SDK doesn't handle return of control for agent actions

[v-giorgetti]

Checkboxes for prior research

• I've gone through Developer Guide and API reference
• I've checked AWS Forums and StackOverflow.
• I've searched for previous similar issues and didn't find any solution.

Describe the bug

Hi, I'm currently implementing a chatbot to help our customers execute actions, and I've setup a test agent in AWS with a single action (defined in a group action) that is triggered when the customer asks to open a ticket.
This action is defined with the option return of control, which means I should get the payload so I can handle the action the user is trying to achieve.
Though I couldn't find any documentation that is providing the ability to do that, except in Python where the SDK provides a field in the InvokeAgent command response.

Regression Issue

• Select this option if this issue appears to be a regression.

SDK version number

@aws-sdk/client-bedrock-agent-runtime@3.679.0

Which JavaScript Runtime is this issue in?

Node.js

Details of the browser/Node.js/ReactNative version

v20.15.0

Reproduction Steps

import { InvokeAgentCommand } from '@aws-sdk/client-bedrock-agent-runtime';
import { getClient } from '../client';
import { AGENT_CONFIG } from '../config';

export const askAgent = async (prompt: string, sessionId: string) => {
    const client = getClient();

    const command = new InvokeAgentCommand({
        agentId: AGENT_CONFIG.agentId,
        agentAliasId: AGENT_CONFIG.agentAlias,
        sessionId: sessionId + '-bedrock',
        inputText: prompt
    });

    try {
        let completion = '';

        const response = await client.send(command);

        // ? There is no field in the response that provide the payload of the action that is triggered

        for await (const chunkEvent of response.completion) {
            const chunk = chunkEvent.chunk;
            console.log(chunk);
            const decodedResponse = new TextDecoder('utf-8').decode(chunk.bytes);
            completion += decodedResponse;
        }
        return { sessionId: sessionId, completion };
    } catch (err) {
        console.error(err);
    }
};

Observed Behavior

There is no field in the response that provide the payload of the action that is triggered
Response type is :

export interface InvokeAgentResponse {
    /**
     * <p>The agent's response to the user prompt.</p>
     * @public
     */
    completion: AsyncIterable<ResponseStream> | undefined;
    /**
     * <p>The MIME type of the input data in the request. The default value is <code>application/json</code>.</p>
     * @public
     */
    contentType: string | undefined;
    /**
     * <p>The unique identifier of the session with the agent.</p>
     * @public
     */
    sessionId: string | undefined;
    /**
     * <p>The unique identifier of the agent memory.</p>
     * @public
     */
    memoryId?: string;
}

Expected Behavior

We should have a nullable returnControl field like the python library does

Possible Solution

No response

Additional Information/Context

No response

[aBurmeseDev]

Hi @v-giorgetti - thanks for reaching out.

It sounds like returnControlPayload is missing in your response. I was able to return it in my response and you can see the code in model file here. Can you share your response? Also when you say Python library, are you referring to boto3?

[v-giorgetti]

Hey @aBurmeseDev, thanks for answering.
I do not get the returnControlPayload when invoking agent, here's the response of the agent when it should have executed my action:

{
  "$metadata": {
    "httpStatusCode": 200,
    "requestId": "c905440f-0e00-45e0-b3b6-7f3f94e81006",
    "attempts": 1,
    "totalRetryDelay": 0
  },
  "contentType": "application/json",
  "sessionId": "playground_user",
  "completion": {
    "options": {
      "messageStream": {
        "options": {
          "inputStream": {},
          "decoder": {
            "headerMarshaller": {},
            "messageBuffer": [],
            "isEndOfStream": false
          }
        }
      }
    }
  }
}

Here is my InvokeAgent input :

let options: InvokeAgentCommandInput = {
  agentId: AGENT_CONFIG.agentId,
  agentAliasId: AGENT_CONFIG.agentAlias,
  sessionId: sessionId,
  inputText: "bla bla bla"
};

const command = new InvokeAgentCommand(options);
const response: InvokeAgentCommandOutput = await client.send(command);

Am I missing something?

I'm also adding my action configuration

[v-giorgetti]

And yes, by Python library I’m referring to boto3

[v-giorgetti]

Ok nvm, I was able to get the return control by extracting it from the response.completion, here's the snippet:

const response: InvokeAgentCommandOutput = await client.send(command);

if (!response || !response.completion) {
  return { sessionId: sessionId, completion: '' };
}

for await (const chunkEvent of response.completion) {

  if (chunkEvent.returnControl) {
    console.log('returnControl', chunkEvent.returnControl);
  } else {
    const chunk = chunkEvent.chunk;
    const decodedResponse = new TextDecoder('utf-8').decode(chunk.bytes);
    completion += decodedResponse;
  }
}

I was expecting to get it from the response as boto3 does, silly me!

You may close the ticket.