Skip to main content
For non-streaming Anthropic Messages calls, Averta checks final assistant text before the wrapped client returns it to your application. Output checks run only when the response does not contain tool_use blocks. If the model asks for a tool, the wrapper checks the tool call and returns control to your app for the native Anthropic tool loop.

​
Example

const message = await client.messages.create({
  model: process.env.ANTHROPIC_MODEL ?? "claude-sonnet-4-5",
  max_tokens: 512,
  messages: [
    {
      role: "user",
      content: "Write a customer-safe summary of the account recovery steps.",
    },
  ],
});
Before this call resolves, the wrapper extracts text content blocks from the assistant response and sends the combined text to the output checkpoint.

​
Outcomes

DecisionWrapper behavior
allowReturns the original Anthropic response.
blockThrows AvertaSdkError.
rewriteAsks Anthropic for a rewritten answer, then checks the rewritten output once more.

​
Rewrite Behavior

For rewrite, the wrapper builds a new non-streaming Anthropic request that:
  • keeps the original conversation
  • appends the unsafe assistant response
  • appends an Averta rewrite instruction as the next user message
  • removes tools, tool_choice, and stream
The rewritten answer is checked again with rewrite_attempt: 1. If that second output is blocked, the wrapper throws.

​
Decision Callback

client = wrapAnthropic(client, {
  onDecision(event) {
    if (event.checkpointType === "output") {
      console.log(event.rewriteAttempt, event.decision.decision);
    }
  },
});

​
Limits

Streaming is not supported in the current Anthropic wrapper, so output checks and rewrites are available only for non-streaming messages.create(...) calls.