dsh-tool-search

In AI agent workflows, growing tool inventories often create challenges in tool discovery and schema disclosure, where too many tools inflate token usage and complicate selection. The dsh-tool-search plugin addresses this by introducing a local tool_search tool for each agent, enabling autonomous searching and progressive selection of global tools. This approach optimizes resource efficiency and streamlines the tool management process, moving beyond static visibility models.

At its core, the plugin offers robust search capabilities and safety controls. Each agent independently uses tool_search, which ranks tools via BM25 scoring on names and descriptions, while alwaysVisible patterns keep critical tools persistently accessible. Progressive schema disclosure ensures only selected tools incur full schema costs, reducing request tokens and KV cache overhead. It integrates with the existing ctx.tools.restrict() seam without altering the agent loop, maintaining compatibility with the DSH ecosystem.

This plugin is ideal for developers, AI engineers, and research teams building or optimizing agent workflows with DeepSeek Harness. It caters to those managing multi-tool environments, requiring granular control over tool visibility, or aiming to lower token expenses. Whether for experimental projects or production tool management, dsh-tool-search provides a lightweight, extensible solution that enhances agent autonomy and system efficiency.