dsh-lan-access

A DeepSeek Harness web plugin that adds a LAN access toggle to the DSH Settings shell (Settings → General). It replaces the manual cordis.patch.yml webserver override:

Tested with DeepSeek Harness v0.1.5-rc.2 — this plugin is verified runnable against that harness version.

• On — the web GUI binds 0.0.0.0, so other machines on the same network can open it at http://<LAN-IP>:3080/?token=…. dsh web prints the full URL (with the per-process ?token= launch token) for the LAN address when it starts — a fresh LAN browser needs that token in the URL to authenticate. The /api trust fence is updated live, so the browser on a LAN machine works fully (chat, tools, workspace).
• Off — the GUI binds 127.0.0.1 again (loopback only — the safe default).

Screenshots

The DSH web GUI opened from another machine on the same network (http://192.168.0.101:3080):

The LAN access toggle in Settings → General, showing the address other devices can open:

How it works

The route fence accepts loopback or the deployment's trusted authorities, read live from the connection row's resolved config — the same boundary the /api gateway uses. Cross-site requests are refused.

Install from GitHub

The built artifacts (lib/) are committed, so installation needs no build step and no modification of the DeepSeek Harness checkout:

# From GitHub (replace <owner>/<repo>)
dsh plugin --profile web add git+https://github.com/<owner>/<repo>.git

# ...or clone and install the local checkout (link: keeps your rebuilds live)
git clone https://github.com/<owner>/<repo>.git
dsh plugin --profile web add link:/path/to/dsh-lan-access

# Restart the GUI
dsh web

The install appends dsh-lan-access to dsh.profile.bundles; its dsh.bundle.patch inserts the host row and overrides the webserver row's host with the lanAccess service expression. The client half is picked up by the client-modules scanner automatically. No harness change is required for the core feature — the toggle, the LAN bind, and the live /api trust fence all ship inside the plugin.

Local development — rebuild with pnpm build (or npm run build) after changing src/, then reinstall/restart. The repo's node_modules mirrors the DSH profile's package farm (TypeScript/tsdown come from the harness checkout).

Migrating from a manual patch — remove any webserver host: 0.0.0.0 override from the profile's cordis.patch.yml (and the bundle patch layers) so the plugin is the single owner of the bind host.

Use

1. Open the GUI, go to Settings (sidebar footer) → General.

2. Flip 局域网访问 / LAN access.

   • Enabling shows the ONE address other devices can open — the IPv4 of the interface that owns the default route (http://192.168.x.x:3080) — with a copy button.
   • That address must include the browser-session ?token=… when opened from another machine. dsh web prints the full URL (with the token) for both the loopback and the LAN address on startup — copy the LAN one, e.g. http://192.168.0.101:3080/?token=ICKD2317KYP…. The token is a per-process launch token that exchanges for a session cookie; a fresh LAN browser cannot authenticate without it.
   • The web server restarts to rebind; the row waits for it and re-reads the state (a network error mid-restart is not reported as failure).
   • The plugin also installs a crypto.randomUUID polyfill on plain-HTTP LAN origins (that Web API only exists in secure contexts, and the DSH API client mints every RPC id with it — without the polyfill a remote browser fails with "crypto.randomUUID is not a function").

3. The choice is persisted in ~/.dsh/settings.yaml:

   lan-access:
     enabled: true
   

Remote Settings pages and workspace — no harness change needed

Everything the plugin serves works from a LAN browser with zero modification of the DSH checkout:

• The current /api gateway trusts the served LAN authority, so the configuration plane (settings.*, credentials.*) reaches the host directly — the Host/Origin fence admits the LAN host and the ordinary browser-session auth authenticates it. The Models page provider directory, the Plugins configuration cards, and the Language/Appearance rows therefore work remotely with no extra hop.
• The client settingsScope degrades to memory mode on non-loopback origins (surfaces render empty). The browser bundle widens connection.isLoopback to "loopback OR served LAN authority" at runtime. The client entry injects connection and is marked dsh.client.immediately, so its bundle is prefetched and its apply runs right after the connection row provides the handle — before any settings surface (which waits on remote / settingsScope) reads remote.$host.isLoopback. That inject-ordered widening is what keeps the scope in host mode on a LAN page: without it, a surface that binds early sees the unpatched isLoopback and its scope stays memory-mode — the plugin configuration cards render nothing.
• crypto.randomUUID does not exist on plain-HTTP LAN origins. The bundle installs a getRandomValues-based polyfill (same CSPRNG).

Harness API changes this plugin tracks

The harness evolved the settings/connection APIs between 0.1.0-rc.5 and 0.1.5-alpha.1; the plugin was updated accordingly:

• @deepseek-ai/dsh-settings dropped the settingsNamespace(ns) helper — settings.register / .update / .replace / .mutate now take the raw namespace string (validated at runtime and by a compile-time guard).
• The client ConnectionHandle no longer carries an api member — remote methods go through connection.rpc.call('/api', '<ns>/<method>', …) — so the plugin no longer patches connection.api.settings.* / connection.api.credentials.*.
• The /api gateway stopped pinning the configuration plane to loopback: it now trusts the served LAN authority, so the settings/credentials RPCs reach the host directly. The plugin therefore widens connection.isLoopback early (via inject: ['connection'] plus a synchronous patch) to keep settingsScope in host mode on a LAN page; the fenced /lan-access/rpc proxy is no longer required for the remote Settings surfaces.

Remaining loopback-only (hardcoded in the harness, not patchable from a plugin): host.pickDirectory / host.openPath (native dialogs and host file opens) and llm.discoverModels (the Models page "discover" button). The workspace's own add/browse flow does not need them, and chat file opens route into the sidebar editor.

Debug aids

The host exposes GET /lan-access/diag (fenced like the other routes) with the latest browser boot reports: slot-registration counts, whether the connection patch is active, and a settingsScope probe bound to the shell namespace (status ready proves the host-mode settings read works end to end). During the first minute after boot the browser also posts a 2-second poll of the Plugins cards' own injected snapshots (available flags), the slot ledger view, and the declared spec — the exact data that separates "cards gone", "cards abdicated", and "cards present but rendering null" when a Settings page misbehaves on a remote machine.

Optional: dsh-better-sidebar compatibility patch

dsh-better-sidebar's trust fence matched the connection row by the wrong name and read the raw !!js config, so its panels (explorer / editor / terminal / git) only ever accepted loopback. The repo ships the fix as a profile-level pnpm patch (no harness change):

./scripts/install-patches.sh web

This copies patches/dsh-better-sidebar.patch into the profile's patches/ directory, registers it under patchedDependencies in pnpm-workspace.yaml, and runs pnpm install.

Security notes

• Default is off (loopback). The DSH launcher itself refuses --host 0.0.0.0 for the same reason: binding all interfaces exposes the agent's tools to the network. Only enable it on a trusted network.
• The toggle is only reachable through the fenced route, and disabling from a remote machine cuts that machine off (expected — re-enable locally).
• The bind survives plugin reloads and patch re-applies; a full process restart re-applies the persisted value at boot.

Development

pnpm build        # tsdown: lib/index.js (host) + lib/client.js (browser bundle)
pnpm typecheck    # tsc --noEmit

The client bundle is a __ModuleLoader__.load closure-factory artifact (same format as the DSH monorepo's tsdown client preset); only the frozen platform-module table words stay external. After changing client code, rebuild and restart dsh web (the client-modules package metadata cache expires only on restart).