Responses API

Use this page when your agent is built on OpenAI’s Responses API. The wrappers preserve the native OpenAI method names and add Averta decisions around supported calls.

Guarded Methods

Method	JavaScript	Python
`client.responses.create(...)`	Supported	Supported
`client.responses.create({ stream: true, ... })`	Supported	Supported as `stream=True`
`client.responses.stream(...)`	Supported	Supported

Non-streaming Request

const response = await client.responses.create({
  model: process.env.OPENAI_MODEL ?? "gpt-5.4-mini",
  input: [
    {
      role: "developer",
      content: [{ type: "input_text", text: "Be concise." }],
    },
    {
      role: "user",
      content: [{ type: "input_text", text: "Summarize this runbook." }],
    },
  ],
  tools,
});

response = client.responses.create(
    model=os.environ.get("OPENAI_MODEL", "gpt-5.4-mini"),
    input=[
        {
            "role": "developer",
            "content": [{"type": "input_text", "text": "Be concise."}],
        },
        {
            "role": "user",
            "content": [{"type": "input_text", "text": "Summarize this runbook."}],
        },
    ],
    tools=tools,
)

Before calling OpenAI, Averta can:

normalize request input
extract supported image Data URLs
normalize tools
return block before the provider call
return restrict_tools and remove blocked tools before forwarding

Tool Loop Continuation

When the response contains function calls, execute the tools in your app and send the results through the same wrapped client.

const functionCall = response.output.find(
  (item) => item.type === "function_call"
);

if (functionCall) {
  const continuedResponse = await client.responses.create({
    model: process.env.OPENAI_MODEL ?? "gpt-5.4-mini",
    input: [
      ...response.output,
      {
        type: "function_call_output",
        call_id: functionCall.call_id,
        output: JSON.stringify({
          matches: ["Password reset links expire after 24 hours."],
        }),
      },
    ],
    tools,
  });

  console.log(continuedResponse.output_text);
}

function_call = next(
    (item for item in response.output if item.type == "function_call"),
    None,
)

if function_call:
    continued_response = client.responses.create(
        model=os.environ.get("OPENAI_MODEL", "gpt-5.4-mini"),
        input=[
            *response.output,
            {
                "type": "function_call_output",
                "call_id": function_call.call_id,
                "output": '{"matches":["Password reset links expire after 24 hours."]}',
            },
        ],
        tools=tools,
    )

    print(continued_response.output_text)

On continuation calls, Averta evaluates new function_call_output items before OpenAI sees them.

Output Checks

If the response is final text and does not contain tool calls, Averta evaluates the output before returning it to your app. For non-streaming output:

allow returns the original result
block throws AvertaSdkError
rewrite asks OpenAI for a rewritten result and checks that rewritten output once more

Streaming

Responses streaming is guarded in both OpenAI wrappers:

const stream = await client.responses.create({
  model: process.env.OPENAI_MODEL ?? "gpt-5.4-mini",
  input: "Summarize password reset link expiration guidance.",
  stream: true,
});

stream = client.responses.create(
    model=os.environ.get("OPENAI_MODEL", "gpt-5.4-mini"),
    input="Summarize password reset link expiration guidance.",
    stream=True,
)

const stream = client.responses.stream({
  model: process.env.OPENAI_MODEL ?? "gpt-5.4-mini",
  input: "Summarize password reset link expiration guidance.",
});

with client.responses.stream(
    model=os.environ.get("OPENAI_MODEL", "gpt-5.4-mini"),
    input="Summarize password reset link expiration guidance.",
) as stream:
    for event in stream:
        ...

For streaming calls, request and tool-result checks run before the stream opens. Streaming output is checked as it is produced.

Current Limits

Responses streaming supports one output text stream per response.
Streaming output rewrite is not supported yet. If policy requires rewrite, the wrapper fails closed.
Rich media request preflight supports Data URL images only.
Remote image URLs, OpenAI file IDs, file content parts, and audio parts are rejected before preflight.

Debugging

Symptom	Check
No request decision	Confirm the OpenAI client is wrapped before `responses.create(...)` is called.
No tool-result decision	Confirm the continuation sends `function_call_output` through the wrapped client.
Missing tool after request	Check request decision `blockedTools` and dashboard tool exposure policy.
Stream fails on unsafe text	Streaming rewrite is unsupported; use non-streaming calls for rewrite behavior.

Tools

Learn how tool exposure filtering works.

Streaming

See streaming-specific limits.

Tool results

Screen returned tool content.

Output checks

Understand output blocks and rewrites.

Get Started

Core Concepts

OpenAI

Anthropic

Dashboard

Advanced

Guarded Methods

Non-streaming Request

Tool Loop Continuation

Output Checks

Streaming

Current Limits

Debugging

Tools

Streaming

Tool results

Output checks

​Guarded Methods

​Non-streaming Request

​Tool Loop Continuation

​Output Checks

​Streaming

​Current Limits

​Debugging

​Related Pages

Tools

Streaming

Tool results

Output checks

Guarded Methods

Non-streaming Request

Tool Loop Continuation

Output Checks

Streaming

Current Limits

Debugging

Related Pages