LizzzMax/dsh-tool-index

Local tool discovery for DeepSeek Harness. The plugin builds a per-agent in-memory keyword index from the tools admitted for the current step. It ranks names, descriptions, parameters, action aliases, and URI schemes. A clear action and object match can expose one deferred tool directly. Otherwise dsh_search_tools returns up to three candidates and dsh_call_tool invokes the chosen tool through DSH's normal executor and step admission.

Token benchmark

Each task ran once in each isolated profile using DSH 0.2.2-rc.1, Web Standard, and gpt-6.1-sol at medium reasoning. Both profiles included OpenViking. The comparison used Lean Context 0.1.0 and dsh-tool-lazy-dispatcher 1.2.0 with a local compatibility patch. Totals include input and output tokens, with cached input counted in the input total.

For the MCP task, the prompt specified a resource URI and a reading goal. Lean Context preselected the reader, so the model called it directly and retrieved the expected value. In the file task, the baseline also spent one request on pwd, while Lean Context started with glob. Request counts can vary with the model's execution path.

Profile setup

In DSH Web, open Settings → Plugins → Add Plugin, then paste the package URL from the latest release. You can also install the downloaded .tgz or this source folder as a local package. Enable the dsh-lean-context bundle in your profile. In DSH Web you can also install by entering dsh-lean-context as the npm package name.

For upgrades from a GitHub release through 0.1.1, remove the old dsh-tool-index bundle before enabling dsh-lean-context.

Use one tool-disclosure bundle per profile. This bundle replaces the default native executor entry and delegates actual calls to DSH's native executor. The verified profile is DSH 0.2.2-rc.1 in Web Standard mode.

The defaults defer mcp__ tools and keep other tools visible. Override deferredPrefixes to include other tool families and pinnedTools to keep specific tools visible.

- id: tool-index
  config:
    deferredPrefixes: ["mcp__"]
    pinnedTools: ["mcp__browser__screenshot"]

Run npm test in this folder to check keyword ranking, conservative preselection, and exact discovery. The retrieval runs in memory and does not require an embedding service or a separate model request.

Example

Ask “Take a screenshot of the current web page” or “Read the memo at this resource URI.” A clear action and object can preselect a matching tool without the user naming it. Ambiguous requests can use the search tool. Each search and call sees only the tools admitted for that DSH step.

Welcome to try it. Issues and feedback are welcome, especially queries that selected an unexpected tool.