context-degradation

Long-context degradation in agent sessions occurs when reasoning quality drops during ongoing conversations. It manifests as lost-in-middle, context poisoning, distraction, confusion, and context clash patterns that compromise multi-task reasoning performance.

You detect lost-in-middle and context poisoning issues by analyzing attention distribution and context structure. This skill assesses risky regions and identifies poisoning indicators to pinpoint degradation patterns in ongoing agent conversations.

You fix context degradation by applying mitigation strategies such as context isolation, selective loading, and proactive truncation. The skill provides actionable guidance to optimize prompts and maintain reliable task performance.

Yes, you can diagnose context clash and distraction for multi-task reasoning. The skill detects these degradation patterns in ongoing conversations with shifting objectives and provides actionable mitigation guidance for robust reasoning.

Analysis of long-context degradation applies to ongoing conversations with shifting objectives or multi-task reasoning. It requires quantifying attention distribution and context structure to identify risky regions and poisoning indicators accurately.