dsh-browser-annotate

When you ask an AI agent to open a page, click a button, or fill a form, it works blind: no visual context, no element grounding, and no way to absorb your offhand remark like "no, the button next to it." dsh-browser-annotate closes that gap by giving the agent twelve real browser tools driven over Chrome DevTools Protocol and Playwright against a live Chromium window — openBrowserPage, navigatePage, readPage, screenshotPage, clickElement, dragElement, runPlaywrightCode, cdp and more — while giving you a right-click annotation gesture that captures the element, a cropped screenshot, CSS selector, DOM path, computed styles, and your comment into a structured record that can be batched into the conversation composer as a precise coordinate for the agent\u0027s next move.\n\nAnnotation has two tempos. Quick mode is a right-click, type, Enter — one annotation done, even on pages that swallow their own events. Inspect mode lets you walk the DOM level by level with arrow keys, hover to preview selector and styles in a DevTools-style panel, and commit when you land on the right node. Every record is self-contained: the screenshot is cropped to the element\u0027s bounding rect, the selector and role are captured, and the page URL and title are stamped in. The agent reads the same queue through its annotations tool — list returns the screenshot inline, inject delivers it as a context block, and clear retires the batch once you have taken ownership.\n\nThe plugin is built for developers who already run DSH web in a terminal. Your conversation lives in the left pane, a real multi-tab browser in the right, and the two talk over loopback HTTP with no shared service. Neither half presupposes the other: without the desktop app the twelve tools answer with a clean "not attached" instead of hanging; without the plugin the app is simply a two-pane browser with annotations and queue delivery. If your workflow is "agent drives the page, I point at the exact element it should act on," this is the bridge.