DeepSeek Harness Phone Remote

Don't watch your agent. Keep its pulse.

A secure, zero-app remote workspace for DeepSeek Harness. Reach the real Harness
web UI from your phone over Tailscale (or LAN), manage files and workspaces, and
keep an ambient eye on your agents through Agent Presence — the floating Orb,
Task Board and notifications.

• 🎈 Agent Presence — floating brand Orb (drag anywhere): Needs You / Failed /
  Possibly Stalled / Running / Done / Idle / Disconnected, with a Task Board and
  browser notifications for what needs you.
• 📁 Files & workspaces — start/resume an agent in any approved folder; read,
  write, upload, download files.
• ↔ Handoff — leave the desk and continue the same session from your phone
  (sessions.open — never a replacement session).
• 🔐 Device pairing — one-time code, per-device credentials, revocation.

This project does not replace the Harness UI. It turns Harness itself into a
remote work environment: the phone opens the real DeepSeek Harness web UI over
Tailscale (or your LAN), and a persistent plugin bridges the gaps a browser
can't close remotely — authenticated file/workspace access and Agent Presence
(Orb / Task Board / notifications).

English | 中文 · Architecture · Security · Contributing

Why

• Harness binds 127.0.0.1 — a deliberate, sane default. The Harness process
  itself stays loopback-bound; only the opt-in forwarders (Tailscale IP and,
  when enabled, the LAN IP) expose selected interfaces.
• A phone browser still can't reach loopback, and the GUI's directory picker is
  a loopback-only privileged method — so this plugin adds a secure path
  (Tailscale + LAN forwarders) and a filesystem/workspace bridge for exactly
  those two gaps.
• Sessions normally die with the page — this plugin is a persistent loader
  entry, so the workbench loads on every page automatically, no per-session
  "run" needed.

Architecture

flowchart LR
  P[Phone / remote browser] -->|Tailscale HTTPS| S[tailscale serve]
  P -->|Tailscale IP| T[TCP forwarder]
  P -->|same Wi-Fi: LAN IP| L[LAN forwarder]
  S --> H[DeepSeek Harness Web<br/>127.0.0.1:3080]
  T --> H
  L --> H
  H --> R[/remfs RPC channel<br/>trusted-host fence/]
  R --> A[Device authentication<br/>pairing + per-device credential]
  A --> F[Filesystem capability layer<br/>allowlist + protected paths + realpath]
  F --> W[(Approved workspace)]

Three independent layers:

1. Transport — who can reach the channel: Tailscale membership or your LAN
   (forwarders only bind the Tailscale IP and the LAN IP; never 0.0.0.0).
2. Application — who may use it: device pairing + per-device credentials.
3. Capability — what they may touch: the allowlist + protected paths.

trusted-host and the tailnet are transport trusts. They are not
authentication. Pairing and the filesystem capability layer are.

Features

• One-click deploy (auto-installs prerequisites) — install.ps1 validates
  the Node version (^22.19 || >=24), installs missing Node.js / Tailscale
  (winget), guides the one-time Tailscale sign-in (re-reads the real MagicDNS
  name — never a fabricated one), writes the launcher, enables HTTPS Serve,
  installs the plugin, registers auto-start on login.
• Walk-on-LAN (opt-in) — off by default. Create %USERPROFILE%\.dsh\lan-on
  (or set DSH_REMFS_LAN=1) to trust the LAN IP and start the LAN forwarder;
  on the same Wi-Fi the phone can then skip Tailscale (http://192.168.x.x:3080).
  /remfs stays device-authenticated. Enabling it widens the network exposure,
  so it is an explicit choice.
• Persistent plugin — loader entry; host channel registers at startup, the
  client module loads on every page. No re-running after refresh.
• Device pairing & management — one-time pairing code (10 min TTL, single
  use); list / revoke / revoke-all devices; credentials stored only as hashes.
• Mobile-first workbench — New Session / Files tabs, breadcrumbs, preview /
  edit / upload / download, workspace badges, floating ball, auto-collapsed
  sidebar, bilingual UI (EN/zh).
• Host-enforced protected paths — system dirs, AppData, credential/key files
  (.credentials.yaml, .ssh, .aws, .gnupg, .env, id_rsa, *.pem …)
  and private data dirs (WeChat/WPS) are blocked regardless of the allowlist.

