DSH-Think-zh

DSH-Think-zh addresses a small but frustrating issue for Chinese users of DeepSeek Harness: model reasoning often stays in English, making the thinking process harder to read and verify. By injecting a concise language directive into every request's system prompt, it encourages the model to reason in Simplified Chinese, reply in the language of the prompt, and keep code, identifiers, file paths, and commands unchanged.

Its core strengths are simplicity and low overhead. The plugin uses a single injection mechanism, avoids extra detection, buffering, warnings, or write-back behavior, and keeps context pollution and token cost to a minimum. It also registers the directive after the system prompt service is ready, reducing the risk of silent degradation. Users can disable injection or customize the directive when needed.

This is best suited for developers, document writers, and DSH users who want Chinese-readable reasoning while preserving natural bilingual responses. It does not try to lock the model completely; instead, it offers a lightweight way to make reasoning more inspectable, replies more language-aware, and technical content safer to read.