DeepSeek Harness Plugin: dsh-skill-bundle
The `dsh-skill-bundle` plugin addresses the context pollution issue caused by skill installation in DeepSeek Harness through the “Bundle Box” concept. It allows developers to package related skills into bundles and load subskills on demand only via the root skill’s `loadSubskills` configuration, reducing interference from irrelevant information while optimizing inference efficiency and lowering costs. Installation requires using the plugin together with `dsh` and supports custom path configuration.
Read Moredsh-token-diet: A Pre-Use Token-Saving Toolkit for DeepSeek Harness
`dsh-token-diet` is a toolkit in DeepSeek Harness (DSH) for handling large-volume inputs, reducing token consumption at the source through structure-preserving summarization. It provides pure-function tools for text, JSON, and CSV, and supports configuring three presets or a manual compression ratio via the Web UI or settings. The plugin provides real token-saving feedback statistics, features a pure-function safety model, requires Node.js 22+, and can significantly improve context window utilization efficiency in long conversations.
Read Moredsh-mcp-adapter: Enables DeepSeek Harness to invoke MCP on demand instead of injecting all schemas into the context
# Introduction Connecting MCP servers to DeepSeek Harness (hereinafter referred to as dsh) is most directly achieved via the official `@deepseek-ai/dsh-mcp-client`. This client connects to the server at startup and registers each declared tool as a native `mcp__<server>__<tool>` function. The trade-off is the schema cost: with multiple servers, hundreds of tool definitions are passed into the context with every request, regardless of whether they are utilized. `dsh-mcp-adapter` provides an alternative solution to this problem by registering only a single `mcp` proxy tool. It employs lazy server startup, allowing the model to `search` or `describe` tools on demand.
Read Moredsh-mcp-proxy: A context-efficient MCP access method for DSH
dsh-mcp-proxy is an MIT-licensed DSH plugin designed to address the overhead increase caused by registering the complete Schema of all tools into the prompt context when configuring multiple MCP servers. By registering only two proxy tools, `mcp_discover` and `mcp_call`, and utilizing a lazy connection mechanism along with a disk-based tool directory cache to manage subsequent discovery and invocation, this plugin significantly reduces resident prompt costs. Its core functions include `mcp_discover`, which searches the tool directory by keyword, and `mcp_call`, which invokes tools by specifying the server and tool name. The plugin does not connect to servers upon startup, supports `stdio` or `streamable-http` types, and has an empty default server list, requiring manual user configuration. Installation requires N
Read Moredsh-rtk-filter: Passing Oversized Command Output to RTK for the Model
`dsh-rtk-filter` is a community plugin for DeepSeek Harness (DSH), maintained by Aeternus123 and licensed under the MIT License. This plugin addresses the issue of excessive token consumption and high context usage caused by overly long tool outputs during agent command execution. Its core mechanism involves condensing text results that meet certain conditions (e.g., byte size exceeding the threshold `minBytes`) via the external `rtk` CLI after tool execution but before the model receives the output. The plugin is registered as a `tools/post-execute` transformer, replacing only the model-visible content without altering the original tool values, while strictly preserving critical status markers such as exit codes, timeouts, and signals. If `rtk` is missing, fails, or produces no changes, the plugin silently degrades to outputting the original content, ensuring stability.
Read More