One-Key Input v0.3.0

中文 / English

本版两件大事：本地离线 ASR（音频不再出网）和智能热键模式（短按 toggle / 长按 push-to-talk）。建议所有 v0.2.x 用户升级。

Two headline features: fully offline ASR (audio never leaves the machine) and smart hotkey behavior (tap to toggle, hold to push-to-talk). Recommended for all v0.2.x users.

---

🇨🇳 中文

✨ 新功能

• 本地离线 ASR：sherpa-onnx + Paraformer-zh 流式识别（#12 / #14）。新增 sherpa-paraformer provider，语音转文字完全本地跑，音频不再发送到任何云端（只有润色这一步还走 LLM）。
  • 引擎：sherpa-onnx v1.13.2，CMake FetchContent 拉取并校验 SHA256
  • 模型路径：默认 models\paraformer-zh-streaming（exe 同目录），支持相对路径 + %LOCALAPPDATA% 等环境变量展开
  • Settings UI：ASR provider 下拉新增「Sherpa Paraformer (local)」，含模型目录 Browse + 线程数 + Test 按钮（不加载模型直接探测目录结构）
  • 模型下载：README 给了 ModelScope（国内）和 HuggingFace 两个镜像
  • 体积：portable bundle ~55 MB（不含模型），中英双语 int8 量化模型再 +~230 MB
• 智能热键行为（#15）。新增 hotkey.behavior 配置，三档可选：
  • push_to_talk（默认，按住说话，老行为）
  • toggle（按一下开始，再按一下结束）
  • smart（短按 = toggle，长按超过 smart_threshold_ms（默认 400ms）= push-to-talk）—— 新装用户默认这档
  • 配套：Esc 强制停止 / max_duration_ms（默认 5 分钟）看门狗自动结束，防止 toggle/sticky 状态忘了关
  • 托盘图标真的会变色了：录音时变红（之前 Redraw() 是空实现）

🐛 修复 & 改进

• 路径解析以 exe 目录为锚（C++ 引擎 + Rust Settings probe 一致），自启动 / 快捷方式 / Win+R 启动都能正确找到模型
• v0.1 / v0.2 的旧 config 升级到 v0.3 自动 backfill 新字段（behavior 默认 push_to_talk，行为不变）
• 未知 behavior 字符串安全降级到 push_to_talk 并打 warning

📦 使用

下方下载 OneKeyInput-0.3.0-portable.zip（~15 MB），解压后跑 onekey-settings.exe。

• 想保持云端 ASR：什么都不用改
• 想试本地 ASR：Settings → ASR Provider 选 Sherpa Paraformer (local) → 下载模型放到 models\paraformer-zh-streaming\ → Test 按钮确认 → 保存
• 想试智能热键：Settings → Hotkey → Behavior 选 Smart

🚧 不在本版（明确范围）

• 模型自动下载器
• GPU 加速（CUDA / DirectML）
• 本地 LLM 润色
• VAD 自动停止 / 流式转写浮窗 / 自定义结束热键

---

🇬🇧 English

✨ New Features

• Fully offline ASR: sherpa-onnx + streaming Paraformer-zh (#12 / #14). New sherpa-paraformer provider runs speech-to-text entirely on-device — audio never leaves the machine; only the polish LLM still goes to the cloud.
  • Engine: sherpa-onnx v1.13.2, fetched via CMake FetchContent with pinned SHA256
  • Model layout: default models\paraformer-zh-streaming next to the exe; supports relative paths + %LOCALAPPDATA%-style env-var expansion
  • Settings UI: new Sherpa Paraformer (local) option with Model-directory Browse, thread count, and a Test button that probes directory shape without loading the model
  • Models: README links both ModelScope (China mirror) and HuggingFace
  • Footprint: portable bundle ~55 MB without the model; the bilingual zh-en int8 model adds ~230 MB
• Smart hotkey behavior (#15). New hotkey.behavior config with three modes:
  • push_to_talk (default for upgrades — legacy behavior preserved)
  • toggle (tap to start, tap to stop)
  • smart (short tap = toggle, hold past smart_threshold_ms (400ms) = push-to-talk) — default for fresh installs
  • Safety: Esc force-stop and a max_duration_ms watchdog (default 5 min) so toggle/sticky sessions can't run away
  • Tray icon actually changes color now: red while recording (previous Redraw() was a no-op)

🐛 Fixes & Improvements

• Relative model paths now anchor on the exe directory (consistent across C++ engine and Rust Settings probe) — autostart, Start-menu shortcut, and Win+R launches all resolve correctly
• v0.1 / v0.2 configs auto-backfill the new hotkey fields on upgrade (behavior defaults to push_to_talk, no behavior change)
• Unknown behavior strings safely fall back to push_to_talk with a warning

📦 Usage

Grab OneKeyInput-0.3.0-portable.zip below, unzip, run onekey-settings.exe.

