Harness Web Bridge

English · 中文

Use your logged-in web AI — DeepSeek, GLM (Z.ai), Kimi, Qwen, Doubao — as model providers inside DeepSeek Harness. Web models emit tool calls, DSH executes them under its native permission and approval system, and the results go back into the same web session.

No API keys. It drives the site through your own browser login.

dsh plugin --profile web add dsh-webcode-bridge

Why

Most "web model" integrations are fragile in the same three ways: they scrape HTML, they lose the conversation when the page reloads, and they bypass the host's tool permissions. This plugin is built the other way around:

Features

• Real site rendering in the DSH right sidebar. Not a screenshot, not a mock — the actual site, loaded through a same-origin mirror so you can type, scroll and click in it. Login state persists.
• Per-site protocol teaching. Each site is taught the tool-call shape it actually accepts (DeepSeek's native token format, GLM's ``json code block,` tags elsewhere). The protocol is recomputed from the live tool list every time, so it never drifts.
• Per-site and global settings. Model, instruction, send gap and prompt-delivery mode can each be set globally and overridden for a single site — changing one site never disturbs the others.
• Multi-account. One site can hold several independent logins; each gets its own browser profile and its own request pacing.
• Observable. Every delivery, fallback and rate-limit event is projected into GET /__webcode/status, so "did it actually work" is answerable from the UI instead of from logs.

Requirements

Install

From npm (recommended)

dsh plugin --profile web add dsh-webcode-bridge

From a release tarball

Download dsh-webcode-bridge-<version>.tgz from Releases, then:

dsh plugin --profile web add C:\path\to\dsh-webcode-bridge-<version>.tgz

From source

cd package/dsh-webcode-bridge
pnpm install            # do NOT add --frozen-lockfile, see doc/ci-cd.md
pnpm pack              # produces dsh-webcode-bridge-<version>.tgz

Restart dsh web after installing. Installing only swaps files on disk; the running process still has the old code loaded. This is the single most common "it does not work" report.

Quick start

1. Start — dsh web, then open http://127.0.0.1:3080.
2. Sign in to a site — open the native Settings → Web Bridge panel and click Log in for the site you want. A real browser window opens; once you are signed in it closes and the session is reused headlessly from then on. Each site keeps its own login.
3. Pick a model — the model picker gains a Harness Web Bridge group listing every model from every signed-in site. Choose the default for new sessions in Settings.
4. Use it — the site appears in the DSH right sidebar (the official @deepseek-ai/dsh-client-ui-sidebar-right tabs; no third-party sidebar plugin involved). Ask a question there and the web model runs with local tools.

Tool calls produced by the web model run on your machine under DSH's normal permission prompts, and the results are returned to the same web conversation.

Configuration

All settings live in Settings → Web Bridge and can also be edited through the standalone settings page. Settings are stored per profile.

Global

Per site

Each site gets its own tab where you can override, independently of the global value:

• Account / login — one row per login on that site, with add/switch.
• Model — which model this site should use.
• Instruction — an extra instruction that only applies to this site.
• Send gap — its own pacing; Follow global clears the override.
• Prompt delivery — Follow global / Attachment / Plain text.

The fallback chain is always site override → global → built-in default, and clearing an override deletes the key rather than writing a third value — so "never configured" and "configured to follow global" stay distinguishable in the settings file.

Environment variables

Site status

DeepSeek is the primary target and holds up over long sessions. GLM works but is unstable.

What "unstable" means for GLM, concretely:

• Thinking cannot be turned off, and it dominates the stream (1051 of 2472 replies were thinking-only).
• The site intercepts tool-call tags in prose through its own native tool layer, so GLM is taught only the ```json code-block shape, with fallback parsing from the thinking stream.
• Deep-link navigation is stopped by an Aliyun slider captcha, so probes go through the driver path.
• Two failure modes have been fixed: a stream that died mid-thinking produced an empty reply, and call-fence tails leaked into the reply body as garbage.

Read it as: DeepSeek is the one to run all day; GLM is a good second opinion. When GLM fails, retry — attributable failures surface as coded messages rather than silent degradation.

How it works

A few design notes worth knowing before you read the source.

Right sidebar, one origin per site

Each site is served from http://<siteId>.localhost:8931/, so the pathname the site sees matches its real site and SPA routers and root-relative assets work unchanged. The exception is DeepSeek, which is always mounted at the relay root — it validates the hostname, and a subdomain gives a blank Unknown hostname page. *.localhost resolves to loopback, so the exposure surface is unchanged.

Sites often load scripts and styles from a different origin with crossorigin, and that origin's Access-Control-Allow-Origin may be an illegal wildcard (https://*.deepseek.com), which browsers refuse outright. lib/mirror.js therefore rewrites those URLs to same-origin /__static/<host>/…, strips integrity and crossorigin, rewrites the runtime fetch/XHR calls that target those origins, and installs an idempotent runtime hook to catch lazily-built chunk URLs.

Order matters: rewrite the site HTML first, then inject the bootstrap. Reversed, the constants inside the bootstrap get rewritten too and the runtime comparison never matches. test/mirror.test.mjs locks this in.

Prompt delivery

Bodies over the threshold go up as an attached .md file instead of being typed into the composer, which sidesteps two measured failure modes: PROMPT_WRITE_STALLED (the composer stops growing while being written) and PROMPT_TRUNCATED (the site accepts only half). Any failure on that path falls back to plain text — delivery is an implementation detail, "the message gets sent" is the contract.

Two sites (DeepSeek, Kimi) have their own lower thresholds because their composers were measured to truncate earlier than the global default. Sites where upload is confirmed to not be read are downgraded at runtime and remembered.

Token accounting

Token estimates use three measured densities (0.19.3 onward) — CJK 0.75, prose ASCII 0.30, other ASCII (digits, punctuation, source) 0.70 — plus a 10% margin. The rule is that the estimate never understates; the single source of truth is TOKEN_DENSITY in lib/metrics.js. Speed figures are measured from the site's own SSE stream and are kept separate from the estimate.

Troubleshooting

GET /__webcode/status exposes the live readouts (last delivery, transport mode, session slot, rate-limit retries, fresh-conversation reasons) — that is the first place to look when the UI is ambiguous.

Development

cd package/dsh-webcode-bridge
pnpm test              # regression, parsing and Harness adapter contracts
pnpm lint:comments     # comment discipline gate
pnpm ci:local          # full local CI

Useful single-file runs:

node test/glm-session-replay.test.mjs        # replay a real GLM session shape offline
node test-mock/real-mirror-matrix.mjs        # right-sidebar acceptance matrix, 10 sites
node test-mock/run-real-longrun.mjs          # long real-browser tool loop
node scripts/verify-pack.mjs                 # tarball vs working tree, byte for byte

Documentation

License

MIT — full text in LICENSE. Third-party notices are kept in permissions and boundaries rather than in the license body, so GitHub's license detection reports MIT instead of NOASSERTION.

SSE decoding protocol references the MIT project three-water666/webcode.