DeepSeek Harness: running dsh in a container
dsh is the launcher for DeepSeek Harness. How its profiles and configuration are laid out, which volumes to mount, how upgrades and rollbacks work, and what is out of scope.
What dsh is
DeepSeek Harness ships a command called dsh. It is not a single chat window; it is a launcher. A profile is an ordered stack of plugin bundles plus your own patch layer. dsh web boots the Web profile in a browser; the headless, sdk and acp profiles are for scripts and automation.
Profiles, plugins and patches live under $DSH_HOME, by default under profiles/<name>/. The profile package.json records the bundles and dependencies, cordis.patch.yml is your override layer, and there is a home-level patch on top of that. Plugins are managed with dsh plugin --profile <name> <pnpm args>, which forwards to pnpm.
Running it by hand is not hard. Keeping it running is the work: it should come back after a reboot, upgrades should not break plugins, and there should be a way back.
Why a container
- Dependencies are locked in the image, so the next machine does not reinstall Node and pnpm.
$DSH_HOMEand the workspace are volumes; rebuilding the container does not lose configuration or sessions.- A pinned image tag, plus an older tag to roll back to.
- CPU, memory and restart policies, so one job cannot take the host down.
What gets delivered
- Image or Compose: pinned Node and dsh versions, running as non-root.
- Volumes:
$DSH_HOMEand a separate workspace, both included in backups. - Ports and entry: the web port on the internal network only, behind Traefik or cloudflared, never straight onto the internet.
- Secrets: model API keys and other credentials injected as environment or runtime secrets; never in the image or in git.
- Plugins: installed as needed, with version ranges and disabled entries recorded;
engines.dshis checked before upgrading. - Health check, restart policy and log rotation.
- Upgrade steps: back up
$DSH_HOME, change the image tag, verify plugins and the web UI, roll back otherwise. - A verification list: the web UI opens, plugins load, sessions survive, restart recovers.
Not included
- Downloading model weights, tuning inference performance or GPU drivers.
- Modifying dsh upstream code; if that is needed, it is scoped separately.
- Applying for model API quota or accounts.
- 24/7 on-call. After handover you can restart and upgrade it yourself.
Security notes
The web UI is not exposed to the internet by default. For remote access, put Access or a proxy authentication layer in front and use HTTPS. Run the container as non-root, mount only what is needed (never the Docker socket casually), and inject secrets from the environment.
Who it fits
You already run dsh web or headless locally and want it available long term, portable, and upgradeable with a way back. Or a small team wants one Harness with pinned versions and backups.
How to start
Email how you run dsh now, which profile, the model backend, which machine it goes on, and whether remote access is needed. Scope first, then a quote.
Questions
Can I skip the container?
Yes. Installing on the host is simpler, but upgrades, rollbacks and dependency conflicts are then yours. A container buys repeatability.
Where does the data live?
Profiles, configuration and sessions under $DSH_HOME; working files in the workspace. Back up both, and a new machine only needs them mounted.
Can it use a local model?
Yes, depending on the provider configuration on the dsh side. Delivered and verified against your backend.
Will an upgrade lose anything?
Back up $DSH_HOME, change the tag, and roll back if a plugin is incompatible. Delivery includes one verified upgrade and rollback.
Related
Not sure where to start?
Run the assessment first, or send your setup and goal by email. The reply says whether it fits and what happens next.