fixed reflexion agent death spiral

[Steve-Dusty]

Problem

ReflexionAgent had the same death spiral issue as IRE agent, causing excessive iterations and API calls:

1. Fragile score extraction - Failed to parse scores from LLM responses with markdown or varied formatting
2. Never triggered early termination - Score defaulted to 0.5 when parsing failed, which never exceeded 0.9 threshold
3. Always ran full iterations - Even simple tasks used all max_loops iterations
4. Exceeded timeout thresholds - Simple tasks: 61s, Complex tasks: 203s
5. Wasted API calls - 9-15 LLM calls per task instead of 3

Root Cause

Identical to IRE agent issue:

• Early termination logic existed but depended on score extraction
• Score parsing used single fragile regex pattern: r"(?:final|overall)\s+score:?\s*(\d+(?:\.\d+)?)"
• LLM responses with markdown (**Score**: 8/10) or different formats (Rating: 8/10) failed to parse
• When extraction failed → defaulted to 0.5 → never met 0.9 threshold → ran all iterations

Solution

Applied the proven fix pattern from IRE agent with robust score extraction and improved termination logic.

1. Robust Score Extraction

Added _extract_score_robust() method with multiple fallback strategies:

def _extract_score_robust(self, evaluation: str) -> float:
    # Strategy 1: Multiple regex patterns (handles markdown, different formats)
    score_patterns = [
        r"(?:final|overall)\s+score:?\s*(\d+(?:\.\d+)?)",
        r"score:?\s*(\d+(?:\.\d+)?)\s*/\s*10",
        r"(?:rating|grade):?\s*(\d+(?:\.\d+)?)\s*/\s*10",
        r"(?:rating|grade):?\s*(\d+(?:\.\d+)?)",
    ]

    # Strategy 2: Context-aware patterns (X/10, X out of 10)
    # Strategy 3: Sentiment analysis fallback
    # Default: 0.6 (better than old 0.5)

Now handles:
- **Final Score**: 8/10 ✅
- Rating: 8.6/10 ✅
- Grade: 8 out of 10 ✅
- Markdown formatting ✅
- Sentiment-based scoring ✅

2. Configuration Constants

EARLY_TERMINATION_THRESHOLD = 0.8  # Lower than 0.9 for realistic termination
DEFAULT_SCORE = 0.6  # Higher than 0.5 to increase termination chance
SCORE_IMPROVEMENT_THRESHOLD = 0.05  # Minimum improvement to continue

3. Dual Termination Conditions

# Condition 1: Score is high enough
if current_score >= EARLY_TERMINATION_THRESHOLD:  # 0.8 instead of 0.9
    logger.info(f"✓ High score achieved ({current_score:.2f}). Stopping early.")
    break

# Condition 2: Score not improving
if iteration > 0 and (current_score - prev_score) < SCORE_IMPROVEMENT_THRESHOLD:
    logger.info(f"✓ Score improvement minimal. Stopping early.")
    break

4. Progress Logging

============================================================
Processing task 1/1
============================================================
Task: Explain photosynthesis in one sentence...

--- Iteration 1/3 ---
Evaluation complete. Score: 0.80
Iteration 1 complete | Score: 0.80 | Best: 0.80
✓ High score achieved (0.80 >= 0.8). Stopping early.

============================================================
Task complete | Iterations used: 1/3 | Best score: 0.80
============================================================

Changes

Modified Files:
- swarms/agents/flexion_agent.py - Core ReflexionAgent implementation

Key Changes:
- Line 2: Added import re
- Lines 11-14: Added configuration constants
- Lines 306-371: Added _extract_score_robust() method
- Lines 439-443: Updated evaluate() to use robust extraction
- Lines 603-673: Enhanced termination logic and progress logging


<!-- readthedocs-preview swarms start -->
----
📚 Documentation preview 📚: https://swarms--1266.org.readthedocs.build/en/1266/

<!-- readthedocs-preview swarms end -->