dsh-think-translate

DeepSeek-class models tend to reason in Chinese, so for non-Chinese-speaking users the expanded Think row can be a wall of foreign text that slows down debugging and review. dsh-think-translate renders the thinking chain, task cards, and final answer in your chosen language in real time—like live subtitles for the model's inner monologue—across 8 target languages including English, Japanese, Korean, Spanish, French, German, and Russian.

By default it calls a local Ollama model (qwen2.5:7b, 14b, or any custom endpoint), keeping everything free, unlimited, and entirely offline. On first selection the model is auto-downloaded with a live progress bar and enabled once ready. Because the plugin works purely at the display layer, the original text always stays in the transcript and model context; translated output never consumes a single token of context window. A Google or Bing fallback kicks in automatically when the local model is unavailable.

Translation quality and resilience are carefully handled: long chains are chunked on paragraph boundaries and further batched by sentence so even a 7B local model maintains coherence; file paths, shell commands, URLs, regex patterns, and pure-code lines are detected and left untouched. Runtime includes 3× exponential-backoff retries, a browser-direct fallback, and failed results are never cached. Users can choose to pre-translate everything, lazy-load historical chains, or translate only when a chain is expanded.

Ideal for developers who need to follow non-native-language reasoning, multilingual teams, and anyone who prefers fully local, offline translation without sending data to a cloud API.