DeepSeek Harness

Install and launch

Complete Install and connect, using tokamak-stag for staging. Have Node.js/npm available; Tokamak installs @deepseek-ai/dsh if missing. Select a model with Chat Completions and the tool capabilities you need.

tokamak launch deepseek-harness --model YOUR_MODEL_ID --check
tokamak launch deepseek-harness --model YOUR_MODEL_ID -- --port 3081

The launcher runs dsh web. Open the local URL printed by Harness and follow its local authentication flow. Extra web options belong after --:

tokamak launch deepseek-harness -- --help

Configuration and models

Tokamak starts from DSH_HOME/settings.yaml, normally ~/.dsh/settings.yaml, and creates a private temporary copy. A final settings patch selects the Tokamak provider, endpoint and model. The key is supplied through TOKAMAK_API_KEY; original sessions, profiles and credentials stay in their existing home.

Existing provider compatibility settings and model capabilities are preserved. Changes made to the temporary settings copy do not persist after the run. Additional --patch flags are rejected because they could replace the managed route. Select a different model with tokamak launch deepseek-harness --model.

Verify and troubleshoot

Try a small task and check Analytics or tokamak usage. Web UI startup alone is not evidence of successful inference. The recorded Windows check started Harness 0.1.2-rc.1 and loaded its loopback UI; inference was not exercised.

If settings cannot be parsed, inspect the existing YAML rather than deleting your profile. If a port is busy, choose a different --port. The developer-preview schema is version-sensitive; use web help for supported arguments.

Native tools, plugins and web search keep their configured services. They are not replaced by a Tokamak search API. The launcher does not configure a continuous Harness OTLP usage exporter; any native feedback-sharing settings remain under the client's control. Server inference records are authoritative. See troubleshooting.