Issues with AI Agent in Server Connect

[karh]

My feedback, just started using it for a bit.

1. Please remember model selection, I just opened a server action started a chat. Realised i'm in the wrong server action. Then I gotta find my model again through this huge list
   

   

2. I asked it to change something, I now have no idea if it did it or not. Agent mode is on and it said "I will remove this block" then stopped the chat.

According to git this file has not been changed.

After pressing the AI to change the file for me it is just telling me I'm paying money and not giving feedback. I guess it's encountering errors and not displaying it to me?

I did a last attempt with claude 3.7 sonnet but now do get an informative error

If it's usable then Keith's feedback sounds great as to make it more useful and possibly replace my windsurf use.

Edit: tried one more time with github as provider opposed to openrouter and that seems to do stuff and give me feedback, seems buggy with openrouter

3. Having some kind of reference we can visually use would be great, like line numbers... it doesn't tell me much that action step j1_49 is changed:

4. It seems unable to add any action steps:

{
  "name": "insert_action_step",
  "args": {
    "action_step": {
      "name": "",
      "module": "core",
      "action": "comment",
      "options": {
        "comment": "User update logic removed; now only insert if user does not exist."
      }
    },
    "insert_pos": "after",
    "ref_node": "j1_64"
  }
}

Invalid action_step. Please provide a valid JSON strinn that has name, module, action and options.

[patrick]

How is your experience with the latest version, do you still have many of the issues you mentioned here?

We do now remember model selection and the list is shorter since we filter them now showing only the models that support tools and have a large enough context window.

There were several updates to the tools and instructions and will be more finetuned in upcoming releases. We don't have full control over the AI output and it can hallucinate sometimes.

[karh]

Just gave it a single shot, but sadly not much better:

I paid $0.43 for Wappler to crash and when opening it again seeing that it didn't even do anything other than reading the json file.

Edit: I then switched to use github copilot, i tried to continue it with the default chatgpt 4o: Error: 400 bad request

Then I tried to switch to claude 3.5 sonnet:
Error: 400 The requested model is not supported.

It has only worked to start a new chat and use GPT 4.1

Which then provided me inaccurate history in the chat:

Then it begs the question: with which chat window should I be chatting? The manager or the agent?

I asked it to change it, to the manager, and it supposedly replaced the action step from 'Throw error' to an 'end' step. However I don't see that in my actual editor.

---

Sadly it feels far from usable

[George]

If you just need to edit a single server connect file, use the local AI Assistant inside the SC edition.

The AI manager is for creating multiple files and it will delegate the work to the dedicated editor AI assistants which are specialized with the knowledge of their specific editor.

If you mind the cost then just use github copilot as a AI model provider in Wappler, it is a fixed $10 monthly only.

[karh]

In my edits you can see I switched back to github copilot. That doesn't sadly fix the other issues (unsupported models, it telling me it did edits but didn't actually do it)

[George]

As for unsupported models make sure you have GitHub copilot pro and the models are enabled in your GitHub profile settings / copilot