Context Compression

This handles context compression for long-running agent sessions where naive summarization causes the agent to forget which files it modified or lose track of decisions. It recommends anchored iterative summarization with structured sections (files modified, decisions made, next steps) rather than regenerating summaries from scratch each time, which causes cumulative detail loss. The key insight is optimizing for tokens per task, not tokens per request. If your agent keeps re-reading files it already processed, you're compressing too aggressively and wasting more tokens on re-fetching than you save. Use probe-based evaluation (can the agent answer "which files did we modify?") instead of metrics like ROUGE, which miss the functional details that actually matter for continuation.

Install to Claude Code

npx -y skills add guanyang/antigravity-skills --skill context-compression --agent claude-code