Function: wrapOutputGuardrails()

wrapOutputGuardrails(service, context, stream, options): AsyncGenerator<AgentOSResponse, void, undefined>

Defined in: packages/agentos/src/safety/guardrails/guardrailDispatcher.ts:289

Wrap a response stream with guardrail filtering.

Creates an async generator that evaluates each chunk through registered guardrails before yielding to the client. Supports both real-time streaming evaluation and final-only evaluation based on guardrail configuration.

Evaluation Strategy:

• Guardrails with config.evaluateStreamingChunks === true evaluate TEXT_DELTA chunks
• Guardrails with config.evaluateToolCalls === true evaluate TOOL_CALL_REQUEST chunks
• All guardrails evaluate the chunks that carry isFinal: true (the FINAL_RESPONSE, an ERROR); other chunks pass through unevaluated
• Rate limiting via config.maxStreamingEvaluations per guardrail

Actions:

• GuardrailAction.BLOCK - Terminates stream immediately with error chunk
• GuardrailAction.SANITIZE - Replaces chunk content with modifiedText; on a TOOL_CALL_REQUEST chunk it is applied as a BLOCK
• GuardrailAction.FLAG / GuardrailAction.ALLOW - Passes through

Parameters

service

Single guardrail or array of guardrails

IGuardrailService | IGuardrailService[] | undefined

context

GuardrailContext

Conversation context for policy decisions

stream

AsyncGenerator<AgentOSResponse, void, undefined>

Source response stream to wrap

options

GuardrailOutputOptions

Stream options and input evaluations to attach

Returns

AsyncGenerator<AgentOSResponse, void, undefined>

Wrapped stream with guardrail filtering applied

Example

// Wrap output stream with PII redaction
const safeStream = wrapOutputGuardrails(
  [piiRedactor, contentFilter],
  guardrailContext,
  orchestratorStream,
  { streamId: 'stream-123', inputEvaluations }
);

for await (const chunk of safeStream) {
  // Chunks are filtered/sanitized before reaching here
  yield chunk;
}