dsh-think-translate

When working with DeepSeek-class models, the reasoning chain is often emitted in Chinese or whatever language the model happened to think in, leaving non-Chinese users unable to follow the model internal deliberation. dsh-think-translate is a client-side plugin for the DeepSeek Harness Web UI that translates thinking chains, task cards, and final answers into your chosen language in real time, purely at the display layer. Originals remain untouched in the transcript and the model context window; translated text never consumes a single token.

It supports 8 target languages and a single setting that applies globally and persists across reloads. The translation provider chain is drag-and-droppable: built-ins include Google GTX, Bing, and local Ollama; you can add any OpenAI-compatible or Anthropic Messages API endpoint; and providers already configured in DSH settings.yaml are auto-discovered with one click, their keys resolved from credentials at request time and never written to plugin config. Paragraph- and sentence-aware chunking keeps quality high even with small local models, and a 3x backoff retry policy ensures resilience. File paths, shell commands, URLs, regex patterns, and pure code lines are never translated.

Built for developers and researchers who want to read a model reasoning in their own language, teams deploying DeepSeek Harness in non-Chinese locales, and privacy-conscious users who prefer fully offline translation. Local Ollama auto-downloads and enables the model on first selection, so nothing leaves the machine.