context-degradation

Lost-in-middle context degradation occurs when language models lose attention to critical mid-context information as conversations grow longer. This skill diagnoses that attention collapse and quantifies its impact on important information in multi-turn dialogues and large documents.

Detect attention collapse by running a context-health analysis to measure attention distribution and identify middle-region risk. This skill quantifies where attention collapses mid-context and generates a health score to reveal degradation patterns in long-running agent sessions.

Yes, you can calculate a context health score using only numpy as a dependency. This skill measures attention distribution, detects poisoning indicators, and quantifies middle-region risk to generate health scores for long-running agent sessions without requiring additional frameworks.

Prevent context confusion and clash patterns by using selective loading and task isolation across sub-agents. This skill identifies these degradation patterns in large-context documents and offers practical mitigation strategies like context compaction to retain critical mid-context information.

Yes, context degradation detection works for interactive assistants, multi-turn conversations, and large-context documents. The skill diagnoses lost-in-middle, poisoning, distraction, confusion, and clash patterns across these formats to improve long-running session reliability.