dsh-builtin-browser

English | 中文

A built-in browser plugin for DeepSeek Harness. The agent and the user look at the same page: the agent drives the page in the conversation's right sidebar, and no separate browser window opens.

Demo video and screenshots are coming soon.

Highlights

• Shared page. Navigation, element snapshots, clicks, typing and screenshots run over the Chrome DevTools Protocol on the sidebar's <webview>, so the user watches (and can take over) exactly what the agent does.
• Numbered snapshots. Every element gets a number that later actions target with target.by: "ref". A click is hit-tested first, so a covered element is a reported failure instead of a click in the wrong place.
• Settled results. Click, type, select and similar actions return the settled page in the same result, so the model needs no follow-up snapshot.
• Human in the loop. Page dialogs, HTTP authentication and downloads become one status line above the page; CAPTCHA and sign-in prompts are handed to the human instead of being retried.
• Safe by default. Arbitrary page scripts, downloads, cookie export and cookie restore are off until you switch them on in the settings panel.
• Robust. Every CDP command has a timeout and a new command can interrupt a stuck navigation, so one bad page cannot stall a turn.
• Desktop-width pages. In a narrow sidebar, pages are laid out 1280 px wide and shrunk to fit; coordinate clicks stay accurate through zoom and display scaling.

33 dsh_browser_* tools are registered; see docs/reference.md for the tools, settings, rows and known limitations.

Requirements

• A DeepSeek Harness checkout (this package is a workspace:* package of it) built from a tree that contains the packages named in package.json peer dependencies.
• The Harness desktop app (apps/desktop), which owns the sidebar <webview> guests. The plugin cannot run without the shell half in shell/browser-bridge.ts: a token-authenticated loopback socket that publishes $DSH_HOME/dsh-builtin-browser-bridge.json and speaks newline-delimited JSON (version / list / cdp / ensureSidebar / showTab / closeSidebarBrowser, protocol generation 2). Without it every page operation fails with a named error; no fallback window is opened.

Install

1. Copy this repository into custom/packages/builtin-browser of your Harness checkout and run pnpm install.
2. Integrate the desktop shell half: apply shell/harness-shell.patch, copy shell/browser-bridge.ts, and wire ipc.ts / main.ts as described in docs/shell-integration.md. The patch adds what the bridge relies on (prompt queue, agent-input tracking, sidebar client service) and is based on upstream commit f3803c96 (0.2.1-rc.3).
3. Add the four rows from cordis.patch.yml to your profile. Keep the inert package-root row: the desktop shell reads dsh.client from it to load the settings panel.
4. Build with pnpm --filter @corp/dsh-builtin-browser bundle.

The @corp scope is the placeholder scope of the custom/ packages; keep it, or rename it consistently in package.json, cordis.patch.yml and client.js.

Origin and license

Vendored from wqty123/dsh-browser (dsh-builtin-browser 0.4.14, MIT) and reduced to the sidebar carrier, with many behavior changes listed in docs/reference.md. Thanks to the upstream authors. MIT licensed; the original LICENSE is retained.