Providers and resilience

GuardProvider

interface GuardProvider {
  name: string
  complete(params: GuardCompletionParams): Promise<string>
}

A provider receives chat-style messages and returns raw text. Detectors own prompt construction and JSON parsing; providers should stay thin.

Built-in providers

Provider	Class	Notes
OpenAI chat.completions	OpenAIProvider	Also covers OpenAI-compatible gateways through baseURL; json: true maps to JSON mode.
OpenAI Responses API	OpenAIResponsesProvider	Maps system messages to instructions and reads output_text / message output text.
Anthropic	AnthropicProvider	Moves system messages into system; JSON behavior is prompt-enforced.

Custom provider

const provider: GuardProvider = {
  name: 'my-gateway',
  async complete({ messages, model, temperature, signal }) {
    const text = await callMyGateway({ messages, model, temperature, signal })
    return text
  },
}

Return only the model text. Throw the original network/HTTP error so withResilience can classify and retry it.

Resilience wrapper

const resilient = withResilience(provider, {
  retries: 2,
  delayMs: 200,
  timeoutMs: 10_000,
  onRetry: (attempt, error) => metrics.increment('provider_retry', { attempt }),
})

• Retries use exponential backoff and do not consume detector logic.
• Abort errors are not retried.
• Timeouts abort the inner provider through AbortSignal and also win a race, so slow providers that ignore signals still fail the call.