dsh-tailscale-console

A control panel for the DeepSeek Harness web GUI that operates secure remote access over Tailscale: one-click health checks, HTTPS entry (Tailscale Serve) toggle, macOS proxy bypass repair, relay-server ops, and an ACL snippet generator.

UI language: Chinese. Docs: English · 中文

---

Features

Card	Description
① Devices & online alerts	Per-device online/offline state; auto-alerts when an offline device comes online
② Health check	One-click verification of HTTPS entry, page, /api session list, /sidebar/api, server direct path, proxy bypass
③ HTTPS entry	Tailscale Serve status / URL + on/off toggle (with verification & retry)
④ Proxy bypass	Checks *.ts.net / 100.64.0.0/10 in the macOS system proxy bypass list; one-click re-apply (macOS only)
⑤ Relay server	Tailnet status of your relay server + enable Peer Relay + ping verify
⑥ Access	Remote HTTPS URL, local URL, and the exact dsh web start command
⑦ ACL snippet	Generates the tailscale.com/cap/relay grants JSON from live device IPs

---

Full Setup Guide (server → local machine → DSH)

Target architecture: the server joins the tailnet as a Peer Relay; the local machine exposes HTTPS via Tailscale Serve; DSH keeps listening on 127.0.0.1 only.

0. Prerequisites

• Tailscale account (enable 2FA), MagicDNS enabled
• A Linux server with a public IP (Ubuntu 22.04 in this guide)
• Local macOS with the tailscale CLI installed and logged in
• A dsh web profile (~/.dsh/profiles/web) — requires dsh ≥ 0.1.0-rc.6 (for --trusted-host and the webRuntime service)
• Node 20.x + pnpm ≥ 9 for the profile (pnpm-lock.yaml is v9; the default pnpm 8 under Node 24 will fail. Verified with Node 20.19.2 + pnpm 10.27.0)
• A phone with the Tailscale app installed and logged into the same account (for remote access testing)

How to find your tailnet values (the <...> placeholders below):

• Local tailnet IP: tailscale ip -4
• All devices: tailscale status
• MagicDNS name (<hostname>.<tailnet>.ts.net): tailscale status --json | python3 -c "import json,sys; print(json.load(sys.stdin)['Self']['DNSName'].rstrip('.'))", or read it from the Tailscale admin console → Machines, or from the output of tailscale serve status once serve is up (section 2.1)
• Server tailnet IP: tailscale status on the Mac (the Linux peer)

Tip: you can bring the panel up locally first with dsh web (loopback default) — card ⑥ then shows the real remote URL, start command and ACL JSON, which you can copy back into the steps below.

1. Server side

1.1 Install & authenticate Tailscale

ssh <server> 'curl -fsSL https://tailscale.com/install.sh | sh'
ssh <server> 'nohup tailscale up >/tmp/ts-up.log 2>&1 & sleep 4; cat /tmp/ts-up.log'
# Open the https://login.tailscale.com/a/xxxx link printed, authorize with your account.
# The server comes online automatically; `tailscale status` on the Mac should list it.

1.2 Enable the Peer Relay port + firewall

ssh <server> 'tailscale set --relay-server-port=40000 && ufw allow 40000/udp && ufw allow 41641/udp'
# 41641 = WireGuard direct port; 40000 = Peer Relay port

⚠️ Your cloud security group must also allow inbound UDP 40000 and 41641 (source 0.0.0.0/0) — the guest ufw cannot protect the cloud edge.

1.3 Verify the direct path (run on the Mac)

tailscale ping -c 3 <server-tailnet-ip>   # expect "via <public-ip>:41641" (tens of ms); DERP means 41641 is blocked

2. Local machine

2.1 HTTPS entry: Tailscale Serve (the core)

tailscale serve --bg 3080    # https://<hostname>.<tailnet>.ts.net → 127.0.0.1:3080, tailnet-only
tailscale serve status       # inspect; `tailscale serve reset` to disable

Why HTTPS is mandatory: browsers expose crypto.randomUUID only in secure contexts (HTTPS or localhost), and every DSH client RPC depends on it. The GUI at http://<tailnet-ip>:3080 will never work — debug only.

