dsh-compaction-instant

In long-running conversations and tool-heavy workflows, traditional LLM-based context summarization often introduces high token costs, noticeable inference latency, and inevitable hallucinations or information loss. dsh-compaction-instant solves this critical bottleneck by completely bypassing model summarization. Instead, it leverages the deterministic conversation-compiler principle of lllyasviel/VCC to deliver instant, near-lossless context compaction for DeepSeek Harness.

The core capability of this plugin lies in its zero model calls and millisecond compilation. By executing a single deterministic pass over shadowed history nodes, it collapses complex tool calls into one-liners and truncates lengthy texts to fit predefined budgets. Crucially, every elided fragment is marked with a precise pointer to the durable session log. Coupled with the built-in recall and search tools, both the agent and human users can losslessly restore any folded historical detail on demand, closing the loop on near-lossless memory management.

If you are building agents that handle million-token histories, rely on frequent external tool calls, or have strict requirements for compression latency, API costs, and information fidelity, dsh-compaction-instant is the perfect fit. It serves as a contract-exact drop-in replacement for the basic compaction module, allowing your agents to maintain blazing-fast response times and exact memory recall in massive context windows.