context-degradation

Agent performance drops in long sessions often stem from context degradation, where excessive tool outputs bury critical instructions in low-attention regions or introduce conflicting information. This Skill diagnoses these specific attention loss and distraction issues.

To diagnose lost-in-middle effects, you analyze the agent's context window to identify when critical information is buried in low-attention regions. This Skill applies attention analysis to detect this drift and pinpoint why earlier requirements are ignored.

Context poisoning occurs when conflicting or corrupted state information degrades agent responses. It is detected using poisoning heuristics and health scoring to evaluate degradation risk and identify specific clash or confusion scenarios within the session.

You can fix context drift by applying mitigation steps like context compaction, isolation, masking, and reordering critical information. This Skill generates practical recovery guidance and recommendations to restore performance in degraded agent sessions.

Yes, you can evaluate context health using numpy to process attention analysis and poisoning heuristics. This Skill leverages numpy to calculate utilization metrics and generate a comprehensive health score for agent debugging.

The best way to assess context health is to evaluate utilization, degradation risk, and poisoning indicators together. This Skill analyzes these factors to produce a health score and targeted recommendations for mitigating context degradation.