• Staying on cloud ASR: nothing to change
• Trying local ASR: Settings → ASR Provider → Sherpa Paraformer (local) → drop the model into models\paraformer-zh-streaming\ → hit Test → Save
• Trying smart hotkey: Settings → Hotkey → Behavior → Smart

🚧 Not in this release (explicit scope)

• Auto model downloader
• GPU acceleration (CUDA / DirectML)
• Local-LLM polish
• VAD auto-stop, streaming-transcript overlay, custom end-hotkey

---

Full Changelog: v0.2.1...v0.3.0

One-Key Input v0.2.1

中文 / English

小版本修复 + GPT-5.x 兼容。建议所有 v0.2.0 用户升级。

Minor fixes + GPT-5.x compatibility. Recommended for all v0.2.0 users.

---

🇨🇳 中文

🐛 修复

• 修复从 v0.1 升级到 v0.2 后 Settings 窗口空白（#6）。原因是新增的 translate 字段没有被 backfill，旧 config 加载后 React 渲染崩溃。现在 load_config 后会用 defaultConfig 补齐所有缺失的 section。
• 修复润色把口述当指令直接回答的问题（#10）。比如说"how do I deploy this"，模型之前会真去回答"你应该这样部署..."，而不是润色成"How do I deploy this?"。现在用 <transcript>...</transcript> 包装用户输入 + system prompt 明确边界，模型不再被绕进去执行口述内容。F8 翻译模式同样修复。

✨ 兼容性

• 支持 GPT-5.x reasoning 系列模型（gpt-5-mini / gpt-5 等）。这类模型在 Azure OpenAI 上禁止自定义 temperature、强制 max_completion_tokens，之前用 gpt-5-mini deployment 会 400 错误。现在：
  • 移除 polish.temperature 配置项（润色任务对温度不敏感，靠 server 默认值更省心）
  • 请求体改用 max_completion_tokens（对老模型 gpt-4o-mini 也兼容）
• 推荐切换到 GPT-5-mini：GPT-4o-mini 在 Azure 上 2026-02-27 退役，建议提前换

📦 使用

下方下载 OneKeyInput-0.2.1-portable.zip，解压后跑 onekey-settings.exe 配 Azure key。如果你之前的 deployment 是 gpt-4o-mini 一切照旧；如果想试 gpt-5-mini，直接改 deployment name 即可。

---

🇬🇧 English

🐛 Fixes

• Fix Settings white-screen on v0.1 → v0.2 upgrade (#6). Root cause: the new translate config section wasn't backfilled, so React crashed when reading cfg.translate.enabled from old configs. load_config now merges all missing sections against defaultConfig().
• Polish/translate no longer answers question-shaped dictations (#10). Saying "how do I deploy this" used to make the LLM actually answer the question instead of polishing it into "How do I deploy this?". User input is now wrapped in <transcript>...</transcript> and the system prompt makes the boundary explicit. F8 translate mode fixed the same way.

✨ Compatibility

• GPT-5.x reasoning models now supported (gpt-5-mini, gpt-5, etc.). These models on Azure OpenAI reject custom temperature and require max_completion_tokens instead of max_tokens — previously every polish call to a gpt-5-mini deployment 400'd. Changes:
  • Removed polish.temperature config field (polish quality isn't temperature-sensitive enough to justify the footgun)
  • Request body and connection-test probe now use max_completion_tokens (also accepted by gpt-4o-mini, no regression)
• Recommended: migrate to gpt-5-mini. Azure is retiring gpt-4o-mini on 2026-02-27.

📦 Download

Grab OneKeyInput-0.2.1-portable.zip below. Existing gpt-4o-mini setups keep working; to try gpt-5-mini just change the deployment name in Settings.

---

Full Changelog: v0.2.0...v0.2.1

One-Key Input v0.2.0

中文 / English

---

🇨🇳 中文

本次发布带来 翻译模式（Translation Mode），与原有的润色（口述）流程并列。

✨ 新功能

• F8 翻译模式：按住 F8（可在 Settings 里改）口述一种语言，自动翻译后注入到光标位置。复用 F9 润色相同的 Azure Speech → OpenAI 管道。
• TranslationPrompt 模块：结构化的翻译提示词，规避 Azure 内容过滤的越狱检测。
• Settings UI 翻译区：在 Tauri 设置面板新增翻译相关配置——启用开关、热键、源/目标语言、提示词覆盖。
• 润色提示词 显式要求英文技术词保持原样，告别 API、JSON 等被乱译的情况。

🐛 修复

• 修复 kPromptTidy 中的乱码字符（U+FFFD 替换字符）。
• 修复翻译提示词触发 Azure 内容过滤越狱检测的问题。
• 翻译配置回填复用 defaultConfig()（来自 PR #5 review）。

🛠 工程

• 新增 scripts/stage-portable.ps1，本地复现 CI 的 portable zip 打包流程，方便 dogfood 不用推 tag。
• 测试栈已就位（#1）：GoogleTest（core） + Vitest（settings UI） + cargo test（Tauri 后端）。

📦 使用

下方下载 OneKeyInput-0.2.0-portable.zip，解压到任意目录：

1. 先跑 onekey-settings.exe 配 Azure key、热键、目标语言
2. 再跑 onekey-core.exe 常驻
3. 首次启动 SmartScreen 会拦截（未签名），点「仍要运行」即可

---

🇬🇧 English

This release introduces Translation Mode alongside the existing polish (dictation) pipeline.

