dsh-local-radiance-server-controller

dsh-local-radiance-server-controller

A DeepSeek Harness (DSH) plugin: remote-control an AMD inference server on your local LAN over SSH — the local Radiance inference server controller.

Adds a "推理服务器 / Inference Server" panel to the DSH sidebar plus an always-visible live-metrics widget in the sidebar footer, so day-to-day operations need no terminal.

Chinese documentation: README.zh-CN.md

Sidebar entry and the always-live metrics widget (bottom of the sidebar):

Features

Card	What it does	Remote command
SSH settings	Host / port / username / password / compose directory / default service file / image name / power limit / sudo toggle	— (saved locally to ~/.dsh/dsh-local-radiance-server-controller.json)
AMD GPU power limit	Enable limit / reset to default; shows current limit state; buttons get a ✓ and disable when the state already matches	sudo amd-smi set -o <W> / sudo amd-smi reset -o
Radiance inference service	Dropdown of yaml files in the compose directory (common.yaml hidden), start / stop / update image (down first, then pull)	cd <dir> && sudo docker compose -f <file> up -d / down; sudo docker pull <image>
Server diagnostics	Fetch full amd-smi output	amd-smi
Danger zone	Shutdown / Reboot — each with its own two-click confirmation (auto-disarms after 4 s)	sudo shutdown -h now / sudo reboot
Sidebar widget	One persistent row: status dot + GPU % CPU % decode/prefill t/s (shrinks to a status dot when the sidebar is collapsed)	see Live metrics below

Opening the panel probes the server immediately (running services, power-limit state), then refreshes every 20 s; every action also triggers an immediate refresh.

Live metrics & connection architecture

The whole plugin keeps exactly one long-lived SSH connection to the server (15 s keepalive, self-healing reconnect with backoff). All traffic multiplexes over it through ssh2 exec channels:

Traffic	Cadence	Remote command
Engine decode/prefill	0.5 s	curl localhost:<PORT>/stats (radiance Web API; PORT and API key are read on the fly from <compose dir>/.env — the key never leaves the server)
CPU / GPU utilization	5 s	/proc/stat double-sample delta; GFX_ACTIVITY from amd-smi metric -u
Service / power state	20 s	docker compose ls --format json; amd-smi static -l (SOCKET vs MAX power to detect an active limit)
User actions	on demand	see Features table

No short-lived connections: the server's journalctl typically shows only one or two SSH logins per day.

The sudo password is piped in via base64 (base64 -d | sudo -S); plaintext never appears in remote process arguments.

Installation

Prerequisites

• This machine: DeepSeek Harness desktop
• Server: sshd with password login, sudo available (Radiance deployed via docker compose; AMD GPU with amd-smi; radiance engine exposing the /stats endpoint)

Steps (first install on any machine)

1. Copy the plugin directory anywhere on the target machine (do not copy node_modules; reinstall dependencies per machine): You only need index.js, client.js, package.json, cordis.patch.yml, icon.svg, locale/ (assets/ is the README screenshot, optional).

2. Install dependencies inside the plugin directory (installing a local directory does not install its dependencies; skipping this fails with Cannot find package 'node-ssh'):

   cd <plugin dir>
   npm install --omit=dev
   

3. Install the bundle into DSH: tell the Harness agent

   use plugin_manager to install <absolute plugin dir> as a bundle

   or run dsh plugin install <absolute plugin dir>.

4. Restart the DSH app (Host-side code loads at app start).

5. Open the sidebar panel, fill in the SSH details, save.

To sync between machines, copy the files listed in step 1 and repeat steps 2–3.

Configuration

Field	Notes	Default
Host / port / username / password	SSH login (the password is also used for sudo)	22
Compose directory	Directory containing the service yaml files (~ allowed)	~/radiance-compose
Default compose file	Preselected service file in the panel	flashnext-tp2.yaml
Radiance image	Image pulled by the Update-image action	stilldeadcode/radiance
Power limit (W)	Target watts for the Enable-limit action	210
Use sudo	When off, commands run without sudo (the remote user needs direct docker/AMD permissions)	on

⚠️ Security: the SSH password is stored in plaintext on this machine at ~/.dsh/dsh-local-radiance-server-controller.json. Don't use this on shared/managed machines, and keep the account's permissions minimal.

Troubleshooting

Symptom	Fix
Errors or stale behavior after code edits	Restart the DSH app (Host module cache)
sudo: a password is required	Re-save the password in the panel (sudo uses the login password)
Widget shows - although the service is fine	Normal during the reconnect cooldown after a server reboot/disconnect; recovers within ~30 s
A yaml file missing from the dropdown	Only plain .yaml/.yml file names without .. are listed; common.yaml is hidden on purpose
Decode reads 0/tiny while idle	Radiance reports a windowed engine-side average; idle is near zero by design

File structure

dsh-local-radiance-server-controller/
├── index.js          # Host half: HTTP endpoints + SSH connection pool + all remote actions
├── client.js         # Client half: panel UI + sidebar icon + live metrics widget
├── cordis.patch.yml  # Loader patch (injects the plugin entry)
├── icon.svg          # App icon
├── assets/           # README screenshots (screenshot.png, sidebar.png)
├── locale/{zh,en}.json
└── package.json      # dsh.bundle / dsh.client manifest; depends on node-ssh

Host endpoints (prefix route on the local DSH web server, protected by DSH login): GET state, GET job?id, GET fast, POST settings, POST action {action,file?} under /dsh-local-radiance-server-controller.

License

MIT — see LICENSE.