Agent-driven compaction for DeepSeek Harness: the agent summarizes its own conversation (KV-cache friendly, no giant replay). Verified in production at ~460k-token sessions.