Introduction

In practical use of DeepSeek Harness (DSH), when dealing with AI-generated responses, developers or users often find it difficult to precisely control subtle style nuances or emphasis with a single prompt (for example: more formal, more concise, or usage of specific terminology). Repeatedly asking for “another version” is usually inefficient.

dsh-possibility-space transforms a DSH response into an explorable two-dimensional semantic map. It does not automatically generate multiple candidates and choose one for you; instead, it first identifies key trade-off axes at the semantic level, then lets you generate a unique variant by dragging coordinates on the map.

Plugin Overview

Name: dsh-possibility-space
Core value: Turn a DSH response into a navigable semantic possibility space.
Author: zhangguiping-xydt
Category: Model inference
License: MIT

Core Features

  1. Semantic trade-off axis discovery: For a specific response, automatically discovers three independent semantic trade-off dimensions (e.g., formality vs. approachability).
  2. Map interaction: Select two dimensions to generate a 2D map. Drag a preview coordinate on the map; releasing the mouse generates the corresponding response variant.
  3. Content locking: Supports verbatim locking of text you are satisfied with on the map, or adding semantic invariants.
  4. Version tracking and comparison: Each generated point is recorded in a version trajectory, allowing you to view diffs and revisit entries anytime.
  5. Personal preference saving: You can save a selected response and its coordinates as personal preferences. When opening a space with the same semantic axes later, the coordinate can be restored directly.

Installation and Activation

Prerequisites:
- DeepSeek Harness (DSH) version 0.1.1-rc.2 or later
- Node.js 22 or later

Installation command:

npx --yes @deepseek-ai/dsh@latest plugin --profile web add github:zhangguiping-xydt/dsh-possibility-space
npx --yes @deepseek-ai/dsh@latest web

Usage Example

In a DSH conversation, after the assistant generates a response, a map entry appears below that response. Click it to expand the map within the conversation flow (you can also use “Full Screen Mode” for focused interaction). The workflow is as follows:

  1. Click the map entry to expand the interactive interface.
  2. The system displays the three semantic axes associated with the current response.
  3. Drag the mouse on the 2D map to select a coordinate. After you release the mouse, DSH generates a new response variant based on that coordinate.
  4. If you are satisfied with the generated variant, you can click “Lock” to preserve the text, or review the diff.

Notes and Constraints

  • No separate API key: The plugin does not require a separate API key. It reuses the Provider and Model information of the currently opened response and calls the model through the configured DSH adapter.
  • Privacy and data: The plugin does not maintain its own cloud service. All requests flow between the browser, the authenticated DSH Host, and the configured model adapter. Responses, variants, locks, caches, and preference data are all stored in the browser’s localStorage.
  • Host constraints: Model input length, number of locks, output token count, and concurrency are limited by the DSH Host, which also validates the JSON format returned by all models.
  • Permission scope: The plugin only runs in the browser/Host environment. It does not execute external tools, edit files, or submit code.

Summary

dsh-possibility-space provides a fine-grained way to control model output, suitable for scenarios that require precise adjustment of AI response style or structure. Through semantic axes and coordinate dragging, developers can quickly iterate on variants until they meet expectations.