What problem does it solve? AI routing systems make static decisions that never improve from past outcomes. This Skill closes the loop by recording every routing decision, batch-learning from outcomes with GRPO, and transferring validated patterns between fast (System 1) and deep (System 2) caches so routing accuracy improves over time. ## Core Features & Use Cases - Experience Collection: Records each routing decision with input, complexity score, outcome, latency, and confidence into a structured JSON log. - GRPO Batch Learning: Groups experiences by domain and complexity, compares System 1 vs System 2 success rates, and adjusts routing thresholds within a clamped [-0.1, 0.1] step. - Knowledge Transfer: Promotes patterns to System 1 after 3 consecutive System 2 successes and demotes them after 2 consecutive System 1 failures, with human checkpoints for review. - Use Case: After a week of sessions, run the learning cycle to discover that medium-complexity refactoring requests succeed more often in System 2, then automatically lower the routing threshold so future similar requests route correctly. ## Quick Start Ask the agent to run the lifelong learning cycle to batch-process this session's routing experiences and update the System 1 and System 2 caches.
