deepseek-harness-phone-remoteDeepSeek Harness plugin
DeepSeek Harness phone remote control via Tailscale - persistent file/workspace plugin - tested on OPPO Find X8 Ultra
- Stars
- 10
- Forks
- 0
- License
- MIT
- Last commit
- Aug 30, 2026
- Latest release
- v1.0.0
Overview
DeepSeek Harness phone remote control via Tailscale - persistent file/workspace plugin - tested on OPPO Find X8 Ultra
Original README
Cached from the project repository on Sep 3, 2026. This is source content, separate from the Agents.md review above.
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 โ the floating Orb, Task Board and notifications.
- ๐ Agent Presence โ 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.
- ๐ Files & workspaces โ start/resume an agent in any approved folder; read, write, upload, download files.
- โ Handoff โ leave the desk and continue the same session from your phone
(
sessions.openโ never a replacement session). - ๐ Device pairing โ 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 โ authenticated file/workspace access and Agent Presence (Orb / Task Board / notifications).
[!WARNING] Device pairing authenticates this plugin's
/remfsand protected/api. Treat tailnet/LAN reachability as access to the Harness control plane; keep walk-on-LAN off unless needed and restrict Tailscale ACLs to trusted devices.
English | ไธญๆ ยท Architecture ยท Security ยท Contributing
Why
- Harness binds
127.0.0.1โ 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 โ 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 โ this plugin is a persistent loader entry, so the workbench loads on every page automatically, no per-session "run" needed.
Architecture
mermaid1flowchart LR 2 P[Phone / remote browser] -->|Tailscale HTTPS| S[tailscale serve] 3 P -->|Tailscale IP| T[TCP forwarder] 4 P -->|same Wi-Fi: LAN IP| L[LAN forwarder] 5 S --> H[DeepSeek Harness Web<br/>127.0.0.1:3080] 6 T --> H 7 L --> H 8 H --> R[/remfs RPC channel<br/>trusted-host fence/] 9 R --> A[Device authentication<br/>pairing + per-device credential] 10 A --> F[Filesystem capability layer<br/>allowlist + protected paths + realpath] 11 F --> W[(Approved workspace)]
Three independent layers:
- Transport โ 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).
- Application โ who may use it: device pairing + per-device credentials.
- Capability โ 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) โ
install.ps1validates 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 โ never a fabricated one), writes the launcher, enables HTTPS Serve, installs the plugin, registers auto-start on login. - Walk-on-LAN (opt-in) โ off by default. Create
%USERPROFILE%\.dsh\lan-on(or setDSH_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)./remfsstays device-authenticated. Enabling it widens the network exposure, so it is an explicit choice. - Persistent plugin โ loader entry; host channel registers at startup, the client module loads on every page. No re-running after refresh.
- PC always-on-top Orb โ
scripts/orb-widget.ps1(double-clickscripts/start-orb-widget.cmd, deployed to~/.dsh/launcher): a zero-dependency WinForms ball that stays above every window, shows the highest-priority agent state, is draggable (position persisted), and opens the harness on click. Polls the same task DTOs (/remfs-presence.json).install.ps1also registers auto-start at logon (a Startup-folder entry; single-instance, so the auto-start and the .cmd never stack โ removedsh-orb-widget.cmdfrom the Startup folder to disable). - Phone push notifications (page closed) โ an opt-in Web Push path: the
phone registers a same-origin service worker (
/remfs-sw.js) and, once paired, subscribes with the host's VAPID key. The host pushes NEEDS_USER/FAILED (and DONE when enabled) transitions even with the tab closed. RFC 8291 encryption implemented from scratch (zero deps), verified against the official RFC test vector. HTTPS required (Tailscale HTTPS or localhost). See docs/presence-push.md. - Device pairing & management โ one-time pairing code (10 min TTL, single use); list / revoke / revoke-all devices; credentials stored only as hashes.
- Mobile-first workbench โ New Session / Files tabs, breadcrumbs, preview / edit / upload / download, workspace badges, floating ball, auto-collapsed sidebar, bilingual UI (EN/zh).
- Host-enforced protected paths, in two tiers:
- Hard-denied โ never reachable, regardless of the allowlist or of any
workspace you register: system dirs (
Windows,System32,SysWOW64) and credential/key files (.credentials.yaml,.ssh,.aws,.gnupg,.env,id_rsa,*.pemโฆ). - Soft-denied โ blocked by default, but reachable if you deliberately
register a workspace exactly there on the PC:
AppData,Program Files,ProgramDataand private data dirs (WeChat/WPS). These are privacy boundaries, not credential boundaries; registering such a folder is a local decision only the PC user can make, and it never unlocks the hard-denied files underneath.
- Hard-denied โ never reachable, regardless of the allowlist or of any
workspace you register: system dirs (
Security model
- Tailscale โ authentication. It proves which network you are on, not who you are. Device pairing is the application boundary.
