dsh-pilot

dsh-pilot addresses a common gap for chat-based agents: they can often explain what to do, but cannot actually open a page, read it, click, type, or verify the result. This plugin gives a DeepSeek Harness agent real browser control, turning web tasks from advice into executable actions while keeping the process visible and auditable.\n\nIts central idea is to make the browser legible to the model. Pages are converted into structured text with numbered element references, so the agent can act by referring to visible elements instead of guessing fragile CSS selectors. Opening, reading, comparing changes, clicking, typing, filling labels, uploading, waiting, asserting, downloading, and taking screenshots are all available as tools. At the same time, the web UI shows a live cockpit panel with a screenshot, current URL, action log, and session controls, so a human can watch what the agent is doing at every step.\n\nIt is most useful for people who want agents to perform practical web work: testing flows, filling forms, reading page content, checking whether an action succeeded, or automating repetitive browsing steps. Because it runs an isolated local browser, uses a text-first approach, and does not depend on an external browser service, it fits local development, internal tools, and privacy-conscious workflows where control and visibility matter.