✨ New features

• F8 Translation Mode — press-and-hold F8 (configurable) to dictate in one language and have the result translated and injected at the cursor. Reuses the same Azure Speech → OpenAI pipeline as F9 polish.
• TranslationPrompt module with a structured translator prompt designed to survive Azure content-filter jailbreak detection.
• Settings UI — new Translation section in the Tauri settings app: enable/disable, hotkey, source & target language, prompt overrides.
• Polish prompt now explicitly keeps English technical terms verbatim (no more mangled API, JSON, etc.).

🐛 Fixes

• Fixed mojibake (U+FFFD replacement char) in kPromptTidy.
• Translate prompt no longer trips Azure content-filter jailbreak heuristics.
• Polish reuses defaultConfig() for translate backfill (review feedback from #5).

🛠 Tooling

• scripts/stage-portable.ps1 — local equivalent of the CI portable-zip staging step, useful for dogfooding without pushing a tag.
• Test stacks landed in #1: GoogleTest (core), Vitest (settings UI), cargo test (Tauri backend).

📦 Download

Grab OneKeyInput-0.2.0-portable.zip below, unzip anywhere:

1. Run onekey-settings.exe first to configure Azure keys, hotkeys, target language
2. Then run onekey-core.exe to keep it resident
3. SmartScreen will block on first launch (binary is unsigned) — click "Run anyway"

---

Full Changelog: https://github.com/zwcih/one-key-input/commits/v0.2.0

One-Key Input v0.1.0

One-Key Input v0.1.0

第一个公开版本 / First public release.

这是什么 / What it is

按住 F9 → 说话 → 松开 → 上下文感知的 AI 润色文字 直接落到光标处。
Hold F9 → speak → release → context-aware AI-polished text typed at your cursor.

不切窗口、不复制粘贴、不打断输入流。
No window switching, no copy-paste, no interrupting your flow.

亮点 / Highlights

• 上下文感知润色 / Context-aware polish
  按下热键时，Core 用 Windows UIAutomation 抓取当前前台窗口的 — 应用名、窗口标题、焦点控件、光标前后字符、同窗口其它可见文本区域 — 附加到 LLM 的 system prompt。同样一句"挺有意思的"，在 YouTube 评论框、Slack 回复、代码注释里会被润成完全不同、贴合场景的句子。
  When you press the hotkey, Core takes a UIAutomation snapshot of the foreground window — app, title, focused control, text around the caret, sibling text regions — and appends it to the LLM's system prompt. The same "interesting take" gets rewritten differently in a YouTube comment box, a Slack reply, or a code comment.

• 零延迟设计 / Zero-latency design
  上下文抓取和录音并行（800ms 预算），不增加总时长。config.json 里 polish.use_context: false 完全关闭。
  Context capture runs in parallel with recording (800ms budget) — no added latency. Set polish.use_context: false in config.json to disable entirely.

• 后端可插拔 / Pluggable backends
  ASR / LLM 都是接口化的。当前默认 Azure Speech + Azure OpenAI，OpenAI 也支持。本地 Whisper / 本地 LLM 接口已留好。
  ASR and LLM are interfaces. Defaults: Azure Speech + Azure OpenAI; OpenAI works too. Local Whisper / local LLM hooks are stubbed.

• 润色失败安全网 / Polish-failure safety net
  LLM 调用失败时自动回落到注入原始识别结果——你不会丢字。
  If polish fails, the raw transcript is injected — you never lose an utterance.

• 完全 BYO Keys / Bring your own keys
  没有云服务、没有遥测。所有调用直接发到你自己的 Azure / OpenAI 账号。
  No server, no telemetry. All calls go directly to your own Azure / OpenAI account.

安装 / Install

下载下面的 OneKeyInput-0.1.0-portable.zip，解压到任意目录，双击 onekey-core.exe。第一次运行会弹出设置窗口让你填 key。完整说明见 README。

Download OneKeyInput-0.1.0-portable.zip below, extract anywhere, double-click onekey-core.exe. The settings window opens automatically on first run. See the README for details.

你需要 / You need

• Windows 10/11 (x64)
• Azure Speech 资源 (免费档每月 5 小时识别足够日常用 / free tier covers normal use)
• 一个 LLM endpoint：Azure OpenAI deployment（推荐 gpt-4o-mini）或 OpenAI API key

已知局限 / Known limits

• UIA 对原生 Win32 / WinUI / WPF 应用读取最佳；Electron / 浏览器内嵌内容深度依赖该应用是否暴露完整 a11y 树。专门的浏览器/IM 适配器在 v0.2 路线图。
  UIA reads native Win32 / WinUI / WPF apps best; Electron and browser content depend on a11y tree exposure. Dedicated browser/IM adapters are on the v0.2 roadmap.
• 仅 Windows。Linux/macOS 抽象层已铺设但未实现。
  Windows only. Linux/macOS abstraction layers are stubbed.

反馈 / Feedback

• 🐛 Issues
• 💡 Discussions
• ⭐ Star if it's useful — that's how I see whether to keep building.