dsh-web-companion

A DeepSeek Harness plugin that puts the DeepSeek and Doubao web apps one click away in the sidebar, and lets the model read answers from the Doubao web app — including image recognition.

Both entries and the recognition tool share one persistent Chromium profile, so a single sign-in survives Harness restarts.

中文说明见 README.zh.md.

Features

• One-click web apps — two entries at the bottom of the left sidebar (root-scoped, so they exist even on the new-conversation screen): DeepSeek opens chat.deepseek.com, Doubao opens www.doubao.com/chat/. No new window per call: one browser, reused.
• Vision through the Doubao web app — doubao_vision pastes an image into the Doubao composer, sends it, and returns the reply text to the harness as a tool result.
• Persistent login — the browser runs with a fixed --user-data-dir, so cookies and site sessions are written to disk and reused across calls and restarts.

Requirements

Item	Value
Harness	DeepSeek Harness Desktop (Electron shell), plugin API as of 2026.11
OS	Windows (paths and browser auto-detection are Windows-specific)
Browser	A local Edge or Chrome build, or set browserPath / DSH_COMPANION_BROWSER
Node	≥ 22 in the harness host (the plugin uses the global WebSocket and fetch)

Install

The plugin is a DSH bundle: it is installed from a directory, not from npm.

# inside the harness
plugin_manager action=install_bundle target=C:\path\to\dsh-web-companion

application: applied means it is live. To reload after editing:

plugin_manager action=set_bundle target=@local/dsh-web-companion enabled=false
plugin_manager action=set_bundle target=@local/dsh-web-companion enabled=true

To remove it: plugin_manager action=remove_bundle target=@local/dsh-web-companion.

The package name stays @local/dsh-web-companion because the bundle row and the installed symlink are keyed by it. The repository itself is distributed under the name dsh-web-companion.

Usage

Sidebar entries

1. Bottom of the left sidebar: DeepSeek and Doubao. Each opens its page in the persistent browser.
2. The right sidebar's Start tab also carries a Web companion card that opens the panel.
3. The panel's own buttons open the same browser; only if the host route is unavailable do they fall back to a Harness Browser tab (that partition is ephemeral — you would sign in again).

Image recognition

Attach an image in the harness, then ask the model to recognise it with Doubao. The model calls:

{ "image_path": "C:\\path\\to\\image.png", "prompt": "What text is in this image? Answer with the text only." }

imageBase64 (raw or a data: URL) is accepted instead of image_path. The tool returns the Doubao reply as the tool result, so the model can answer from it.

Managing the browser

{ "action": "status" }                       // running?, profile dir, open pages
{ "action": "open", "target": "doubao" }     // open or switch to a page
{ "action": "login", "target": "deepseek" }  // open the sign-in page once

How it works

sidebar entry / panel  ──fetch──▶  GET /dsh-companion/open?target=…
                                        │
                                        ▼
                              CompanionBrowser (index.js)
                              ├── spawns Edge/Chrome with a persistent profile
                              └── drives it over CDP (127.0.0.1:9334)
                                        │
model ──doubao_vision──▶ paste image + prompt ▶ read reply bubble

• No private APIs. The Doubao side is driven through the page: a synthetic paste event carries the image into the ProseMirror composer, Input.insertText + Enter submit it, and the reply is read from [data-testid=union_message] → receive_message → message_text_content.
• One browser for everything. pageFor() reuses the page already on the target origin, so the entry, the panel and doubao_vision operate on the same tab and the same cookies.
• Why not an iframe? chat.deepseek.com sends Content-Security-Policy: frame-ancestors 'none', so it cannot be embedded. The plugin opens a real browser page instead.
• Why a loopback route? The Web UI's renderer cannot call harness tools directly, but it can fetch() a route registered through ctx.webServer. The route accepts loopback or null origins only.

Configuration

config in cordis.patch.yml:

Key	Default	Meaning
port	9334	CDP port for the plugin browser, bound to 127.0.0.1
browserPath	empty	Explicit executable; empty auto-detects Edge, then Chrome
profileDir	~/.dsh/web-companion/profile	Persistent profile directory
hidden	false	Start the window off-screen; keep false until the first sign-in
doubaoUrl / deepseekUrl	vendor URLs	Replace if a site moves
replyTimeoutMs	180000	Per-recognition timeout

Verification

1. Entries — both icons appear at the bottom of the left sidebar, including on the new-conversation screen. Clicking one opens the matching page in the persistent browser.
2. Recognition — attach an image and ask the model to read it with Doubao; the text comes back as a tool result.
3. Same session — click Doubao, leave the page open, then run a recognition; the conversation appears in Doubao's Recent list, proving one browser and one profile are shared.
4. Login reuse — sign in once, restart the harness, { "action": "open" } again; the page is still signed in. Cookies live in profile/Default/Network/Cookies.

Known limitations

• Site changes break selectors. Doubao UI updates may require adjusting the data-testid names in recognize() in index.js.
• Captchas may appear; the tool then times out and the browser window needs one manual action.
• Windows only for now: browser auto-detection assumes Windows install paths.
• Web harness has no Electron <webview>; pages open in the system browser instead.
• Recognition polls the DOM and returns once the reply text is stable (~1.2 s), up to replyTimeoutMs.

Development

Path	Purpose
index.js	Host half: browser lifecycle, CDP client, tools, loopback route
client.js	Client half: sidebar entries, right-sidebar panel, inlined logos
cordis.patch.yml	Loader row and default config
tools/fetch-logos.mjs	Re-download the vendor favicon assets
tools/embed-logos.mjs	Inline those assets into client.js as data URIs
assets/	Source artwork for the icons (not loaded at runtime)

Trademarks

DeepSeek and Doubao names and logos are trademarks of their respective owners. They are used here only to identify the services this plugin opens; the favicon copies live in assets/ and are inlined for identification purposes.

License

MIT