ck:context-engineering

To optimize LLM context, analyze token utilization to detect degradation, then apply compaction, masking, or partitioning. This reduces token consumption while preventing reasoning degradation in long-running sessions and multi-agent coordination.

Lost-in-middle degradation occurs when an LLM ignores central information in long inputs. Detect this issue and error-dense poisoning regions by running a context health analysis to generate a health score and actionable recommendations.

Yes, you can evaluate context compression by generating probe sets and evaluating summaries. This measures information retention and continuity, ensuring compaction strategies preserve critical reasoning details.

Design persistent memory layers by applying multi-agent and memory guidance patterns that isolate work across sub-agents. This maintains context health and prevents reasoning degradation during extended operational sessions.

Trigger context compaction when runtime awareness monitoring detects operational thresholds approaching context limits. This prevents agent failures by executing warnings and compaction actions before task quality degrades.