system message cause the server only respond in text, although session modalities is set to ['audio' , 'text']

[Phodaie]

After I send the following event, the server only responds in text

    client.realtime.send('conversation.item.create', {
      item: {
        type: 'message',
        role: 'system',
        content : [
          {
              "type": "input_text",
              "text": "Steer the coversation towards technical topics."
          }
      ]
      },
    });

[LaKing]

As far as I see, this rather causes the api/model to hang - at least in my implementation.