Introduction

When developing or debugging agents in the DeepSeek Harness (DSH) plugin ecosystem, you may often encounter situations where the model is “stubborn.” Even after receiving the correct instructions, the model may continue following its own logic during response generation or reasoning, causing the output to be unexpected or even refusing to perform the task. The traditional approach is to keep adjusting prompts or restarting the conversation, which is inefficient when debugging deep logic.

dsh-chat-thinking-editor solves this problem. It allows developers to directly modify the body content of any AI message and its underlying thinking chain in the DSH chat interface, ensuring that subsequent AI generation is fully based on the content you modified.

What It Is

This is a DeepSeek Harness plugin maintained by developer yumiao-dev.

Its core feature is providing direct editing capability for AI responses. By modifying an AI response’s body text and thinking chain, developers can preserve the model’s basic reasoning ability (“smart”) while forcing output that meets specific requirements (“compliant”), enabling precise control over AI behavior during the conversation.

Core Features

The plugin provides the following capabilities:

  1. Visual editing entry point: A pencil icon appears in the action bar of each AI message; click it to enter editing mode.
  2. Full text and thinking chain editing: The editor supports modifying both the AI response body and the thinking chain.
  3. Historical message management: All AI responses provided in the session (including intermediate tool-call information) are listed for you to modify.
  4. Version override: After modifying and saving, the AI will recognize only your version in subsequent generation, without needing to rerun the flow.
  5. Content preservation: Non-text content such as images and tool calls is preserved as-is; only the text parts are modified.
  6. Empty value handling: The body or thinking block can be left empty, effectively deleting that part of the content.
  7. Runtime locking: While the Agent is running, editing is locked to prevent conflicts caused by concurrent modifications.

Installation and Activation

Using this plugin requires the following environment:

  • DeepSeek Harness: Version 0.1.5-rc.2 or later (dependency range ^0.1.1-rc.2).
  • Node.js: 22.19+ or 24+.
  • pnpm: 11.

Install from the official npm registry:

dsh plugin --profile web add dsh-chat-thinking-editor

If you need to inspect source code or modify configuration, you can also install it from a local path:

git clone https://github.com/yumiao-dev/dsh-chat-thinking-editor.git
dsh plugin --profile web add /path/to/dsh-chat-thinking-editor

After installation is complete, refresh the DSH page and you can see the pencil icon in the chat interface.

Typical Usage

  1. In the DSH conversation, locate an AI response whose body content or reasoning process you are not satisfied with.
  2. Hover the mouse over that message and click the pencil icon that appears.
  3. In the pop-up editor, modify the response content or its underlying thinking chain.
  4. Click the Save button to apply the changes. All subsequent AI generation will be based on the version you modified.

Note: In the editing pop-up, you must click the close button in the top-left corner to close the window. Clicking the backdrop or pressing Esc will not close it; this is to prevent accidental touches on mobile devices.

Use Cases and Caveats

This plugin is suitable for developers who need to adjust AI output format in real time during a conversation, correct reasoning direction, or handle specific technical details.

Important notes:

  • Dependency environment: Ensure your DSH version is 0.1.5-rc.2 or later, and the runtime environment is Node 22.19+/24+ with pnpm 11.
  • Interface change risk: The plugin uses the official surface replacement event, but DSH official interfaces may change, and long-term compatibility is not guaranteed.
  • Permissions and source: The plugin runs with DSH process permissions. Before installing, it is recommended to check its source code and license (MIT).

Conclusion

dsh-chat-thinking-editor provides DSH users with a more fine-grained control mechanism. Rather than repeatedly struggling with the model, directly take over the output and let the AI operate within your logical framework. This can be very helpful when debugging complex agent behavior or performing technical deep interaction.