@yoza10635

0-LLM compaction for DeepSeek Harness (dsh): deterministic atomic reference-graph pruning of agent history — exact compression ratio, selective forgetting, recall instead of rewrite. 压缩阶段零 LLM 调用的确定性上下文压缩插件：原子引用图剪枝、选择性遗忘、压缩率精确兑现，即插即用 CompactionEngine。