dsh-openai-responses-bridge

DeepSeek Harness builds its provider system around native protocols, so plugging in Azure OpenAI, a third-party Responses gateway, or Gemini typically means patching the host source tree or maintaining a separate shim. dsh-openai-responses-bridge turns that layer into a standalone plugin that speaks entirely through DSH plugin interfaces and never modifies the source tree.

The bridge is a thin compatibility layer on top of the native Pi protocol. On the OpenAI Responses route it strips fields that some gateways reject, injects a hosted web_search tool on demand, and renders a search card in the conversation showing status, queries, citations, sources, and errors. On the Gemini route it delegates straight to the native generateContent adapter, preserving the request body and stream behavior. In Code Mode, nested web_search dispatches are transparently forwarded as a minimal hosted request that returns the native DSH web_search value, keeping the existing SDK contract intact without exposing a second provider-specific tool. The plugin also performs one-shot model discovery via an OpenAI-compatible /models endpoint, stores API keys through DSH credentials so the settings file never contains secrets, and supports hot-reloading provider configuration.

If you already develop daily inside DSH and need to fold company gateways, Azure OpenAI, or Gemini into the same workflow-specially when you want a single provider setting shared across Native and Code Mode, with hosted search cards appearing naturally in the conversation-this plugin was built for exactly that.