dsh-asc

In long-running conversations, context windows can quickly expand, leading to degraded model performance or information loss. dsh-asc serves as a vital enhancement for DeepSeek Harness, addressing these challenges through a philosophy of "model-driven decisions." Rather than relying on rigid external scripts, this plugin empowers the model itself to autonomously determine when and what to compact, enabling intelligent context management from the very first turn.

Built upon an event-sourced architecture, dsh-asc ensures that every compaction operation is a durable, reversible log replacement event. It features a unique "tiered compaction" strategy—transitioning from Tier-1 raw details to Tier-2 distilled decisions and finally to Tier-3 fact indexes—which minimizes token usage while preserving the full history. With tools like context_search and context_decompress, developers gain full control over a searchable, replayable session log.

This plugin is an ideal choice for developers and engineers building complex agents or applications requiring robust long-context handling. By injecting a disciplined context-management loop directly into the system prompt, dsh-asc allows agents to maintain coherence and efficiency without manual intervention, making it a powerful tool for anyone looking to push the boundaries of AI agent capabilities.