ck:context-engineering

Debug context degradation by detecting lost-in-the-middle patterns and context poisoning risks that cause performance drops and persistent hallucinations. Identifying these issues helps recover model accuracy by isolating high-signal information from corrupted or irrelevant context.

Evaluate context compression using probe-based heuristics to judge whether summaries retain artifact tracking, continuity, and instruction compliance. Running the provided scripts measures summary accuracy, ensuring compressed context maintains the required reasoning fidelity.

Set explicit WARNING and CRITICAL thresholds for runtime context utilization so agents take compaction or reset actions on time. Defining these thresholds prevents degraded performance by ensuring agents proactively manage context pressure before hitting hard token limits.

Isolate high-signal information in multi-agent pipelines by curating and compressing context to prevent cross-agent poisoning and lost-in-the-middle patterns. This approach maintains reasoning quality by ensuring each agent operates only on the most relevant data for its specific step.