Tyan66666/billion-context-dsh
152Last commit Aug 16, 2026
billion-context-dsh DSH plugin
Billion-Context DSH is a plugin that replaces the default compaction backend in DeepSeek Harness. It gives the model a compress tool to write high-fidelity summaries of old messages, reclaiming context space without hard truncation. The plugin is in early beta and should not be used in production.
How to install the billion-context-dsh DSH plugin
dsh plugin --profile web add billion-context-dshCopying does not run this command. Review the repository and version before installing the billion-context-dsh DSH plugin.
billion-context-dsh DSH plugin data source
billion-context-dsh DSH plugin snapshot date: Aug 16, 2026
discovered
What the billion-context-dsh DSH plugin can do
- Provides four model tools: compress, decompress, search_context, and acp_status
- Supports `/acp` command for manual compression management
- Auto-nudge when context usage exceeds thresholds, letting the model decide whether to compress
- Persists compressed blocks in the append-only session log, with original text recoverable
- Supports layered distillation (tier 2/3) by re-compressing summary nodes
Where the billion-context-dsh DSH plugin fits
- Extending conversation length in agent sessions without losing earlier context
- Reducing token usage during long-running coding or research tasks
- Enabling the model to manage its own context window proactively
- Recovering compressed content through decompress or search_context when needed
Who the billion-context-dsh DSH plugin is for
- Developers using DeepSeek Harness for building agent applications
- Users who need finer-grained control over context compression than the default compaction-basic
billion-context-dsh DSH plugin limitations
- Both this plugin (v0.2.1) and DeepSeek Harness are in public beta; do not use in production
- Requires disabling the built-in compaction-basic backend to avoid conflicts
- Depends on acp-kernel and billion-context-pi upstream projects; may inherit their limitations
- Nudge thresholds and prompts are configurable but require understanding of the ACP design
billion-context-dsh DSH plugin: from the repository README
Quoted from the Tyan66666/billion-context-dsh README, the upstream source of the billion-context-dsh DSH plugin. Copyright remains with the original authors.
> **⚠️ 测试版声明——请勿用于生产环境** > 本项目（**v0.2.1**）仍处于开发中的测试版。[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 本身也处于**公开测试版**阶段。**请勿将两者用于工程化 / 生产环境**——预期会有破坏性变更与粗糙之处。 <p align="center"> <strong>衷心感谢以下项目——请给它们一个 ⭐：</strong> <br /> <a href="https://github.com/deepseek-ai/deepseek-harness">DeepSeek Harness</a> · <a href="https://github.com/ranxianglei/billion-context-pi">billion-context-pi</a> · <a href="https://github.com/ranxianglei/acp-kernel">acp-kernel</a> · <a href="https://github.com/ranxianglei/opencode-acp">opencode-acp</a> </p> <p align="center"> <strong>Billion-Context</strong> for <a href="https://github.com/deepseek-ai/deepseek-harness">DeepSeek Harness</a> <br /> 由模型决定<em>何时</em>压缩、<em>压缩什么</em>——而不是一个硬性上限。 </p> --- <p align="center"> <a href="https://www.npmjs.com/package/billion-context-dsh"><img src="https://img.shields.io/npm/v/billion-context-dsh.svg?style=flat-square" alt="npm"></a> <a href="https://github.com/Tyan66666/billion-context-dsh/blob/main/LICENSE"><img src="https://img.shields.io/npm/l/billion-context-dsh.svg?style=flat-square" alt="license"></a> <a href="https://github.com/Tyan66666/billion-context-dsh"><img src="https://img.shields.io/bad
Read the full READMERepository license: MIT
billion-context-dsh DSH plugin questions
How do I install billion-context-dsh?
Run `npm install billion-context-dsh` in your project. Alternatively, use `dsh plugin --profile web add billion-context-dsh` for automatic bundle installation. After installation, disable the default compaction-basic backend in your profile to avoid conflicts.
What does the compress tool do?
The compress tool allows the model to replace a range of conversation history with a concise summary written by the model itself. The original text is preserved in the append-only session log and can be recovered later via decompress. This helps reclaim context space without losing information.
Can I customize the compression prompts?
Yes, you can provide custom prompts via the `config.prompts` field in the plugin configuration. You can override nudge text, tool descriptions, and system prompt segments. Template placeholders like `{pct}` are supported. If not configured, the plugin uses the default prompts from acp-kernel.
Does it work with all models?
The plugin is designed to work with any model supported by DeepSeek Harness. It automatically detects the model's context window size via `agent.ctx.llm.resolveModelInfo`. You can also manually set `modelContextLimit` in the config. However, the plugin's effectiveness depends on the model's ability to follow instructions for compression.
How is billion-context-dsh different from the built-in compaction-basic?
The built-in compaction-basic uses automatic truncation with a hard limit. Billion-context-dsh is model-driven: the LLM decides when and what to compress using a compress tool. It also provides decompress, search_context, and acp_status tools, and preserves the original text in the session log, unlike the hard truncation approach.