cancellation is the exception: it rejects the batch.

4. Provider errors

Everything else is surfaced unchanged, for example an authentication or rate-limit error from the vendor SDK. With Jev, an SDK timeout is normalized to an error named TimeoutError so the default retry policy treats it as transient.