dsh

dsh is a pure-config Cordis patch layer that solves one specific problem: routing DeepSeek Harness model inference through a locally running OpenLLM daemon, with no API key required, no adapter code needed, and dsh itself never aware of the upstream origin. It does exactly two things, repoints the built-in llm-pi-ai adapter to the loopback gateway at 127.0.0.1:8787/v1 and makes it the default model, and registers the OpenLLM MCP stdio server carrying three tool groups (openllm, claude-context, supermemory). The entire plugin has zero runtime logic; every behavior is declared in cordis.patch.yml.

The practical effect is that dsh becomes an ultra-thin, key-free proxy whose only prerequisite is that openllmd is already running on the host. The daemon reads its signed-plan credentials from ~/.openllm/.env and forwards upstream on its own, while dsh simply talks to 127.0.0.1. If you want your Harness browser UI or headless CLI to instantly gain OpenLLM model routing and MCP tool access without writing adapters or managing credential chains yourself, this patch is the shortest path. It is distributed from GitHub with the built lib/ already committed, so a git install loads immediately with no build step.

This is aimed at developers who have already deployed OpenLLM locally, treat dsh as their unified inference entry point, and want the OpenLLM MCP toolchain folded into their daily workflow. The bundle is pre-1.0; pin a specific release and re-test after each dsh upgrade.