- trusted-host โ 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) โ those bind specific addresses, never 0.0.0.0.
- Pairing protects
/remfsonly, 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.jsonon 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: truein~/.dsh/remfs-options.jsonrequires a valid device credential for every/pocketcall (and for/remfs-presence.json). - Push subscriptions are paired-device-only: registering a subscription
requires a valid device credential; revoked devices are pruned within a
minute; task titles/summaries are pushed only to paired devices. VAPID keys
live in
~/.dsh/remfs-push.json(never in the repo). - Tailscale ACLs should be hardened to phone-only access on 443/3080 โ 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 โ the ecosystem has other community projects for those directions, and they are complementary rather than competing.
Installation
Advanced users (npm):
bashdsh 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, sandboxPolicy, workspaceRegistry] # restart dsh web
Windows users (one-click): double-click ไธ้ฎ้จ็ฝฒ.cmd โ 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)
- 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). - The workbench shows the pairing screen.
- On the PC, read the pairing code from
%USERPROFILE%\.dsh\remfs-pairing.txt(or the harness log). - Enter the code + a device name on the phone โ paired. Credentials are stored on the phone; the PC stores only the hash.
- Revoke devices anytime from the workbench โฏ โ 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:
- Verifies our dsh process actually owns
127.0.0.1:3080โ the owning process's command line must contain the deployed dsh bin path (the watchdog reuses the launcher'sGet-OwnedHarnessPidownership check; a bare open port is never trusted, so an unrelated localhost service is never mistaken for the harness). - If our harness is down and the port is free, restarts it headlessly via
restart_harness_once.ps1(DSH_HEADLESS=1โ no browser, no dialogs) and appends every step to%USERPROFILE%\.dsh\launcher\watchdog.log. - If a foreign process occupies the port, it logs the conflict and stands down โ it never kills or restarts over a process it does not own.
Re-run install.ps1 (or ไธ้ฎ้จ็ฝฒ.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 โ keep the network boundary
tight), a compromised host, or a compromised Tailscale account. Details in
SECURITY.md.
Troubleshooting
| Symptom | Fix |
|---|---|
| Phone shows the pairing screen forever | Read the code from %USERPROFILE%\.dsh\remfs-pairing.txt; codes expire after 10 min โ restart the harness to generate a new one |
| Device revoked / re-pairing fails | Pairing codes are single-use; restart the harness for a fresh code |
| Phone gets 403 | Use the printed HTTPS/Tailscale/LAN URL; the GUI must run with those hosts trusted (one-click deploy does it) |
| LAN URL unreachable | Phone must be on the same Wi-Fi; re-run the launcher so the current LAN IP is detected |
npm.ps1 blocked by execution policy | Use npm.cmd, or Set-ExecutionPolicy -Scope CurrentUser RemoteSigned |
npm view 404s right after a publish | CDN edge cache โ wait a minute or query with Cache-Control: no-cache |
Plugin never appears after dsh plugin add | You must also append the loader row and restart dsh web |
| PC sleeps | keep_awake + power plan are handled by the deploy; see keep_awake.ps1 |
| Harness keeps dying / phone unreachable | Check %USERPROFILE%\.dsh\launcher\watchdog.log; re-run install.ps1 to (re)register the watchdog task |
Orb widget shows ? / "Disconnected" | The harness or the forwarder is down โ check the watchdog log and 127.0.0.1:3080 |
| Push toggle disabled ("needs HTTPS") | Open the harness via Tailscale HTTPS or http://localhost:3080; plain-LAN HTTP can't host a service worker โ see docs/presence-push.md |
| No phone push with the page closed | See docs/presence-push.md โ 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 โ sidebar collapse, floating ball, panel width, no overflow.
See
docs/device-tests/.
Roadmap
- Tailscale HTTPS + IP access, walk-on-LAN
- Persistent plugin (no per-session run)
- Device pairing + credential auth + revocation
- Capability-bounded allowlist + protected paths + path-escape tests
- Bilingual UI, security tests, CI
- Startup quarantine of corrupt demo sessions + self-healing watchdog
- Session size hints + "suggest archiving" in the phone UI
- Demo-presence behavior tests (idempotent add, fixed cwd, clean-only-demos)
- UTF-8 encoding guard on upload/edit + BOM/newline preservation
- /pocket strict mode (opt-in via
~/.dsh/remfs-options.json) - Tailscale ACL hardening guide
- Desktop shortcut automation (install.ps1)
- PC always-on-top Orb widget (scripts/orb-widget.ps1, zero-dependency WinForms)
- Phone Web Push notifications with the page closed (RFC 8291, opt-in per paired device)
- More device resolutions validation
- Upstream contributions (see below)
License
MIT