context-degradation

Context degradation in multi-turn agent conversations is detected by analyzing attention distribution, Lost-in-Middle patterns, and context poisoning indicators. This skill quantifies these degradation risks and provides a health score with actionable recommendations to preserve context integrity.

Lost-in-Middle risk during long-running agent sessions occurs when central context information drifts or becomes inaccessible to the model's attention. This skill identifies and quantifies these attention distribution patterns to highlight at-risk information in multi-turn workflows.

Yes, you can monitor context poisoning signals in multi-agent workflows using this skill. It applies structured detection to multi-agent orchestration sessions, analyzing context structure to guide debugging and sustain production performance.

Attention distribution is analyzed by applying modular detection mechanisms to the current session context. This skill surfaces degradation risks and generates a health score that provides actionable recommendations to debug and preserve agent reliability.

Yes, numpy is required as a dependency to run context health analysis on long-running sessions. It supports the underlying numerical computations needed to quantify degradation patterns and calculate the overall context health score.