context-assembler-DSH

Every turn, a cloud LLM must re-ingest the entire conversation history. The longer the session grows, the more diluted the information becomes and the steeper the token bill. Context Assembler DSH inverts the economics: the local small model already has idle compute between turns, so it spends that time compressing and reassembling context on the client side, offloading what would have been a peak-cost cloud operation onto local idle cycles.

The core mechanism revolves around topic blocks. The session is split into blocks in real time; within each block a summarized version of the history is built and its prefix kept stable so the cloud prompt cache stays warm. As accumulated characters rise, a water-pressure model progressively lowers the Jaccard similarity threshold, force-splitting at peak to prevent an over-long single-topic session from forfeiting the cache advantage. On top of that, the plugin provides deterministic tool-round compression, topic grading and freezing on switch, reality-recall background injection, and pressure-triggered handoff planning, all as pure computation with zero host dependencies.

It is built for developers running DeepSeek Harness who already have an Ollama-compatible 4B-class model locally and are feeling the token pinch in long multi-turn agentic workflows. A GPU with roughly 10 GB of VRAM runs comfortably; CPU-only setups also work, at the cost of slower assembly.