A major highlight of DSH is its “Trace” system.
With the typical Agent Harness familiar to most users, the backend during task execution is often a black box. In DSH, every interaction with the large language model, message sending/receiving, tool calls, and more are transparently visible in the “Trace”. This allows users to clearly understand exactly what the AI is doing while performing tasks.
As shown in the screenshot above, details like how much time each step took, how many tokens were consumed, which skills were consulted, which tools were used, what commands were executed, etc., are all clearly disclosed.
Through it, you can theoretically review and optimize all the successful and failed details in the AI’s task execution process, making your usage increasingly smoother.
However, this form of AI execution trace presentation still seems a bit too geeky. I believe most ordinary users share my feeling—these traces are certainly valuable, but they’re also quite difficult to read.
Rather than struggling through them manually, why not let AI help read them?
So I created this sidebar plugin.
It integrates AI to directly translate what these execution traces actually mean, allowing human users to straightforwardly understand what the AI performing the task specifically did, how it did it, what errors occurred along the way, what risky behaviors were present, and what the final outcome was.
Compared to traditional black-box harnesses where you can only wait on the sidelines, with this plugin, you can let the AI execute tasks while simultaneously supervising how it’s executing them.
As shown below, I asked my locally deployed Qwen3.8-27b to research the latest MacBook Pro configurations and prices online. However, this is a locally deployed, quantized small model that hasn’t been fully broken in yet, so I don’t have complete trust in it. Therefore, I can open the sidebar interpretation and choose to use the affordable, capable, and fast Gemini 3.7 Flash to interpret its execution trace.
From the plugin’s interpretation, we can see that the Agent first made two parallel search calls, obtaining release information and starting prices for the M5 and M5 Pro/Max series.
It then checked IT Home, Macworld, and Apple’s official website to supplement pricing and specification information from both domestic and international sources.
Afterward, by comparing data from the official website, it discovered that the current prices didn’t match the launch prices. It likely identified Apple’s price increase after this year’s 618 shopping festival.
After confirming the price increase indeed existed, it began structurally extracting product specifications and prices…
Furthermore, each execution step includes an overall execution judgment provided by Gemini 3.7 Flash, assessing whether the current executing Agent’s behavior is reasonable, giving a risk level rating, and offering a reference idea for the next action.
I won’t include all the lengthy case execution details. In short, after all that operation, it successfully obtained the MacBook Pro configurations and prices from both the Chinese and US official websites, specifically noted the price changes, and even provided a brief purchasing suggestion.
Still have to marvel—Qwen3.8-27b is truly a little powerhouse. The entire task execution resulted in correct outcomes, and the process logic was sound. For a locally deployed, quantized version to achieve this level is really impressive.
Back to this plugin.
If you suspect there are critical, questionable steps during the Agent’s execution, or if you think the Gemini model isn’t reliable, you can use different models to analyze the trace twice for comparison.
If you still have doubts, you can add another layer, summoning an external judge, until you figure out the problem.
Actually, using it within Codex is my most common scenario for DSH—not just summoning GPT 5.6 to audit model traces, but there are other ways to play with it too. We’ll talk about that another time when we have the chance.
You can also, after completing a long execution, summarize and expand on the issues that arose, carefully analyzing what optimization-worthy problems exist.
As expected, PowerShell on Windows is a rope that can trip up everything.
In summary, this plugin’s purpose is to help you quickly understand what the model has done and how it did it, thereby judging whether its execution has problems, so you’re no longer left in the dark.
Installing the plugin is also very simple. You don’t even need to install it manually—just tell the AI:
Please help me install this DSH plugin: https://github.com/Liu-Bot24/dsh-trace-insightThat’s it. The AI will install it for you.
Have fun!