context-degradation

To detect context degradation, analyze extended chat transcripts to identify lost-in-middle, context poisoning, and distraction patterns. The skill computes health scores and simulates attention distributions to quantify long-context failure severity.

Lost-in-middle and context poisoning occur when attention distributions fluctuate across extended interactions, causing ignored central context or corrupted states. This skill diagnoses these phenomena by simulating attention distributions and detecting contradictions.

Diagnose long-context failures in multi-task agents by running the context health analyzer on active chat transcripts. It identifies distraction and context clash phenomena, then generates actionable mitigation steps and architecture-oriented recommendations.

Yes, you can integrate context health analysis into agent monitoring dashboards. The skill provides lightweight tooling for run-time checks, allowing you to continuously monitor degradation scores and surface context clash events.

Yes, numpy is required to analyze attention distributions and generate context health scores. It serves as the foundational dependency for simulating numerical degradation patterns and detecting poisoning indicators.

The best way to mitigate context clash and distraction is to apply automated recommendations generated from health score analysis. The skill outputs actionable mitigation steps and architecture-oriented solutions to proactively resolve long-context failures.