2.2 Proxy bypass (required if you run Clash-family proxies on macOS)

A local Clash proxy hijacks the private ts.net domain and breaks HTTPS in the browser. Add *.ts.net and 100.64.0.0/10 to the bypass list in Clash Verge → Settings → System proxy; or apply immediately:

networksetup -getproxybypassdomains "Wi-Fi"   # first: back up the current list
networksetup -setproxybypassdomains "Wi-Fi" "*.ts.net" "100.64.0.0/10" "*.local" "<local>" "localhost" "127.0.0.1" "192.168.0.0/16" "10.0.0.0/8" "172.16.0.0/12"

⚠️ -setproxybypassdomains replaces the whole exception list of that service — back it up first with -getproxybypassdomains.
Non-macOS: the panel hides card ④ automatically; for Linux/Windows set the equivalent bypass, e.g. export NO_PROXY="*.ts.net,100.64.0.0/10" in your proxy tool or environment.

2.3 (Optional) socat debug forward

brew install socat
socat TCP-LISTEN:3080,bind=<local-tailnet-ip>,reuseaddr,fork TCP:127.0.0.1:3080
# curl/API debugging only; the GUI is unusable over plain HTTP (non-secure context)

3. DSH configuration

3.1 Start command (use this every time)

dsh web --trusted-host <local-tailnet-ip>:3080 --trusted-host <hostname>.<tailnet>.ts.net

⚠️ Every change below (cordis.patch.yml, plugin code, better-sidebar patch) requires restarting dsh web, which interrupts the currently running session — save/finish your work first. The web profile has HMR disabled.

3.2 Static trustedHosts (belt & braces)

Append to ~/.dsh/profiles/web/cordis.patch.yml:

- id: connection
  config:
    trustedHosts: ['<local-tailnet-ip>:3080', '<hostname>.<tailnet>.ts.net']

3.3 Known third-party issue: dsh-better-sidebar

If your profile installs dsh-better-sidebar, its /sidebar/api fence has a bug in version 0.10.3 (the latest published): trustedHostsOf() matches entry.options.name against "connection", but name is the package name, so the trust list is always empty and /sidebar/api/* is loopback-only (403 remotely — the sidebar breaks). No fixed upstream version exists yet.

Patch trustedHostsOf() in node_modules/dsh-better-sidebar/lib/index.js (back up the file first) to read the webRuntime service:

function trustedHostsOf(ctx) {
  const hosts = []
  const runtime = ctx.get("webRuntime")
  if (runtime !== void 0 && Array.isArray(runtime.trustedHosts)) hosts.push(...runtime.trustedHosts)
  for (const entry of ctx.loader.entries()) if (entry.options.id === "connection") {
    const cfg = entry.options.config?.trustedHosts
    if (Array.isArray(cfg)) hosts.push(...cfg)
  }
  return [...new Set(hosts)]
}

⚠️ The patch lives in node_modules and is lost on the next pnpm install — re-apply after every reinstall (check with grep -c 'options.id === "connection"').

4. ACL grant (required for phone relaying)

console.tailscale.com/admin/acls → JSON editor, add to grants:

{
  "grants": [
    {
      "src": ["<device-A-ip>", "<device-B-ip>"],
      "dst": ["<server-tailnet-ip>"],
      "app": { "tailscale.com/cap/relay": [] }
    }
  ]
}

(src = devices allowed to relay, dst = the relay server. Without it, phone traffic falls back to official DERP, 400ms+. Card ⑦ generates this for you.)

5. Install this plugin

Installed from npm: dsh-tailscale-console. For local development (editing this repo in place), use the file: reference instead.

cd ~/.dsh/profiles/web
# Node 20.x + pnpm >= 9 required (lockfile v9). Verified: Node 20.19.2 + pnpm 10.27.0.
pnpm add dsh-tailscale-console
# 本地开发：pnpm add "dsh-tailscale-console@file:./packages/dsh-tailscale-console"

Append to cordis.patch.yml:

- insert:
    - id: tailscale-console
      name: 'dsh-tailscale-console'
      config:
        sshAlias: my-server    # optional: ssh alias for server-side mutations (install/relay)

Restart dsh web → Settings → Tailscale Console.

