dsh-llm-grok

If you already have a Grok subscription and want to plug it into DeepSeek Harness as a first-class LLM provider without extra infrastructure, dsh-llm-grok is built for exactly that. It routes your Grok CLI session token through the DSH credential system, talks to the subscription endpoint over a local HTTP proxy, and avoids both the Python sidecar and the fragile local auth.json file, removing a common source of 401 errors and format conflicts.

Under the hood the plugin covers grok-4.6 and grok-4.5, both accepting text plus image in a single native multimodal pass: images you paste into the chat or returned in a tool-result are encoded as image_url data URLs and shipped in the same /chat/completions call to the subscription endpoint — there is no OCR step and no separate vision tool. Reasoning effort ranges from low to xhigh, with a 500K context window and 128K output ceiling that comfortably hold long documents and multi-turn conversations. Session logs store only a sha256 reference to each attachment; raw pixels are read from ctx.attachments at request time and never persisted as redundant bitmaps.

The plugin suits developers who have already purchased a Grok subscription and prefer subscription-based billing over per-token API charges, as well as anyone who needs mixed text-image input without spinning up a separate vision pipeline. As long as your workflow runs on DSH 0.1.2-rc.1 or later and the session token is synced into the DSH credential space, Grok and DeepSeek can share the same canvas.