How do I just chat???

[itzrnvr]

Like I send a query and the model just responds. Optionally calls tools etc.

Right now, it always responds then uses task completion tool this results in an unnecessary request.

[mikhailsal]

I achieved this by creating a separate mode with the following system prompt:

You must adhere to the instructions provided by users in their messages.

Always answer in this format:
<attempt_completion>
<result>
...here is your answer
</result>
</attempt_completion>

In my case without tools, but this can be fixed.

[3C0D]

I think it's time for a complete rewrite of this application. The basics are totally broken. It's like a mad horse running straight into walls. I thought it was just because of Grok Fast (which is super bad, btw), but no. After comparing with Cline using the same model, the problem is definitely here.
It really pisses me off when applications don't work properly - they make things unnecessarily complex and then fail at the most basic tasks. Augment is super advanced to this. It has only one mode...

[3C0D]

ok thanks for your solution it seems to work. At least I can use it to chat.

[NaccOll]

RooCode is designed specifically for agents, and your request require breaking the event loop. For details, see #6637, which is a pure chat mode I wrote, but it clearly deviates from the existing RooCode design direction.

[3C0D]

direction right in the wall

[3C0D]

If an LLM is not able to have a conversation without doing some edit before or after to correct its actions, you will build only some one-shot apps. I guess you never tried Augment or Trae or even Kilocode. You can talk and then ask to act, or ask to act directly. It will precisely do what you are asking to do without tricky system of prompts. You just have to toggle confirm before edit or not... Very simple, very efficient. A lot of options doesn't mean an efficient app, it means potentially a lot of bugs difficult to fix and an absolute incapacity to improve real efficiency.
And on top of that, the worst part is that they multiply the profiles but there are no common instructions across all the profiles. For example, speak in French, put the code in English. Because whether it's editing or just chatting, the response should be formatted the same way. It's not the operation behind that matters. After that, if we want to make specialized prompts, why not? But first, master the basics.

[3C0D]

And during modification, we can't see a diff showing the modifications. We're forced to wait until the end of operations to see a diff, when it's often too late. Because on top of that, most other programs allow you to interrupt and resume, to make even more modifications to the action in progress. So if we say stop, it does that instead. Here, what happens? It keeps trying to do the same thing. We tell it, "Explain this code to me," and it tries to modify it. We're completely backwards here.

[NaccOll]

I've used all of the things you mentioned. Kilo is even a fork of RooCode.

The inability to see diffs you mentioned, I suspect you have enabled experimental features, such as background or concurrent editing, which will prevent you from seeing the changes in the process. At this time, you can only browse diffs through checkpoints.

I understand some of your confusion. I prefer to manually build the complete task context. After completing the context analysis, I switch from chat mode to agent mode, allowing LLM to invoke tools to view or edit more files.. So my tasks usually start with simple Q&A, and at that point, I don't want LLM to invoke any tools. It might misunderstand my meaning and query a bunch of irrelevant files, draining my wallet and messing up the task.

This is not the direction of RooCode . Perhaps you could check out roomote. RooCode is moving towards a non manual programming mode, where users complete tasks through natural language communication. Of course, this requires top-tier models and rich money.

[3C0D]

An LLM allows you to add intelligence to an action, modulated with discernment. Whereas here, the program constrains the LLM to react in a certain way, which ultimately breaks the possibilities offered by the LLM. And when you try Claude under Augment, you realize that wow, it goes very far. But well, they also lack control over the visible result of actions. Trae, which is less polished and yet better on this point. As well as Copilot.