6. Verification checklist

# HTTPS entry
curl -s -o /dev/null -w "%{http_code}\n" --noproxy '*' https://<hostname>.<tailnet>.ts.net/      # 200
# /api fence
curl -s --noproxy '*' -X POST https://<hostname>.<tailnet>.ts.net/api/session.list \
  -H 'content-type: application/json' \
  -d '{"type":"client-request","rpcId":"t-1","method":"session.list","payload":{}}' | head -c 120  # ok:true
# panel host route
curl -s -o /dev/null -w "%{http_code}\n" --noproxy '*' -X POST \
  https://<hostname>.<tailnet>.ts.net/tsctl/api/config -H 'content-type: application/json' -d '{}'   # 200
# direct path
tailscale ping -c 3 <server-tailnet-ip>
# Final: open https://<hostname>.<tailnet>.ts.net on any tailnet device — the sidebar lists all sessions

---

Configuration

All values come from the plugin Config; the code contains no personal identifiers. MagicDNS URLs and tailnet IPs are auto-derived from tailscale status --json when not configured.

Key	Type	Default	Meaning
dshPort	number	3080	DSH web GUI port
relayPort	number	40000	Relay server's Peer Relay UDP port
tailnetUrl	string	auto	HTTPS base, e.g. https://myhost.tail1234.ts.net
sshAlias	string	—	SSH alias for server-side mutations (install / relay)
serverPeerPattern	string	first Linux peer	Regex to identify the relay server peer
proxyServices	string[]	macOS set	Network services for the bypass apply

Security model

• Fixed command allowlist: every button maps to a hard-coded command; the client cannot pass free-form input.
• Sandbox: commands run with an explicit danger-full-access sandbox policy (same as the model's bash tool session policy). The default workspace-write confinement kills the detached child of tailscale serve --bg, breaking the HTTPS "on" toggle. Blast radius stays bounded by the allowlist.
• Browser-trust fence: the /tsctl/api prefix accepts only loopback Hosts or hosts in dsh's --trusted-host list (same fence as dsh's own /api).
• Platform: proxy bypass cards are macOS-only; other platforms degrade gracefully.

Troubleshooting

Symptom	Cause	Fix
Panel missing after restart, host route 200	exports lacks "./package.json" — client scan (require.resolve('<pkg>/package.json')) is blocked by Node exports encapsulation	add "./package.json": "./package.json" to exports; restart
Stale code served after editing sources	pnpm file: deps are hard-linked copies	re-symlink: ln -s ../packages/dsh-tailscale-console node_modules/dsh-tailscale-console
HTTPS "on" fails silently	sandbox kills serve --bg's daemon	fixed in code (explicit policy + verify/retry)
"On" button dead after HTTPS entry is off	chicken-and-egg: the HTTPS page itself is unreachable once serve is off (connection refused)	open the panel at http://127.0.0.1:<port> on the local machine and toggle on (remote page shows a hint)
Settings/credentials page 403 remotely	dsh PRIVILEGED_METHODS design limit	operate locally; do not relax
Phone access is slow	ACL grants not saved	see section 4
tailscale ping (TSMP) works, but real TCP/ICMP all time out	control-plane ACL lost the default allow rule (only custom rules like the Peer Relay grant remain); tailscaled logs show Drop: ... no rules matched	re-add the default rule at the top of grants (below) — propagates in ~30s, no restart

Default ACL rule (new grants format — three requirements)

When configuring a Peer Relay, the default "allow all connections" rule is easily overwritten. Once lost, every real packet is dropped by the ACL filter with no rules matched, while TSMP ping keeps working (internal path, bypasses ACL) — the classic "ping works, data doesn't" illusion. Always keep the default rule first:

{
  "grants": [
    { "src": ["*"], "dst": ["*"], "ip": ["*"] },
    { "src": ["<device IP>"], "dst": ["<relay server IP>"],
      "app": { "tailscale.com/cap/relay": [] } }
  ]
}

New grants requirements: dst takes no port; an explicit ip field is mandatory; app and ip cannot both be empty.

Inspect commands:

• Server: journalctl -u tailscaled | grep "no rules matched"
• macOS: log show --last 30m --predicate 'eventMessage CONTAINS "no rules matched"'

License

MIT. The browser-trust fence mirrors dsh's api-request-trust logic (see @deepseek-ai/dsh-client-connection).