Intervention Detector Skill
Purpose
Detect problematic patterns in agent performance during CFN Loop iterations that may require intervention.
Detection Mechanisms
Confidence Plateau:
- Minimal confidence improvement over multiple iterations
- Threshold: Δ < 0.05 for 2+ consecutive iterations
Recurring Feedback:
- Same feedback theme repeated 3+ times
- Indicates underlying issue not being addressed
Deliverables Stuck:
- No new files created in 2+ consecutive iterations
Configuration
- Configurable thresholds
- Flexible input parsing
- JSON output for easy integration
Usage
./detect-intervention.sh \
--iteration N \
--confidence-history "0.72,0.75,0.76" \
--feedback-history "theme1;theme2;theme3"
Output Formats
- Trigger type
- Detailed reasoning
- Recommended action
Implementation Notes
- Use bash for performance
- Minimize external dependencies
- Provide clear, actionable output