What problem does it solve? LLM agents degrade unpredictably in long conversations: they ignore information buried mid-context, compound errors from poisoned context, and mix constraints across tasks. This Skill provides detection signals, thresholds, and mitigation strategies for five degradation patterns: lost-in-middle, poisoning, distraction, confusion, and clash. ## Core Features & Use Cases - Degradation Detection: Measure attention distribution, detect lost-in-middle placement of critical information, and flag context poisoning indicators such as error accumulation and contradictions. - Composite Health Scoring: Run ContextHealthAnalyzer to combine utilization, degradation, and poisoning metrics into a single 0-1 health score with actionable recommendations. - Mitigation Frameworks: Apply the Write/Select/Compress/Isolate four-bucket strategy and architectural patterns like just-in-time loading and sub-agent isolation. - Use Case: An agent's output quality drops after 40K tokens of conversation. Use analyze_agent_context to get a health score, identify that critical instructions sit in the attention-degraded middle region, and restructure the context to place key information at the edges. ## Quick Start Analyze my agent's conversation context for degradation patterns and recommend whether to compact, truncate, or restructure it.
