dsh-compaction-instant: An LLM-free Context Compaction Engine for DeepSeek Harness

What Does It Do?

dsh-compaction-instant is a DSH plugin for DeepSeek Harness that replaces LLM summarization with a deterministic conversation-compiler algorithm, achieving millisecond, near-lossless context compaction. It solves the latency and cost problems of model-based compaction by performing zero model calls and zero token spend, while every cut remains recoverable via seq pointers into the durable session log. Maintained by TsFreddie under the MIT license, it was last updated in 2026-08.

Core Features

• LLM-free compaction: never invokes a model; deterministic text processing compresses a million-token history in milliseconds.
• Near-lossless: output contains only original tokens; every cut is marked and points to its durable seq, and prior checkpoints are copied verbatim.
• Drop-in replacement: same seam, events, provenance, and failure vocabulary as compaction-basic; all built-in presets load unchanged.
• One-line tool calls: whitelisted tools (toolArgTools) keep key arguments, others show only names; tool results never occupy entries, with result pointers for recall.
• Recall layer: recall tool, search tool, and /recall command restore elided content by seq, result, or checkpoint, and support keyword/regex search over the whole durable log.

How to Use This Plugin?

After enabling it in DSH, compaction runs automatically on shadowed history spans without manual steps. To recover compressed content, the model can call the recall tool (specifying type and id) or the search tool (keyword/regex), while human users can use the /recall command to append matching events as a durable message. Adjust recall budgets via maxRecallTokens (default 16000) and search hit caps via maxSearchHits (default 50).