Security model

• Tailscale ≠ authentication. It proves which network you are on, not
  who you are. Device pairing is the application boundary.
• trusted-host ≠ authentication. It is the browser-trust fence (Host header +
  cross-site checks). Pairing is the boundary.
• The Harness process stays loopback-only. Exposing the web service on a
  network interface happens ONLY through the explicit forwarders (the Tailscale
  IP, and the LAN IP when walk-on-LAN is opted in) — those bind specific
  addresses, never 0.0.0.0.
• Pairing protects /remfs only, not the native Harness /api. The GUI's
  own API surface has no user login; keep the network boundary (tailnet / LAN)
  tight and review which devices can reach it.
• The filesystem allowlist is the primary file-permission boundary. Remote
  clients can only narrow it; widening (C:\, new drives) requires editing
  .remfs-roots.json on the PC.
• Path escape is defended twice: raw paths with ../UNC are rejected, and
  the canonical realpath must stay inside the allowlist (symlink/junction
  escapes fail).
• Remote writes are encoding-guarded: uploading or editing a file that is
  not UTF-8 (UTF-16 BOM, GBK/ANSI byte sequences) is rejected instead of
  corrupting it, and the UTF-8 BOM + dominant newline style (CRLF/LF) of an
  edited file are preserved on write-back.
• The presence read-only fence is an operator switch: by default the
  Orb/Task Board work unauthenticated inside the browser-trust fence; setting
  pocketStrict: true in ~/.dsh/remfs-options.json requires a valid device
  credential for every /pocket call.
• Tailscale ACLs should be hardened to phone-only access on 443/3080 — see
  docs/tailscale-acls.md.
• See SECURITY.md for the full threat model (what we do and do
  not protect), and docs/upgrade.md for the pre-upgrade
  backup + verification checklist.

Positioning

This project is a secure remote workspace & filesystem bridge for DeepSeek
Harness: it keeps the native web UI and adds authenticated remote access plus a
capability-bounded file/workspace layer. It is not a UI replacement, skin, or
alternative frontend — the ecosystem has other community projects for those
directions, and they are complementary rather than competing.

Installation

Advanced users (npm):

dsh plugin --profile web add @zetaluolang/remfs-persistent
# append to %USERPROFILE%\.dsh\profiles\web\cordis.patch.yml:
#   - insert:
#       - id: remfs-persistent
#         name: '@zetaluolang/remfs-persistent'
#         inject: [connection, fs]
# restart dsh web

Windows users (one-click): double-click 一键部署.cmd — it validates
the Node version (^22.19 || >=24), auto-installs Node.js + Tailscale, guides the
Tailscale sign-in, writes the launcher, registers the self-healing watchdog,
enables HTTPS Serve, installs the plugin and prints the phone URLs (HTTPS,
Tailscale IP, and the LAN IP when walk-on-LAN is enabled).

First use on the phone (pairing)

1. Open the phone URL (https://<pc-name>.<tailnet>.ts.net, or the LAN URL when
   walk-on-LAN is enabled and you are on the same Wi-Fi).
2. The workbench shows the pairing screen.
3. On the PC, read the pairing code from
   %USERPROFILE%\.dsh\remfs-pairing.txt (or the harness log).
4. Enter the code + a device name on the phone → paired. Credentials are stored
   on the phone; the PC stores only the hash.
5. Revoke devices anytime from the workbench ⋯ → Devices.

Self-healing watchdog

install.ps1 registers a Task Scheduler task (dsh_harness_watchdog, every
5 minutes, current user, hidden window) that runs
%USERPROFILE%\.dsh\launcher\watchdog.ps1. Each run:

1. Verifies our dsh process actually owns 127.0.0.1:3080 — the owning
   process's command line must contain the deployed dsh bin path (the watchdog
   reuses the launcher's Get-OwnedHarnessPid ownership check; a bare open
   port is never trusted, so an unrelated localhost service is never
   mistaken for the harness).
2. If our harness is down and the port is free, restarts it headlessly via
   restart_harness_once.ps1 (DSH_HEADLESS=1 — no browser, no dialogs) and
   appends every step to %USERPROFILE%\.dsh\launcher\watchdog.log.
3. If a foreign process occupies the port, it logs the conflict and stands
   down — it never kills or restarts over a process it does not own.

Re-run install.ps1 (or 一键部署.cmd) to update the task definition; the
watchdog itself checks in with one log line every 5 minutes when healthy.

Threat model

We defend against: unauthenticated RPC, remote allowlist widening, path escape,
credential theft at rest, accidental LAN/public exposure of the GUI.

We do not (yet) defend against: the harness GUI /api itself having no user
login (pairing protects /remfs, not the GUI — keep the network boundary
tight), a compromised host, or a compromised Tailscale account. Details in
SECURITY.md.

Troubleshooting

Tested devices

• OPPO Find X8 Ultra (real hardware).
• Emulated matrix: iPhone 16 Pro/SE, Pixel 8, Galaxy S24, Redmi Note, iPad Air,
  iPhone landscape — sidebar collapse, floating ball, panel width, no overflow.
  See docs/device-tests/.

Roadmap

• [x] Tailscale HTTPS + IP access, walk-on-LAN
• [x] Persistent plugin (no per-session run)
• [x] Device pairing + credential auth + revocation
• [x] Capability-bounded allowlist + protected paths + path-escape tests
• [x] Bilingual UI, security tests, CI
• [x] Startup quarantine of corrupt demo sessions + self-healing watchdog
• [x] Session size hints + "suggest archiving" in the phone UI
• [x] Demo-presence behavior tests (idempotent add, fixed cwd, clean-only-demos)
• [x] UTF-8 encoding guard on upload/edit + BOM/newline preservation
• [x] /pocket strict mode (opt-in via ~/.dsh/remfs-options.json)
• [x] Tailscale ACL hardening guide
• [x] Desktop shortcut automation (install.ps1)
• [ ] More device resolutions validation
• [ ] Upstream contributions (see below)

License

MIT