跪拜 Guibai
← All articles
AI Programming · DeepSeek · Frontend

DeepSeek Harness + dsh-browser Kills the Screenshot-to-Codex Debugging Loop

By 阿灿在吗 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Frontend debugging with AI assistants has been bottlenecked by session state: built-in browsers lose logins, and screenshot-based workflows are slow and lossy. A lightweight browser plugin that preserves real session state and accepts DOM-level commands from an agent removes that friction entirely.

Summary

DeepSeek V4 Pro's benchmark scores depend heavily on first-round tool anchoring, and the community has settled on a specific configuration — the "Liangzi" transformer position — to hit official numbers. A separate finding shows that running the model in minimal mode on non-Windows command-line environments also reaches those published scores.

More practically, the combination of DeepSeek Harness and the dsh-browser plugin changes frontend debugging workflow. Instead of taking screenshots of a broken page and pasting them into Codex with a description, a developer can now let the agent open the actual page, reuse the existing login session, and interact via DOM-level synthetic events to find and fix issues.

dsh-browser is still a small project with just over 100 GitHub stars, but its approach mirrors mature automation frameworks like Playwright. The difference is the tight coupling with an LLM agent that can read a page, navigate comments, and locate answers autonomously — a workflow the author tested live when the agent found a community solution on its own.

Takeaways
DeepSeek V4 Pro's benchmark performance is sensitive to first-round tool anchoring; the community converged on a specific transformer setting called the Liangzi position to reproduce official scores.
Running DeepSeek V4 Pro in minimal mode on non-Windows command-line environments also achieves the published benchmark results.
dsh-browser reuses the existing browser login session, so the agent never hits a re-authentication wall when debugging stateful pages.
The plugin uses DOM-level synthetic events rather than screenshot-based interaction, making it closer to Playwright-style automation than to visual-only agents.
During a live test, DeepSeek Harness autonomously opened a Zhihu answer, scrolled comments, and extracted the relevant community fix without manual guidance.
Conclusions

Session-state reuse is the unglamorous detail that separates a demo from a daily driver in AI-assisted debugging. dsh-browser solves it by never leaving the real browser.

The jump from screenshot-and-annotate workflows to direct DOM automation mirrors the earlier shift from manual QA scripts to Selenium — the agent becomes a participant in the page, not an observer of a picture of the page.

A plugin with 100 stars can already change a developer's daily workflow when it plugs the right gap in an LLM toolchain; maturity of the individual component matters less than how it completes the loop.

Concepts & terms
DeepSeek Harness
An open-source harness for running and evaluating DeepSeek models, providing a standardized environment to benchmark and configure model behavior.
dsh-browser
A lightweight browser plugin that integrates with DeepSeek Harness, allowing an AI agent to control a real browser tab with DOM-level synthetic events while preserving the user's login session.
First-round tool anchoring
A phenomenon where a model's initial tool-calling decision heavily determines downstream performance; if the first tool choice is suboptimal, subsequent results degrade.
Minimal mode
A stripped-down execution mode for DeepSeek V4 Pro that reduces environmental overhead, found by the community to recover benchmark scores that were otherwise lost in standard configurations.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