跪拜 Guibai
← Back to the summary

DeepSeek Harness's Officially Praised Plugins, Plus One That Automates Your Terminal

Hi everyone, I'm ErGe.

I never expected DeepSeek Harness's daily active users calling the official API to have grown to 1.54 million.

That's insane.

All told, DeepSeek Harness hasn't even been online for a few months — a month or two? Third-party plugins have already reached 3.5 ten-thousand units. It's truly a case of "you love it, I love it, everyone loves AI."

The official team also selected 4 of these plugins to praise. Let's see if any of them catch your eye.

If these aren't enough, you can use your smart brain and craft your own.

01. Media Workbench

Nowadays everyone is diving into self-media. In a certain well-known community, I've seen many big-company employees running self-media accounts — and doing quite well.

So the self-media workbench dsh-media-desk is definitely worth installing. It's by tech blogger "Snail's Tech Notes."

When you use an Agent to write articles or video scripts, you often need dozens of rounds of conversation in a single session. Drafts, revision feedback, and storyboards all get mixed into the chat history, making it hard to find things later. This plugin adds an independent workbench to the left sidebar. From recording topic ideas, generating drafts, to scheduling publication, everything can be done inside DeepSeek Harness.

Installation is simple.

Open the desktop app, go to the "Plugins" page in the left sidebar, click "Add Plugin" in the upper-right, fill in dsh-media-desk, and after installation, click "Enable Now."

After enabling, a "Self-media Workbench" entry appears in the left sidebar. On first entry, it shows sample data. Click "Clear Sample Data" in the upper-right to remove it. The workbench has five modules at the top:

Pixel-style animation of the media workbench: recording topics, one-click drafting, and conversation fine-tuning

I saved "Explain Transformer's attention mechanism in one sentence" to the inspiration wall, clicked "One-Click Draft" on the card, and the article area on the right generated candidate titles, spoken script, and storyboard in one go.

One caveat from testing: the plugin itself doesn't come with an image-generation API key. To generate cover images and the six storyboard images, you need to fill in an OpenAI-compatible image-generation API key in the settings.

02. Guizang Design

When you ask an Agent to write web pages or PPTs, the colors and font sizes it generates are different every time, like opening a blind box.

dsh-guizang-design bundles five commonly used content-generation Skills and a unified set of color and layout specs into a single plugin. After installing and enabling it on the Plugins page, a "Guizang Design" entry appears in the left sidebar.

Typing / in the conversation input shows the five Skills. You can also start them by clicking the corresponding card in the plugin panel:

The first three visual Skills share a single design spec, with six magazine-style themes (Ink Classic, Indigo Porcelain, etc.) and four Swiss-style themes (Klein Blue, Lemon Yellow, etc.).

In testing, I opened "Guizang Design," chose "Web Layout," and sent two sentences about the attention mechanism to generate a single-page web page.

03. Context Insight

Below the conversation input, you usually only see a total Token count and a percentage ring.

You can't tell whether a jump to over a hundred thousand tokens comes from long chat history, large file reads, or tool definitions taking up a big chunk — the interface doesn't show it.

Bowen's dsh-context exists to solve this. It breaks down exactly where tokens are spent, item by item.

Search for dsh-context on the Plugins page, install, and enable. It offers four viewing entry points:

I used dsh-context on the session where I generated the "One-Sentence Understanding Attention" web page, switched to the Context tab, and immediately saw where the tokens were mostly spent.

04. Paper Reading

For anyone working on papers, this one is extremely useful.

Say you want to read the original Transformer paper, "Attention Is All You Need."

You can use agentero-dsh by the Phil-Fan team, a paper reading and research plugin.

It brings the research tool Agentero into the DeepSeek Harness client. Agentero's key feature is that all data stays local. The plugin consists of four modules: communication with the client, managing the literature library, note-taking, and reading PDFs. After installation, the window center is the conversation area with the Agent, and the right side holds the literature library, PDF reader, and notes panel.

agentero-dsh needs one extra environment step compared to the previous three. Follow these three steps:

Imported PDFs, highlights, and notes are all saved in this folder.

05. Automation Terminal Plugin

After seeing these four plugins, my hands were itching.

So I threw together an automation terminal plugin called paiwarp-dsh. You can describe your needs in natural language and let DeepSeek Harness drive the plugin to handle terminal input, screenshots, and screen recording.

Fully automated.

Normally, tools like Codex, Claude Code, and DeepSeek Harness can't directly drive the terminal; they lack terminal use.

General PaiWarp enabled on the desktop

paiwarp-dsh is not limited to a particular project. Whether it's a Java-based PaiCLI, a Python script, or a build task with a colorful progress bar, you can use natural language to have DeepSeek Harness open a real terminal on your machine, step by step input commands or key presses, read the on-screen output, check the program's exit code and generated files, and simultaneously save high-resolution PNG screenshots and MP4 recordings.

To try paiwarp-dsh yourself, follow these four steps:

git clone https://github.com/itwanger/paiwarp-dsh.git
cd paiwarp-dsh
python3 -m venv .venv
.venv/bin/python -m pip install playwright
.venv/bin/python -m playwright install chromium
.venv/bin/python lib/doctor.py

open -a "DeepSeek Harness" \
  --env "PAIWARP_PYTHON=$PWD/.venv/bin/python" \
  --env "PATH=$PATH"

You can start with the plugin's built-in example.

Create a new directory, copy todo.py and the test instructions CASE.md from examples/python-cli/, open that directory as the workspace, and send this prompt in the chat:

Please use PaiWarp to test todo.py step by step in a real Bash terminal according to CASE.md in the current workspace.
Decide next steps based on live output, verify the JSON file, error input, and the program's state after restart, 
save key screenshots and full recordings, finally close the terminal, and report real results and asset locations.

After sending, DeepSeek Harness automatically opens a real Bash terminal and starts recording, then based on the on-screen output, it inputs add 买咖啡, add review pull request, done 1, list, and finally done 99 to test whether the program correctly reports an error for a non-existent ID.

When the Agent operates the terminal, the "Terminal" tab on the right side displays the output in real time. If you want to run a few commands yourself mid-run, click "Take Over Keyboard" on the panel — the plugin immediately pauses the Agent, no longer accepts its input, and lets you take over. After you're done, click "Return to Agent" and the Agent continues testing from the current state.

All screenshots, recordings, and check reports are saved under .paiwarp/runs/<run-id>/ in the workspace.

Play real terminal recording in Harness asset tab

ending

An operating system for the Agent era.

After working through these five plugins, I have a more concrete understanding of DeepSeek Harness's official slogan.

In other words, you can customize DeepSeek Harness into anything you want it to be. Any of your previously tedious workflows can be accomplished with its help.

Like before, I tested PaiCLI by typing commands in the terminal myself. Now I let DeepSeek Harness directly do the automated testing and collect the evidence.

No more typing commands, no more taking screenshots — the AI just handles it.

See you next time.