@creait/dsh-tailnet-gateway

Reach a loopback-bound DeepSeek Harness from your own Tailscale devices — with Settings, Models and Plugins actually working — gated on Tailscale identity and, optionally, on which specific machine is asking.

The problem

Point a browser on your tailnet at a dsh instance bound to 0.0.0.0 and you get an app that is two-thirds working. Chat is fine. Settings says settings are unavailable in this browser, Models will not load, Plugins → Plugin configuration is empty, and the agent-preset picker is greyed out.

That is not a bug to route around. @deepseek-ai/dsh-client-connection puts a deliberate fence in front of a set of privileged RPC methods — settings.describe/update/replace/mutate, credentials.describe/set/unset, agentPreset.*, host.pickDirectory, host.openPath, llm.discoverModels — and lets them through only from loopback. It is defending against DNS rebinding and cross-site calls, and it means it: the privileged branch re-runs the trust check with an empty trusted-host list, so --trusted-host cannot open it. Anything else gets 403 forbidden.

The client half compounds it. It decides isLoopback from window.location.hostname alone, so over a tailnet URL the settings pages build a throwaway in-memory store and never send the RPC at all — which is why the symptom reads as a broken page rather than a permission error.

So binding to 0.0.0.0 buys a half-working app and exposes dsh to the whole LAN. Neither half of that trade is good.

The shape of the fix

dsh stays on 127.0.0.1 and is never directly reachable. This plugin listens on a second loopback port; tailscale serve publishes that port to the tailnet over TLS. Every request arriving there has already been through tailscaled, which terminates the TLS, stamps Tailscale-User-Login and X-Forwarded-For, and overwrites whatever the client sent — which is what makes the identity unforgeable. Requests that pass the gate are forwarded to dsh as what they genuinely are: a loopback call from the same machine. Requests that fail are refused with 403 and never touch dsh at all.

▼
text
复制代码
tailnet device ──TLS──▶ tailscale serve ──▶ gateway :7242 ──▶ dsh :7241
                        (stamps identity)   (gate + rewrite)   (127.0.0.1)

Two gates

Login. The Tailscale account behind the request must be allowed. An empty allowlist means the account that owns this node, so a single-user tailnet needs no configuration to be correct. This gate alone already excludes tagged servers — a machine joined with an auth key has no human owner and its login reads as tagged-devices.

Device. The specific machine must be on the allowlist, matched by its short tailnet name (laptop, not an IP, which moves). Off by default, because an empty allowlist with the gate on locks out the device that would turn it on. Fill the list first, then switch it on. This is what makes "my phone and my laptop, never that VPS" expressible even when the VPS is signed in as you.

Both gates fail closed on a missing fact: no identity header means no admission, and an unresolvable peer means no admission while the device gate is on — a request that did not come through tailscale serve therefore fails both. If tailscale status cannot be read at all, the last known peer table is kept rather than an empty one, because a tailscaled restart must not silently suspend the allowlist.

Settings → Tailnet Access

A top-level settings page: gateway on/off with its live listening address and the tailscale serve line to publish it, the login gate and its allowlist, the device gate and every peer on your tailnet with a one-click Allowed/Blocked control, and the loopback override below.

The config route refuses a write that would lock you out — turning the device gate on without ticking your own laptop is answered with an error rather than an unreachable machine, and the only way back from that would be a shell on the box.

The loopback override

trustGatewayClients (on by default) is what fixes the client half. The client cannot be told the truth at runtime: dsh-client-ui-settings reads connection.isLoopback inside its own apply(), and the connection plugin computes it inside apply() too, so a third plugin mutating the service afterwards would be racing a value that has already been read. The honest place to state it is the served bundle, so the gateway substitutes that one line as it proxies /plugins/@deepseek-ai/dsh-client-connection/client.js.

It is a single literal substitution, not a parse, on a file dsh serves uncompressed. If a dsh upgrade changes the line the substitution simply does not match: the app keeps working, the settings pages fall back to their stub store, and the log says so once.

Turning this off leaves the gateway doing its access half and gives the settings pages back their remote behaviour — which is to say, back to not working.

Why it must stay on loopback

The gateway's whole job is to convert an authenticated tailnet request into a loopback one, so it holds loopback trust. Bound to a public interface it would hand that trust to anyone who asks, because a direct caller can set both headers itself. resolveConfig clamps the bind address to 127.0.0.1/::1 and startGateway refuses to listen anywhere else.

Install

▼
bash
复制代码
dsh plugin --profile web add @creait/dsh-tailnet-gateway

Restart dsh web afterwards: the boot manifest is built at startup.

Then bind dsh itself to loopback — it must not be reachable on the tailnet directly, or the gate is just a second door into an already-open room — and publish the gateway:

▼
bash
复制代码
tailscale serve --bg <gateway port>

Settings → Tailnet Access shows that command with the port it actually bound.

Settings

Key	Default	Meaning
enabled	true	Run the listener at all.
host	127.0.0.1	Bind address; loopback only, enforced.
port	7242	The port tailscale serve publishes.
requireLogin	true	Demand a Tailscale identity on every request.
allowedLogins	[]	Empty means the login that owns this node.
deviceAllowlist	false	Also gate on which machine is asking.
allowedDevices	[]	Short tailnet names, e.g. laptop.
trustGatewayClients	true	Tell the served client this hop is loopback.
statusTtlMs	30000	How long one tailscale status read is reused.

Tests

▼
sh
复制代码
node --test test/*.test.js

gate.test.js pins the admission table and the header rewriting; proxy.test.js drives a real gateway over real sockets against a stand-in dsh, and checks the things only a running server shows — that a refused request never reaches the upstream at all, that the bundle is rewritten in flight, and that an upgrade is tunnelled rather than answered.

What breaks this

The loopback override is a text substitution against a build artifact. The client bundle is minified but not mangled beyond recognition, and the marker it looks for is the compiled form of one property initialiser in dsh-client-connection. A harness upgrade that renames the helper, reorders the object, or changes how the hostname is read moves that marker; the rewrite then matches nothing. It fails open in the safe direction — the bundle is passed through byte-for-byte, the access half keeps working, and the settings pages go back to being read-only over the tailnet — and it says so once in the log rather than silently. If that happens, either update the marker or turn trustGatewayClients off until you do.

Both gates rest on tailscale serve stamping Tailscale-User-Login and X-Forwarded-For and, crucially, on it overwriting whatever the client sent. That is what makes the headers evidence rather than a claim. Anything else in front of the gateway — a reverse proxy, a port-forward, a second hop — removes that guarantee, which is why the bind address is clamped rather than configurable. A gateway reachable from anywhere but loopback is a gateway that hands loopback trust to whoever asks for it.

Identity is resolved by shelling out to tailscale status --json and reading Self, Peer and User. Those field names are stable across the versions this was built against but are not a documented API. A read that fails keeps the last good table rather than falling back to an empty one, so a transient tailscaled restart refuses nothing; a permanently broken read means the device gate refuses everything, which is the direction you want to fail in.

settings.section is a pre-1.0 slot and the config routes live on the plugin's own paths rather than the settings RPC, because that RPC is behind the same browser-trust fence this plugin exists to cross. peerDependencies pins the versions this was built against; a harness upgrade can move them.

License

MIT