dsh-browser

@kasenri/dsh-browser

This repository is an installable release mirror for @kasenri/dsh-browser 0.1.0. Canonical source: https://github.com/KasenRi/dsh-orbit-browser-plugins/tree/main/packages/browser Do not develop features here; publish changes from the canonical source monorepo. Community plugin for DeepSeek Harness (DSH). Not affiliated with or endorsed by DeepSeek.

让 AI 真正操作浏览器完成任务。 支持打开网页、点击、输入、翻页、读取内容和下载文件等操作，可用于网页测试、信息采集和自动化流程，并提供域名限制等安全控制。

Technical integration:

• a BrowserAutomationService on ctx.browserAutomation
• a model-facing agent_browser tool that drives the agent-browser CLI

Harness native infrastructure (agents, tools, sessions)
        │
        ▼
@kasenri/dsh-browser
        │  BrowserAutomationService + agent_browser tool
        ▼
agent-browser CLI  →  Chromium / Chrome  (or an explicit CDP endpoint)

Requirements

DSH is in developer preview and evolves quickly; this package is tested with the versions above, not with every future DSH release.

Install

The package is published to the npm Registry; the dedicated Git distribution mirror remains available when a Git source is preferred:

dsh plugin --profile web add @kasenri/dsh-browser
# Git source alternative, tracks the mirror repository HEAD:
# dsh plugin --profile web add github:KasenRi/dsh-browser

The package declares a dsh.bundle patch, so dsh plugin add registers it as a profile layer automatically. Restart the profile (or start a new session) after installing.

The canonical source is maintained in the source monorepo. The source monorepo's versioned Release tarballs remain available for manual or offline fallback, but the Market uses this dedicated Git repository so updates can compare the locked commit with HEAD.

If agent-browser is not on PATH, point the plugin at an executable with the command config key, the executablePath config key, or the DSH_BROWSER_EXECUTABLE_PATH environment variable.

Tool input modes

agent_browser accepts exactly one of:

Standard workflow: open → snapshot -i → use the current @refs → click/fill/select → snapshot -i again after the page changes.

Other options: stdin (only for batch, eval --stdin, auth save --password-stdin), outputPath (atomic 0600 write of the structured result), timeoutMs, sessionMode (auto | fresh).

What it protects

• Stale refs — refs are page-scoped; mutations against refs that are not in the latest snapshot of the same target are blocked before any browser call.
• Tab drift — an unexpected active-target change (including about:blank) invalidates refs and triggers exactly one deterministic recovery via tab list + tab <id>; ambiguity fails as tab-drift instead of guessing.
• Artifacts — screenshots, downloads, PDFs, HAR/trace/record files are verified on disk (existence, size, type) before success is reported.
• Secrets and protected state — .agent-browser state, password stores and cookie databases are blocked; credential-shaped values are redacted from all tool output.
• Domain containment — when allowedDomains is configured, known URLs are preflighted, the managed context launches with upstream --allowed-domains, and the final URL is verified. CDP attach (electron/connect) is refused while containment is enabled, because it cannot be enforced on an existing browser context.
• Human verification — CAPTCHA, OTP, passkey, WebAuthn and 2FA are never bypassed; the tool reports what it saw and stops.

Large outputs are compacted into a bounded preview plus a spill file instead of being pushed into the model context.

Configuration

License

MIT