dsh-llm-motomoto

DeepSeek Harness ships its model selector with the chat/completions path only, yet the MotoMoto relay speaks the OpenAI Responses API and demends a Codex CLI client identity to stream correctly. Hand-tweaking routes and request headers is brittle and hard to maintain. dsh-llm-motomoto packages all three concerns into a single plugin: it injects a default route, fences globalThis.fetch at runtime to rewrite identity headers and strip rejected body fields, and surfaces a live status card under Settings – Plugins so you can see the active group, endpoint, and credential reference at a glance.

The request-layer fence is the heart of the bundle. It rewrites the User-Agent to the Codex CLI default, appends the originator: codex_cli_rs header to match the official client, and removes body fields the upstream rejects by default (max_output_tokens included). Every other host passes through untouched, and unloading the plugin restores the original fetch. Route configuration uses a mergeLayers scheme: the plugin supplies the base patch and per-key overrides in your settings.yaml win field by field without a full rewrite.

Built for developers running the DeepSeek Harness desktop app who want zero-config access to the MotoMoto relay (default model gpt-5.6-terra, 262K context window), or anyone looking to extend their available model sources without standing up a global network proxy.