Optimizes AI agent token usage through advanced context summarization and structured information preservation.
This skill provides a comprehensive framework for managing long-running agent sessions by implementing context compression strategies that prioritize 'tokens-per-task' over simple token reduction. It enables agents to maintain critical project state, file modification history, and decision-making logic even as conversation history exceeds context window limits. By using techniques like anchored iterative summarization and probe-based evaluation, it ensures that your Claude Code agents remain effective and accurate during complex development tasks without losing track of previous code changes or architectural decisions.
主な機能
01Tokens-per-task optimization to reduce re-fetching costs and minimize hallucinations.
02Probe-based evaluation frameworks to measure functional quality and factual retention.
03Structured summary sections for tracking session intent, modified files, and next steps.
047,140 GitHub stars
05Anchored Iterative Summarization to preserve file-specific changes and session intent.
06Adaptive compression triggers based on context utilization and logical task boundaries.
ユースケース
01Architecting multi-agent systems that require high-fidelity state handoffs between cycles.
02Managing long-running coding sessions that exceed standard model context windows.
03Debugging complex multi-file issues where the agent must remember previous file modifications.
