What problem does it solve? Long-running LLM agents silently degrade as context accumulates: critical instructions get lost in the middle of long prompts, hallucinated claims poison downstream reasoning, irrelevant documents dilute attention, and contradictory sources produce unpredictable outputs. This Skill provides a systematic framework to detect, diagnose, and fix these failures before they cascade. ## Core Features & Use Cases - Five-Pattern Diagnosis: Identify lost-in-middle, context poisoning, distraction, confusion, and clash failures with specific detection signals for each. - Health Analysis Toolkit: Python utilities to measure attention distribution, detect poisoning indicators, and compute a composite context health score with actionable recommendations. - Mitigation Framework: Apply the Write/Select/Compress/Isolate strategies matched to the active degradation pattern, with compaction triggers set before degradation cliffs. - Use Case: An agent's output quality drops after 40 conversation turns. Run the health analyzer to find utilization at 85% with poisoning indicators, then truncate to before the poisoning point and reload only verified context. ## Quick Start Ask the agent to analyze the current conversation context for degradation symptoms and recommend whether to compact, truncate, or isolate the context.
