context-degradation

Agents ignore instructions due to context degradation patterns like lost-in-the-middle risk or context poisoning. This skill diagnoses these failures by evaluating attention placement and recommending mitigations such as repositioning critical facts or truncating poisoned segments.

Detect context poisoning by running composite health analysis to identify corrupted or overridden information segments. This skill evaluates structural middle-zone risk and guides selective loading, compression, or isolation to mitigate poisoned context windows.

Lost-in-the-middle risk occurs when LLMs fail to process information placed in the center of long prompts. This skill identifies this attention failure during production monitoring and recommends actionable mitigations like selection, compression, or recovery via truncation.

Debug tool-output-heavy agent failures by diagnosing context degradation patterns like distraction, confusion, and clash. This skill evaluates attention placement risk across growing prompts and produces actionable mitigation recommendations for context isolation.

Yes, this skill supports production monitoring to set operational thresholds. It detects attention, poisoning, distraction, confusion, and clash degradation patterns, evaluating composite health to guide recovery via truncation and structural isolation.

Diagnosing confusion and clash requires evaluating composite health and attention placement risk across long prompts. Limitations arise when prompts lack identifiable structural middle-zones, making selective loading and isolation mitigations less effective for recovery.