dsh-think-translate

DeepSeek-class models often reason in Chinese, which means the expanded Think row can be an unreadable wall of text for English-speaking users trying to debug or review responses. dsh-think-translate renders the thinking chain, task cards, and final answer in real time in one of 8 target languages, effectively adding live subtitles to the model's internal monologue while keeping the original transcript completely intact.

Out of the box the plugin defaults to a local Ollama model such as qwen2.5:7b or 14b, so translation is free, unlimited, and never leaves your machine. Selecting the local provider for the first time triggers an automatic download with a live progress bar, and the model is enabled the moment it finishes. Because it operates purely at the display layer, the model context always contains the original text; no translated token ever occupies the context window. Google and Bing are available as automatic fallbacks if the local model is unavailable.

Quality and resilience are first-class concerns. Long thinking chains are split on paragraph boundaries and further batched by sentence, so even a small local model produces coherent output. File paths, shell commands, URLs, regular expressions, and pure code lines are detected and never translated. The runtime features 3x exponential-backoff retries, a browser-direct fallback path, and a guarantee that failed results are never cached. You can pre-translate everything, lazy-load historical chains, or translate only on expansion.

The plugin is best suited for developers who need to follow reasoning in a non-native language, multilingual engineering teams, and privacy-conscious users who want fully offline translation without shipping data to a cloud API.