integrations/deepseek-harness
tt-a1i/archify从仓库或系统描述生成经过校验的自包含交互式架构图、流程图、时序图、数据流图和生命周期图。
翻译 CPU:Windows x64 专属翻译界面,侧栏独立入口,支持工作台/聊天切换、原文与译文双栏、尽量保留原格式的文档翻译、屏幕框选翻译及最近记录,复用宿主模型、语音输入与预览。使用 CPU OCR,无需 NVIDIA 运行库。不要同时启用 CPU 与 CUDA 两个版本。
在 DSH 侧栏点击“翻译”,打开独立的专属翻译界面。工作台提供原文与译文双栏,支持切换到聊天模式;文本翻译、文档上传、译文下载和最近翻译记录集中在同一页面。
侧栏中的独立“翻译”入口:
专属翻译工作台,包含语言选择、工作台/聊天切换、屏幕翻译和原文/译文双栏:
截图展示 CUDA 版;CPU 版使用相同界面,运行库不同。
Open Translation from the DSH sidebar to enter a dedicated interface with Workbench / Chat modes, side-by-side source and translation, document uploads and downloads, recent translations, and screen-region translation. Uses the models configured in DSH, with host voice input and document preview. Document translation aims to preserve the original formatting.
Windows x64. DSH version numbers are unrestricted (engines.dsh = *); required host APIs are checked at startup. Future API changes may still require updates.
This small npm package automatically installs its pinned document runtime from GitHub Releases. The CUDA edition also installs its pinned GPU component. No postinstall scripts are required. CPU/CUDA share settings and translation history; enable only one edition.
CPU downloads about 364 MB in total; CUDA about 2.04 GB including CUDA/cuDNN. A short versioned runtime directory under LOCALAPPDATA avoids Windows DLL path limits, using hard links where possible.
Source, installation details and validation: https://github.com/piaomiaozhilei/dsh-sacc-translator
Version 0.1.19 defaults to 5 concurrent translation requests; saved preferences take precedence. Bundle overrides support PDF/general file reads up to 1 GiB, Excel up to 256 MiB/1,000,000 cells/120 seconds, and Office inputs up to 512 MiB with PDF output up to 1 GiB and a 300-second conversion timeout. Large previews still load full files and need sufficient memory. Profile overrides take precedence.
Version 0.2.2 includes native Rust screen translation. Enable it in the translator panel, press Ctrl+Shift+S to select an area, then Enter to translate. Selection persists while translating; the overlay uses a continuous Gaussian-blurred backdrop. Screen OCR uses CPU in both editions and releases models after 45 seconds idle. Full live-desktop/model visual verification remains incomplete.