dsh-thinking-language-zh

In multi-agent DeepSeek Harness deployments, a model's reasoning language often drifts between Chinese and English, leaving inconsistent thinking fragments in logs and user-facing replies. dsh-thinking-language-zh pins the agent's thinking and reply language with a single, stable system-prompt section so every agent in the deployment reasons and responds in the chosen language by default.

The plugin does exactly one thing: it injects a thinking:language instruction between the persona and the tool-guidance band. It contributes no tools, no shell access, and no model changes. Configuration is minimal—an editable text field plus an optional complete flag—and hot-reloading a config edit swaps the section instantly. The default Chinese instruction costs roughly 60 tokens per request and keeps prompt-prefix (KV) caching intact, making it safe to leave mounted in production.

It is a good fit for teams running Web UI, headless, or ACP DeepSeek Harness deployments who need a uniform reply language and want to eliminate reasoning drift. Because the text field is a template, the same plugin also serves as a starting point for English or any other language by overriding the row in a cordis.patch.yml layer.