dsh-tabbit — Tabbit's Official DeepSeek Harness Plugin

English | 简体中文 | Changelog

The DeepSeek Harness (dsh) plugin bundle for Tabbit Browser. dsh can use this plugin to drive Tabbit to complete Agent tasks: real pages, real login sessions, real interactions, powered by native code-first tools (no shell forwarding). Suitable for web automation, information extraction, QA, and evaluation.

Capabilities

Capability	Description
tabbit_browser	Users can perform web reading and web operations via the real CLI mode built into Tabbit.
Browser-agent web_fetch	Reroutes dsh's web_fetch to Tabbit (dsh's built-in direct fetcher has no JS rendering, no login session, and does not use the system proxy — under a fake-ip proxy environment it would reject every domain).
"Web tab" mentions with @ in the input box	Type @ in the dsh Web UI input box to list pages opened in the current session's browser task as well as all tabs in the user's browser, which can be selected as context.
Standalone "page read" permission	Robust permission control. Because it shares cookies with the user, dsh confirms permissions when accessing Tabbit web pages: pageAccess (prompted once per session by default) + intranetFetch (web_fetch access to intranet targets is approved per call by default).
tabbit_browser_install tool	Environment pre-check: detects installed stable Tabbit, validates launcher and Runtime Service; when missing or outdated, downloads Tabbit as a dsh background task.
tabbit_plugin_update tool	Plugin update check: queries the latest release on npm at most once per day; if a suitable version exists, it installs offline and silently.
/tabbit-info command	Type /tabbit-info in the dsh input box to diagnose: plugin and host DSH versions, launcher, instance list (including product names), active instance and its source, permission settings, and task occupancy.
tabbit skill	Informs the model of Tabbit best practices. By default it uses the official skill shipped with Tabbit at ~/.agents/skills/tabbit/ (which evolves in sync with the browser Runtime). This plugin ships a fallback version.

Installation

Prerequisites

• A stable version of Tabbit Browser is installed (International or China, ≥1.9.0) and has been launched at least once (registers the CLI launcher on first launch);
• Node ≥ 22.19, dsh ≥ 0.1.1-rc.2 (install dsh: npm install -g @deepseek-ai/dsh). On hosts below 0.1.2-alpha.1, the built-in web_fetch tool is not available in web-based sessions (tabbit_browser is unaffected).

Install dsh-tabbit

dsh plugin --profile web add dsh-tabbit                 # npm 主路线

Other Installation Modes

dsh plugin --profile web add github:Tabbit-Browser/dsh-tabbit # npm 不可达时的回退
dsh plugin --profile web add link:/path/to/dsh-tabbit   # 本地开发

This package supersedes the early tabbit-browser skill-only plugin; the 0.2.x versions on npm are also continued by this release — 0.2.x users will receive an upgrade prompt via the daily update check; just re-run the install command to upgrade in place.

Community

Scan the QR code below to join the dsh-tabbit developer community and share experiences, report issues, and discuss new features:

Setup

Basic Configuration

dsh Settings → tabbit, or $DSH_HOME/settings.yaml

tabbit:
  instance: ""            # 显式指定 16 位大写 hex 实例 id（/tabbit-info 可列出）；通常留空即可
  launcherPath: ""        # 默认自动发现：优先 ~/.local/bin/tabbit-cli，回退 tabbit-playwright；Windows 为 %LOCALAPPDATA%\Tabbit\LocalAgent\bin\tabbit-cli.exe
  pageAccess: ask         # ask（每会话询问一次）| always | never
  intranetFetch: ask      # web_fetch 访问内网/回环目标：ask（每会话每 origin 询问一次）| always | never

Instance Resolution Priority

Priority config when multiple Tabbit versions are present

1. tabbit.instance explicitly set;
2. The Tabbit instance currently viewing dsh-web (auto-detected: on page load, the client plugin pings /tabbit/instance-hint; the server matches the instance registry's browserPid via the peer process of the loopback socket; macOS only — naturally not matched when opened in a non-Tabbit browser) — "whichever Tabbit you view it in is the one it runs in";
3. Inherited TABBIT_PLAYWRIGHT_INSTANCE environment variable (embedded form: when Tabbit launches the bundled dsh, it injects its own instance id, which is the authoritative channel);
4. Registry auto-selection (unique online instance; when ambiguous, reports a guided error with the instance list. On Windows, if the registry is unreadable, delegates to the native CLI for selection).

The currently effective source can be viewed via the /tabbit-info command (execution instance: ... (via ...)).

Full access explanation: dsh's danger-full-access permission preset writes the approval policy never into the session (dsh defines this as "automatically deny all prompts"). When this plugin's permission gate detects this override, it auto-passes instead of issuing prompts that would inevitably be denied — Full access means full access (in this mode bash is already unrestricted, so blocking only browser tools provides no defensive value). Exception: when deployment-level defaults set approval to never (not a session override), the public API cannot read it, so it will still be denied; combinations without an approval channel (e.g., certain headless orchestrations) will receive a denial message pointing to tabbit.pageAccess: always as a workaround.

Permissions

• An independent pageAccess permission is designed for data security and login-state control, separate from filesystem/sandbox permissions, requiring user confirmation by default.
• Authorization records: after a successful tool call, subsequent calls in the same session are not re-prompted; failures are not recorded (retry will prompt again).

Development

Project development and testing depend on a local deepseek-harness environment. Since published @deepseek-ai/* packages on npm are generally lagging behind, they cannot be installed directly as dependencies. A local harness checkout must first be built, then the .dsh-harness symbolic link in the repo is pointed to it via the script below (this path is already in .gitignore). The entire process only requires this single configuration — neither tsconfig.json nor package.json needs changes.

Deepseek Harness code retrieval and installation (skip if already installed)

git clone https://github.com/deepseek-ai/deepseek-harness && cd deepseek-harness && pnpm install && pnpm build && cd ..

Modify the local Deepseek environment path and ln it to this project, replacing /path/to/deepseek-harness with your actual directory.

npm run link-harness -- /path/to/deepseek-harness   # 或设置环境变量 DSH_HARNESS_PATH

安装和编译

pnpm install && pnpm build   # tsc → lib/
npm test                     # 构建 + node --test tests/

Known Limitations / Roadmap

• Favorites/bookmark mentions are not yet supported.
• Screenshots into context require the current model routing to support image input.
• Limited regression testing on Windows.