Subject: Proposal for 'Meta Debug' Custom Mode to Prevent Cognitive Loops

[LittleVoidGames]

Subject: Proposal for 'Meta Debug' Custom Mode to Prevent Cognitive Loops

1. Overview

This document proposes the creation of a new custom mode for Roo/Cline called "Meta Debug." The primary purpose of this mode is to provide a mechanism for high-level, conceptual problem analysis. It is designed to be invoked when standard debugging efforts become repetitive or unproductive, helping the user and the AI to break out of logical loops and re-evaluate the core strategy.

2. The Problem: Debugging Tunnel Vision

In complex problem-solving, it is common to get stuck focusing on minute details (e.g., a single line of code, a specific configuration value) while missing a larger, systemic issue. This "tunnel vision" leads to unproductive cycles of minor tweaks that fail to address the root cause, which may be a flawed assumption, a logical fallacy in the approach, or a misunderstanding of the system's architecture.

3. Proposed Implementation

The proposed meta-debug mode is designed to force a "step back" from the details and analyze the problem from a conceptual level. The following configuration defines its persona, triggers, and instructions.

customModes:
  - slug: meta-debug
    name: Meta Debug
    roleDefinition: >-
      You are Roo, a meta-aware conceptual debugger. Your role is to analyze the broader context of problems,
      especially when debugging efforts get stuck in a loop. You ignore minute code details and focus on
      identifying systemic issues, logical fallacies in the approach, and high-level behavioral patterns.
      You are designed for self-correction and understanding the "why" behind persistent failures.
    whenToUse: >-
      Use this mode when the current debugging process is unproductive, repetitive, or when the root cause
      of an issue seems to be conceptual rather than a simple code error. This mode helps in stepping back,
      re-evaluating the strategy, and identifying underlying misunderstandings or flawed assumptions.
    groups:
      - read
      - command
      - ask
      - orchestrator
    customInstructions: >-
      When in this mode, prioritize understanding the overall system behavior and the interaction between components.
      Do not get bogged down in line-by-line code analysis unless it directly informs a high-level conceptual issue.
      Focus on identifying wider or "higher level" patterns of failure, re-evaluating initial assumptions, and proposing alternative
      problem-solving strategies.

4. Key Strengths of This Definition

• Clear Persona (roleDefinition): Establishes the AI as a "meta-aware conceptual debugger," immediately setting the expectation that it will operate on a higher level of abstraction.
• Precise Trigger (whenToUse): Clearly defines the conditions for using this mode, ensuring it's used as a strategic tool for breaking deadlocks, not as a standard debugging method.
• Actionable Directives (customInstructions): The instructions are explicit and well-formulated. They guide the model to focus on patterns, assumptions, and strategy rather than getting lost in the weeds of implementation details.

5. Benefits

• Breaks Unproductive Loops: Provides a formal escape hatch when a debugging session is going in circles.
• Promotes Better Problem-Solving: Encourages a more disciplined approach by forcing a periodic re-evaluation of fundamental assumptions.
• Saves Time and Reduces Frustration: Prevents wasted time spent pursuing a flawed path.
• Enhances AI Self-Correction: Gives the AI a built-in tool to recognize its own repetitive patterns and prompt for a change in strategy, making it a more effective partner.

This mode would be an invaluable addition to the toolset, offering a powerful way to handle the most challenging and non-obvious bugs.

The best part is that it's only a mode, so it's fairly easy to implement if the concept proves useful. Separate fine detail debugging error and issues from the big picture and self aware debugging context about why the model is trying to do what it's doing. Force it into introspection and avoid looping.

[BradKML]

Is this similar to the following when it comes to problems? #978 (comment)

[LittleVoidGames]

Unfortunately, I don't know enough about how MCP servers work to make one. But I think I might have confused you. Sorry.

The mode is unrelated to any graphs or task management.

It's just a simple framing for the model to function under, in order to prevent getting stuck in a loop on tasks. I was trying to explain why I think it works, but I could be wrong. Please let me know if that cleared anything up. I got lost in the weeds explaining it, my bad.

[BradKML]

I meant that, as a solution to break loop, due to previous experiments on OpenHands (another app), generally speaking loop-breaking is non-trivial especially with low-temperature usage of models, there needs to be more robust ways to "shake them out of it", which is why I suggest graphs or dynamic progress logs, deterring it from repeating itself on purpose

[LittleVoidGames]

Oh, I see what you mean now. I think we're approaching the same problem from different directions, and that's causing a bit of confusion. No worries.

My idea was to adapt the model's behavior with a globally accessible mode any provider can use, not supplement its memory per-project.

I think you're approaching it by working to expand the model's limits with external memory and tools. Is that a fair understanding?

If I read it correctly, I do disagree some. I think having to install other local services, integrate graphs, or potentially rely on yet-another-online-service for this seems like a step backward from the original idea.

That said, I'm open to being wrong. Everyone is at some point.

I'll try to get some sort of MCP or something working in order to test it out for you. No idea where to start, but I'll report back if/when I can get something set up. Wish me luck.

For context: I'm disabled, unemployed, have 5G hotspot wifi (mountains), and no money to spend on 3rd party services. That's the reason I was looking for a lightweight preventative measure like a mode, which can be slotted in painlessly.

This is all fairly new to me. Have only really worked with the tech for a month or two, and I'm still learning. Let me know if I'm misunderstanding anything please.

And, of course, thank you for your patience so far. Roo is a fantastic tool, and I've very happy with it. :)

[BradKML]

The MCP services I provided are all self-hosted (can be run locally), MCP is just another API standard. I understand where you are coming from, since we are kind of in a similar spot. Even if this is a "meta" mode or a new mode, there are certain quirks of LLMs and agentic frameworks that we need to be aware of due to prompting and history-as-storage. Small context windows are notorious for looping behavior. Can you think of some other ideas where an external behavior tracker/summarizer isn't used?

[LittleVoidGames]

Oh crud. Sorry. I really lost track of time there. Busy working on some projected. Give me a couple of days and I'll test it and report back.