dsh-context-enhancement

Coding agents in long-running tasks can go through dozens or even hundreds of tool calls. Tool outputs, code snippets, and error logs eventually fill the limited context window. When /compact fires, the input is often too large and both fine-grained details and global structure are lost at once, making the agent's behavior feel like a different model before and after. dsh-context-enhancement is a DSH memory plugin built specifically to govern context within a single long session.\n\nIt replaces the blunt one-shot compaction with two parallel mechanisms. The main-thread compression chain partitions history into a memory zone, a tool-compression zone, and a recent zone: the oldest content undergoes batched semantic archival, tool outputs are summarized or pruned before being retained, and the current working context stays intact. Meanwhile, an independent checkpoint sidecar continuously updates a single task summary in the background using its own model request, input budget, output budget, and retry policy. The main agent always reads the latest successfully committed stable version, and background failures never block the primary task. After a service restart, task state can be restored directly from storage without re-invoking the model.\n\nThis plugin suits developers who run multi-turn long coding sessions in DSH, repeatedly hit information gaps after compaction, and want more robust context continuity on top of the full standard tooling and planning capabilities.