dsh-compaction-instant: An Instant, Near-Lossless Context Compaction DSH plugin for DeepSeek Harness
kitdoesit/dsh-compaction-instant
LLM-free near-lossless context compaction for DeepSeek Harness, with instant compile and recall recovery.
dsh-compaction-instant provides LLM-free instant compaction for DeepSeek Harness, replacing LLM summarization, keeping original tokens, and enabling full recall.
dsh plugin --profile web add github:kitdoesit/dsh-compaction-instant- Category
- Memory & Context
- Platform
- DSH-Plugin
- Author
- kitdoesit
- Distribution
- Plugin
dsh-compaction-instant Key Features
dsh-compaction-instant Repository
dsh-compaction-instant Repo Summary
dsh-compaction-instant is a context compaction plugin (DSH plugin) for DeepSeek Harness, maintained by KitDoesIt under the MIT license, last updated in 2026-08. It serves as a drop-in replacement for @deepseek-ai/dsh-compaction-basic, replacing LLM summarization with the deterministic conversation-compiler principle of lllyasviel/VCC, achieving millisecond, near-lossless context compaction. The plugin solves the latency and cost issues of model-dependent compression, with core capabilities including: LLM-free instant compaction, near-lossless output containing only original tokens, and companion recall and search tools ensuring compressed content is fully recoverable via (seq N) pointers into the durable session log.
What core features does the dsh-compaction-instant DSH plugin support?
The plugin provides an LLM-free compaction engine that never invokes a model—no summarizer prompt, no inference latency, no token spend; a million-token history compresses in milliseconds. Output contains only original tokens; every cut is marked and points at its durable seq, and prior checkpoints are copied verbatim. It is a contract-exact drop-in for compaction-basic, maintaining the same seam, events, provenance, and failure vocabulary, so every built-in preset loads unchanged. Additionally, each tool call in the compiled output occupies one line: whitelisted tools show key arguments, others show name only, and tool results never occupy entries—the -> result N pointer keeps them one recall(type:"result") away.
How does dsh-compaction-instant achieve near-lossless read-back (recall)?
The plugin ships recall and search tools that close the near-lossless loop. The recall tool (model-facing) supports typed restore by seq, result, and checkpoint; the search tool (model-facing, grep) performs keyword/regex search over the whole durable log, including content elided by compaction, returning matching events with their (seq N) pointers. Humans can use the /recall command (/recall <keyword|regex>) to append a durable form: "recall" user message with matching events and seq pointers, making them visible to the next model turn. Shared cores dsh-compaction-instant/recall and dsh-compaction-instant/search handle seq parsing, log expansion, budgets, and projection. Everything is kept: text, reasoning, raw tool-call arguments, nested tool-result content; log-only events render as labeled data dumps; missing seqs are reported; a maxRecallTokens budget (default 16000) cuts with a provenance marker and counts the skipped remainder; searches cap shown hits (maxSearchHits, default 50).
How to install dsh-compaction-instant?
Installing the plugin is a single command via the DSH plugin manager, adding the GitHub repository:
dsh plugin --profile web add github:KitDoesIt/dsh-compaction-instant
Once installed, the plugin mounts as separate rows and can coexist with any compaction backend, as it only reads the durable log.
Which DeepSeek Harness versions and platforms are compatible with dsh-compaction-instant?
The plugin is built on official DeepSeek Harness dependencies, including @deepseek-ai/dsh-agent, @deepseek-ai/dsh-compaction, @deepseek-ai/dsh-session, etc., ensuring seamless integration with the DSH ecosystem. As a drop-in replacement for compaction-basic, all built-in presets load unchanged without configuration modifications. Written in JavaScript, the plugin is cross-platform and works in any environment supporting DSH.