billion-context-dsh

As conversations grow, context windows inevitably fill up. Many tools resort to hard truncation, silently discarding earlier messages and causing the model to lose critical background. billion-context-dsh takes the opposite approach: it hands compression decisions back to the model, providing a set of context-compression tools that let the LLM decide when to compress and what to compress, writing high-fidelity summaries that preserve key details like file paths, decisions, and error messages while reclaiming space without losing important information.

At its core are model tools—compress, decompress, search_context, and acp_status—along with the /acp command and gentle nudges. Compression creates checkpoint nodes, while the original content remains in the append-only session log, making it recoverable and searchable. Repeated compression can also form hierarchical distillation, allowing key conclusions to accumulate over time, so each step builds on prior work and very long tasks become more likely to complete successfully.

Unlike DSH's built-in automatic compression, this plugin only suggests and never forces; summaries are written by the model itself, with no second LLM summarization call. The context therefore stays lean on every request, using fewer tokens without losing detail through coarse wholesale summarization.

It suits users running long conversations, complex coding sessions, or research tasks on DeepSeek Harness—especially those who need the model to remember key decisions and errors while worrying about silent context truncation. If you want context management to be a model-driven, recoverable process rather than a passive hard limit, this plugin is worth trying.