Aquesta és la versió actual. · Data de publicació:
Fitxers d’instal·lació 0.4.5
| Arquitectura | Tipus | Àmbit | Mida del fitxer | Baixa | SHA256 |
|---|---|---|---|---|---|
| 64 bits | EXE (NSIS) | Usuari actual | — | github.com | VirusTotal |
L’enllaç i l’SHA256 provenen del manifest comunitari del Gestor de paquets del Windows. Convé comparar el hash del fitxer baixat abans d’executar-lo.
Notes de la versió
- Improved macOS selection compatibility with screenshot overlays. Clipboard fallback no longer injects copy shortcuts when an overlay covers the selection gesture. Moved host accessibility queries off the main thread to keep result-window updates responsive.
- Improved streaming results. The first text chunk appears immediately, later updates are batched per frame, and hidden-window timers ensure the stream continues to flush. Refined scrolling, waiting, copying, and Markdown presentation.
- Added model-aware thinking controls for each tool. Detection uses model, protocol, endpoint, and capability metadata for OpenAI Chat/Responses, Anthropic, Gemini, OpenRouter, and common Qwen, GLM, Kimi, and DeepSeek models. Added the max level; when a model cannot disable thinking, selecting off uses its lowest supported level.
- Moved thinking settings to each tool’s model selection, with controls to enable or disable thinking, select a level, and rerun automatic detection. Different tools can use different levels for the same model.
- Improved the default translation and explanation prompts. Translation better preserves meaning, tone, protected content, and document structure. Explanations can use necessary field terminology while remaining accessible. Only untouched built-in prompts are upgraded.
- Improved language detection with shared TypeScript and Rust rules. Code, URLs, commands, and logs are excluded from language detection without changing the text sent to the model. Fixed inline-code lines being mistaken for fenced code.
- Improved PDF text handling by merging semantically continuous wrapped lines into paragraphs while preserving paragraph, list, table, and code boundaries.
发布版本 0.4.5
- 提升 macOS 划词与截图遮罩的兼容性。截图浮层遮住手势位置时，不再向宿主注入复制快捷键；将宿主辅助功能查询移出主线程，保持结果窗口响应流畅。
- 改进流式结果显示：首段文本立即出现，后续增量按帧合并，并通过隐藏窗口定时器确保内容持续刷新；优化滚动、等待、复制和 Markdown 呈现。
- 为各工具加入模型感知的思考控制。结合模型、协议、端点和能力元数据识别 OpenAI Chat/Responses、Anthropic、Gemini、OpenRouter 及常见 Qwen、GLM、Kimi、DeepSeek 模型。增加 max 档位；模型不支持关闭思考时，选择关闭会使用其最低支持档位。
- 将思考设置移至各工具的模型选择处，支持启用或关闭思考、选择档位和重新自动识别。同一模型可按工具分别设置不同档位。
- 优化默认翻译与解释提示词。翻译更注重保留语义、语气、受保护内容和文档结构；解释可使用必要的学科术语，同时保持易懂。仅更新未被用户编辑的内置提示词。
- 提高语言识别准确度，TypeScript 与 Rust 共用识别规则。识别时过滤代码、URL、命令和日志，但不改动发送给模型的原文；修复行内代码被误判为代码围栏的问题。
- 改进 PDF 折行文本处理，自动合并语义连续的文本行并恢复段落，同时保留段落、列表、表格和代码边界。
