dsh-llm-grok

If you hold an xAI Grok subscription and want to fold Grok models into a multi-model DeepSeek Harness workflow, dsh-llm-grok provides a dedicated grok provider route. Authentication is a Host-side PKCE flow: the browser opens only for the sign-in page and never receives tokens, the session file is stored under $DSH_HOME with 0600 permissions, and the entire chain stays on the Host. There is no user-editable base URL and no apiKeyEnv declaration, keeping the credential boundary clean and auditable.

On the conversation side, every request goes through Grok Responses with server-side web_search and x_search attached automatically, so grounding works out of the box without configuring a separate search provider. You can dial reasoning effort across low, medium, high, and xhigh (the last tier exclusive to grok-4.6), and each model entry in the catalog carries its own context-window budget used as the DSH compaction ceiling (500,000 tokens by default for the current generation). Signed-in cards also surface live subscription usage. Need image generation? Flip on grok_image_gen to reuse the same OAuth session and call Grok Imagine; output bytes land directly in the workspace attachment store. Paired with dsh-model-switch, Grok also registers an optional Image adapter for the stable generate_image router.

The plugin targets Harness users running multiple models, whether headless or Web, who want a single subscription-authenticated Grok route alongside their existing providers. It contributes only its own card, model catalog, credentials, and Host route; the LLM Providers page and shared order store belong to dsh-llm-providers-ui. Load order does not matter, rollback is a single fixed-version tgz swap, and headless routing works without the owner package at all.