DeepSeek Harness + dsh-browser Kills the Screenshot-to-Codex Debugging Loop
Keywords
DeepSeek Harness, dsh-browser, Minimal Mode
Background
While out shopping this afternoon, I came across the question How to evaluate that DeepSeek Harness's minimal mode on Linux seems to significantly improve DSV4PRO performance?. I didn't look closely at the time, but after getting home this evening, I dug into the details and figured out what was going on.
I won't recount the entire process in detail. Roughly, after the release of Deepseek V4 Pro and DeepSeek Harness, extensive community testing led to a consensus to set the sliding transformer to the Liangzi position. However, a blogger posted a video yesterday claiming that in a non-Windows command-line environment, using Deepseek V4 Pro's minimal mode could achieve the benchmark scores officially published for Deepseek V4 Pro.
The community's current discussion has reached a fairly clear conclusion: Deepseek V4 Pro's performance heavily depends on first-round tool anchoring, as shown in the image below.
While scrolling through the comments, I saw a user mention that a community plugin already existed, but I initially missed the image above.
I only saw the Liang Sheng mode. I searched the community but couldn't find it, so I let DeepSeek Harness do the work for me.
It then opened the answer on its own, opened the comment section, and found the answer highlighted in the red box in the image below by itself.
At that moment, I felt the laziest era for frontend development had arrived.
Previously, when using Codex to debug a webpage, Codex would either open its built-in browser — but if the target page had a lot of stored state, you'd likely need to log in again to reproduce the real environment — or rely on the ChatGPT Chrome extension connected to Codex to control the browser. I didn't use that feature much, and the few times I did, the experience wasn't great. So in practice, most of the time, if I found describing the problem too troublesome, I would just screenshot the page content, annotate it, and give it to Codex.
But now I've discovered that with DeepSeek Harness + dsh-browser, I probably won't need to do that anymore.
dsh-browser directly reuses the login session and uses DOM-level synthetic events, which is already sufficient for most daily webpage debugging. Its implementation is essentially a form of automation, similar in nature to automation frameworks like Microsoft's Playwright. Although the former is still a small plugin with just over 100 stars, far behind veterans like Playwright, DeepSeek Harness + dsh-browser might completely liberate me from those past days of taking screenshots and describing requirements to Codex.
Final Thoughts
I'll try this out more in real development going forward to further verify this possibility and see if it can lead to a future of even greater hands-free work.