dsh-argp

Traditional summarization-based compression in LLM inference often leads to high costs, information distortion, and unpredictable compression ratios due to LLM calls that rewrite history. dsh-argp offers a novel solution: zero-LLM-call context compression through atomic reference graph pruning. By analyzing logical chain dependency topologies in conversations, it selectively retains referenced content and removes isolated information, ensuring a deterministic, transparent process with precise compression ratios. This selective forgetting mechanism—rather than rewriting—significantly reduces inference costs, while built-in recall tools preserve information traceability.

The plugin's core strengths lie in its engine-agnostic interface and strict compression budget enforcement. As a third-party CompactionEngine for DeepSeek Harness, it seamlessly replaces default backends without relying on model-based history rewriting, achieving deterministic convergence during compression. Suitable for developers, researchers, and users managing ultra-long contexts under cost constraints, dsh-argp delivers stable performance in high-throughput scenarios or tasks requiring high information fidelity, making it an efficient choice for precise and economical context management.