core/metacognition/

Overview

The metacognition/ submodule provides self-reflection, loop detection, and prompt management for the agent.

Files

• metacognition.py: Contains the MetaCognition class for internal monologue, task analysis, and intelligent stopping decisions.
• loop_detector.py: Implements the LoopDetector class for detecting and breaking repetitive loops in agent behavior.
• prompts.py: Centralizes all prompt templates and formatting logic for system, metacognitive, and loop-breaking messages.

Key Classes

• MetaCognition: Analyzes tasks, reflects on actions, and decides when to continue or stop.
• LoopDetector: Detects exact and semantic repetition, confusion, and no-action loops.

Responsibilities

• Enable the agent to reflect on its own progress
• Detect and break out of unproductive loops
• Centralize and manage all prompt templates

How Metacognition Works

1. Task Analysis: The agent thinks about what the user really wants and what success looks like.
2. Action Reflection: After each step, the agent reflects on what it just did and how much progress it made.
3. Loop Detection: The agent checks if it's repeating itself or getting stuck.
4. Stopping Decision: The agent decides if it should keep going or stop, based on progress and confidence.

Example: Avoiding a Loop

• The agent notices it's giving the same answer multiple times.
• It triggers a loop-breaking prompt and tries a new approach or stops.

In Plain English

• This part of the agent is like an inner voice: it thinks about the task, checks its own work, and knows when to stop or try something different.