Introduction

The plugin-based architecture of DeepSeek Harness (DSH) allows users to extend the functionality of the Web interface. In scenarios involving long conversations or specific style transitions, manually typing the next line can sometimes be inefficient. dsh-suggest-reply is a sidebar plugin based on dsh-better-sidebar that allows users to generate a list of candidate replies for the latest AI response in the main conversation using a custom system prompt. Clicking a list item directly fills it into the main conversation input box, assisting content creation.

Core Features

The plugin registers a tab named “Help Me Think” in the sidebar, with the following capabilities:

  • Three input fields and custom Prompt:
    • System Prompt (required): fully authored by the user, used to define the AI’s identity, task, and output format (must include a JSON format requirement, such as {"candidates":[{"style":"风格","text":"内容"}]}). The plugin does not include built-in templates and has no preset styles.
    • Latest Message in Main Conversation (required): automatically reads the latest AI response in the current session. While the panel is open, it polls every 2 seconds to stay current; if manually edited, refresh is paused.
    • Director Requirements (optional): acts as an anchor placed before the latest message and as part of the user message.
  • Dynamic result rendering: parses the [{style, text}] structure from the model’s JSON output and dynamically renders result rows. There is no fixed style list and no checkbox; clicking a result text fills it into the main conversation input box (automatically falling back to copying if the Composer is unavailable).
  • Independence and purity: generation is fully isolated from the main session; no new session is created, no fork is performed, and no session logs are written. Each generation is a brand-new off-loop completion.
  • Streaming debugging and full logs: during generation, the panel scrolls the AI’s reasoning stream and raw result in real time. After generation, a collapsible area at the bottom displays the full system prompt, user message, reasoning stream, raw model output, and parsed results, helping troubleshoot Prompt issues.
  • Sampling parameters and persistence: supports a collapsible sampling parameters section (temperature / max output / reasoning effort / stop sequences), passed through to the model. The system prompt, sampling parameters, and panel height are automatically persisted to DSH settings (the suggestreply namespace).

Installation and Activation

Prerequisite: dsh-better-sidebar (version >= 0.14.0) must be installed first; otherwise this plugin will not activate.

dsh plugin --profile web add dsh-better-sidebar@latest

Install the plugin:

dsh plugin --profile web add dsh-suggest-reply

Start and refresh:

After installation, restart the DSH Web process serving this profile (e.g., using PM2):

pm2 restart dsh-web

Hard refresh the page in the browser (Cmd/Ctrl + Shift + R) to load the new plugin. The “Help Me Think” tab will appear in the sidebar “+” menu.

Usage Workflow

  1. Open the “Help Me Think” tab in the sidebar.
  2. Write the Prompt in the System Prompt field, specifying the AI’s identity, task, and JSON output format.
  3. The Latest Message in Main Conversation field is automatically filled with the latest reply in the current session; modify it directly if needed.
  4. (Optional) Enter an anchor for the story direction in the Director Requirements field.
  5. Click the “Generate Candidate Replies” button.
  6. The panel displays the AI’s reasoning stream and raw result in real time. After generation completes, the list shows candidate replies.
  7. Click any candidate text; its content is directly filled into the main conversation input box. Press Enter to send.

Notes

  • Single call with no retry: The plugin does not include a retry mechanism. The model’s output is parsed as-is; if parsing fails or the output is empty, check the Prompt and retry manually.
  • Sampling parameter limits: Supported sampling parameters include temperature, max output, reasoning effort, and stop sequences. Parameters such as top_p, penalty, or seed are not supported.
  • Language follows the browser: The interface supports Chinese and English, and follows the browser language by default.
  • Debugging oriented: Since there are no built-in templates, Prompt writing is entirely up to the user. The debugging information panel is the primary tool for troubleshooting output format errors.