[Feature] Critic/Refinement System for Agent Improvement

[VascoSch92]

Summary

Add a critic system that evaluates agent outputs and triggers iterative refinement when quality is below threshold.

Feature Details (from OpenHands-CLI)

Critic System

• Built-in critic that evaluates agent task completion
• Configurable success threshold (default: 0.6 / 60%)
• Issue detection threshold (default: 0.75 / 75%)
• Maximum refinement iterations (default: 3)

Issue Detection

• Insufficient testing detection
• Loop behavior detection
• Agent behavioral issues
• Categorized feature analysis

Iterative Refinement

• Automatic follow-up prompts when score is low
• Build refinement messages with specific feedback
• Track refinement iterations per user turn
• Stop when max iterations reached or quality improves

Configuration

• enable_critic: bool - Toggle critic on/off
• enable_iterative_refinement: bool - Toggle auto-refinement
• critic_threshold: float - Overall success threshold (0.0-1.0)
• issue_threshold: float - Individual issue threshold (0.0-1.0)
• max_refinement_iterations: int - Max iterations (1-10)

TUI Integration

• Critic settings tab in settings modal
• Critic notice in splash screen
• Display critic scores in conversation

Reference

See OpenHands-CLI implementation in:

• openhands_cli/tui/utils/critic/refinement.py
• openhands_cli/tui/utils/critic/feedback.py
• openhands_cli/stores/cli_settings.py (CriticSettings)

---

This issue was created by an AI assistant (OpenHands) based on a feature comparison with OpenHands-CLI.

[VascoSch92]

📋 Tracked in: #18 (Feature Parity with OpenHands-CLI)