dsh-llm-ollama

If you run Ollama on a local machine or a self-hosted container and want to route those models through DeepSeek Harness (dsh), dsh-llm-ollama is the missing bridge. It skips every OpenAI-compatible shim and talks directly to Ollama's native /api/chat NDJSON streaming endpoint, so Ollama-specific parameters—num_ctx, num_predict, keep_alive, temperature—pass through verbatim: whatever you type in the settings page is exactly what lands in the request body. There is no mapping loss, no truncated option set, no silent parameter drift from a compatibility layer. Model discovery is one click: the plugin calls GET /api/tags, lists everything the server already has, and lets you check them into the model catalog in a single dialog. Local ollama serve or a remote container—the only difference is the API address you paste in; the configuration flow stays identical either way.

For reasoning models (Qwen3, DeepSeek-R1/v3.1, GPT-OSS, and others), the plugin auto-detects capability from the /api/tags response in the background without ever blocking the model picker, and surfaces a reasoning-level control only when it belongs. Boolean-think models get a simple On/Off toggle; models with graded levels can switch to a levels menu with a configurable default, and you can always override the auto-detection per model if the inference is off. The streaming payload's message.thinking trace is rendered as a readable reasoning block, tool calls and image attachments (multimodal models) ride the native channel, and the right-bottom context-usage indicator updates in sync.

Need a quick per-session context bump? Type /ollama-context to open a panel with 4096-to-128K presets or a fully custom value; the override applies only to the current session and reverts to the model's configured window when you start a new one. The plugin ships as a bundle patch that hooks itself into the profile composition tree automatically at install time, so there is no manual cordis.patch.yml editing to worry about.

This is the right plugin for solo developers running Ollama locally, teams that self-host LLMs in containers, and anyone who wants fine-grained control over context length, reasoning depth, and model residency inside dsh without fighting an OpenAI-shaped API.