dsh-plugins

Connecting dozens of third-party text models to DeepSeek Harness used to mean copy-pasting API keys, hand-writing provider blocks, and figuring out on your own whether a model should go through Anthropic or OpenAI. The dsh-plugins package replaces all of that with a single browser sign-in: OAuth tokens refresh automatically, no secret ever touches your config file, and a handful of slash commands let you check status, refresh the catalog, or log out.

The current bundled catalog carries 150 searchable text models and routes each one to the right protocol at request time: 126 of them go through the Anthropic Messages channel with prompt-cache controls enabled, so the first eligible call can create a cache and later calls sharing the same prefix can report cache reads; the remaining 24 OpenAI-only models are routed to Chat Completions. Every inference request also carries the active DSH session ID in the x-zenmux-session-id header for upstream tracing and rate-limiting. The catalog is a release snapshot generated from ZenMux official model-list endpoints, and a single refresh command pulls the latest listing while preserving existing OAuth credentials and other provider settings.

Built for heavy Agent operators who need to schedule both Claude and GPT-class models inside a DSH workflow, care about prompt-cache hit rates, and would rather not juggle multiple API keys in configuration files.