The response.mcp_call.completed is not handle

[Panha-Sim]

Description

Objective

Understand whether it is expected behavior that no audio response is triggered after an MCP call completes when using the Realtime API.

Context

We are using the Realtime API with MCP enabled.
When an MCP call completes, we receive the event:

⚠️ Unhandled message type: response.mcp_call.completed

However, no responseAudioDelta event is emitted afterward. As a result, audio playback is not triggered automatically.

To get audio output, we currently have to send another user message (for example, “Do you have a result yet?”) to trigger responseAudioDelta.

What We Observed

In OpenAIRealtimeSession.swift, there does not appear to be a handler for:
response.mcp_call.completed Reference:

SwiftOpenAI/Sources/OpenAI/Private/Realtime/OpenAIRealtimeSession.swift

Line 214 in b138a46

// MCP (Model Context Protocol) message types

Suggested Improvement

If response.mcp_call.completed were handled, we could use it to manually trigger a response using:

session.sendMessage(OpenAIRealtimeResponseCreate())

This would in theory allow us to invoke a response and start audio playback without requiring an additional user prompt.

Request

Is this the expected behavior?
If not, should support for handling response.mcp_call.completed be added?

[Panha-Sim]

For your reference here is the code:

import Foundation
import AVFoundation
import Combine
import SwiftOpenAI

@MainActor
final class RealtimeVoiceVM: ObservableObject {
    
    @Published var status: String = "idle"
    
    private let service: OpenAIService
    private var session: OpenAIRealtimeSession?
    private var audioController: AudioController?
    private var tasks: [Task<Void, Never>] = []
    
    init(apiKey: String) {
        self.service = OpenAIServiceFactory.service(apiKey: apiKey)
    }
    
    func start() {
        tasks.append(Task { [weak self] in
            guard let self else { return }
            
            do {
                let configuration = OpenAIRealtimeSessionConfiguration(
                    inputAudioTranscription: .init(model: "whisper-1"),
                    instructions: "You are a helpful AI assistant. Be concise and friendly.",
                    tools: [
                        .mcp(
                            Tool.MCPTool(
                                serverLabel: "",
                                requireApproval: Tool.RequireApproval.never,
                                serverUrl: "",
                            )
                        )
                        ],
                    turnDetection: .init(
                        type: .serverVAD(prefixPaddingMs: 300, silenceDurationMs: 900, threshold: 0.8)
                    ),
                    voice: "alloy"
                )
                
                let session = try await service.realtimeSession(
                    model: "gpt-realtime",
                    configuration: configuration
                )
                self.session = session
                
                let audioController = try await AudioController(modes: [.record, .playback])
                self.audioController = audioController
                
                self.status = "connected"
                
                // Receive messages
                tasks.append(Task { @RealtimeActor [weak self] in
                    guard let self else { return }
                    
                    await self.session?.sendMessage(OpenAIRealtimeResponseCreate())
              
                    for await message in session.receiver {
                        switch message {
                            case .responseAudioDelta(let audio):
                                audioController.playPCM16Audio(base64String: audio)
                                
                            case .inputAudioBufferSpeechStarted:
                                audioController.interruptPlayback()
                                
                            case .responseTranscriptDone(let text):
                                print("Assistant (final): \(text)")
                                
                            case .inputAudioTranscriptionCompleted(let text):
                                print("User (final): \(text)")
                            case .responseFunctionCallArgumentsDone(<#T##String#>, <#T##String#>, <#T##String#>):
                                
                            case .error(let err):
                                await MainActor.run {
                                    self.status = "error: \(err ?? "unknown")"
                                }
                            default:
                                break
                        }
                    }
                })
                
                // Stream mic -> OpenAI
                tasks.append(Task { @RealtimeActor [weak self] in
                    guard let self else { return }
                    do {
                        // micStream() is actor-isolated AND can throw when creating the stream
                        for try await buffer in try audioController.micStream() {
                            guard let base64 = AudioUtils.base64EncodeAudioPCMBuffer(from: buffer) else { continue }
                            
                            // sendMessage is likely also RealtimeActor-isolated, so this is safe here
                            try await session.sendMessage(OpenAIRealtimeInputAudioBufferAppend(audio: base64))
                        }
                    } catch {
                        Task { @MainActor in
                            self.status = "mic error: \(error)"
                        }
                    }
                })
                
            } catch {
                self.status = "start failed: \(error)"
            }
        })
    }
    
    func stop() async {
      tasks.forEach { $0.cancel() }
      tasks.removeAll()

      let audio = audioController
      let sess  = session

      await Task { @RealtimeActor in
        audio?.stop()
        sess?.disconnect()
      }.value

      audioController = nil
      session = nil
      status = "stopped"
    }
}

[Panha-Sim]

I might also be doing something wrong.

update: I also fork the repo and added the missing response.mcp_call.completed and it worked let me know if you want me to create a PR.

[jamesrochabrun]

yes please.

[Panha-Sim]

@jamesrochabrun here is the PR:
#193