In the plugin ecosystem of DeepSeek Harness (DSH), building domain-specific agents often requires handling structured data. In the chemistry domain, the traditional approach is to hard-code or manually enter SMILES strings, which is not only cumbersome but also prone to parsing failures caused by typos. The dsh-synomega plugin solves this problem by providing DSH with localized organic reaction prediction capabilities.

Plugin Positioning

This is a plugin designed for DeepSeek Harness and built based on SynOmega. It supports retrosynthetic analysis, forward reaction prediction, synthesis route planning, synthesis feasibility scoring, and multicomponent reaction evolution. All calculations are performed locally and do not depend on remote APIs.

Core Capabilities

The plugin provides the following five model-facing tools, which can be invoked through conversation:

  1. Retrosynthetic Analysis: Answers “what reaction can produce X” and lists candidate single-bond-disconnection reactions.
  2. Forward Reaction Prediction: Answers “what can these reactants produce” and ranks the products.
  3. Synthesis Route Planning: Answers “how do I prepare X” and plans a route toward purchasable starting materials.
  4. Synthesis Feasibility Scoring: Answers “can X be made, and how difficult is it” and provides a SynScore.
  5. Multicomponent Reaction Evolution: Answers “what can this mixture form” and performs forward network evolution from a reactant pool.

Installation and Configuration

Before installing, ensure that your system has Node.js installed (version 22.19 or higher is required) and Python 3.10 or a newer version.

The installation command is as follows:

dsh plugin --profile web add github:zbc0315/dsh-synomega

On the first startup of Harness, the plugin automatically configures the Python environment in the background (located at $DSH_HOME/plugins/synomega) and downloads the necessary model files (about 450 MB). If the system’s python3 version is too old, it is recommended to install uv to ensure the plugin can obtain a suitable interpreter.

Usage Examples

After installation, simply ask in the conversation. The plugin automatically selects the appropriate tool based on the question.

  • “What reacts to give X?”
  • “What do these reactants give?”
  • “How do I make X?”
  • “Can X be made, and how hard?”
  • “What can this mixture form?”

The plugin directly renders molecular structures, reaction equations, and route tree cards in the conversation. To reduce errors in SMILES string input, the plugin integrates the Ketcher molecular editor. Click the benzene ring icon in the Composer control bar to open the drawing interface. After drawing is complete, click Insert, and the structure will be embedded in the message as SMILES instead of manually typing text.

Notes

  • Local execution: All inference is performed locally, and no API Key is required.
  • No Web server environment: If using a configuration without a Web server (such as pure CLI or Headless mode), the plugin cannot register routes. This plugin relies on the browser UI for rendering and must be used with a Web-supporting configuration.
  • Dependencies: Python 3.10 or a newer version is required.