Repository navigation
Releases
v0.9.0 · 后台语音想法与鼠标侧键
简体中文
- 修复保存快捷键时的 Content 属性报错。
- 支持鼠标侧键 1 / 2 和带修饰键的侧键组合,按住录音、松开识别。
- 录音与本地识别全程在后台进行,自动保存到「我的想法」,不弹窗、不抢焦点。
- 在「我的想法」中稍后修改文字,或加入待办;旧版待办与设置保留。
- 设置改为左侧分类导航:分析服务、语音记录、网页采集、隐私与数据。
仅发布中国大陆预览版。首次语音使用仍需下载 SenseVoiceSmall 模型,随后离线识别;安装包无需另装 Node.js 或 Python。公共作者 API 服务仍未开放。应用代码与说明由 AI 生成,第三方模型与工具属于其作者。
验证:后台录音流程(假麦克风)、真实公开音频的识别与自动保存、无窗口/焦点切换、原文编辑与持久化、鼠标按下/松开与键盘冲突、设置导航、待办回归、模型下载完整性与增量分析测试。真实鼠标硬件仍欢迎试用反馈。
English
- Fix the Content property error when saving a shortcut.
- Support mouse side buttons 1 / 2, optionally with modifiers; hold to record and release to recognize.
- Record and transcribe locally in the background, then automatically save to My Ideas without opening a window or stealing focus.
- Edit ideas later or add them to tasks. Existing tasks and preferences are preserved.
- Settings now uses a left sidebar for Analysis, Voice capture, Web collection, and Privacy.
Mainland China preview only. SenseVoiceSmall downloads on first voice use and then runs offline. Node.js is included; no Python installation is required. The public author-paid API service remains unavailable. Application code and documentation are AI-generated; third-party models and tools belong to their authors.
Validated with fake-microphone capture, real public-audio recognition and automatic persistence, window/focus checks, editing, mouse press/release and keyboard conflicts, sidebar navigation, task regression, model integrity, and incremental analysis tests. Physical mouse testing feedback is welcome.
v0.8.0 · 本地语音记录 / Local voice capture
v0.8.0 · 本地语音记录 / Local voice capture
按住 F2 说出想法,松开后使用本机 SenseVoiceSmall 识别,确认或修改文字后保存为待办。支持普通话、粤语、英语及自动识别;语音设置允许更换快捷键。麦克风只在主动按键时录音。首次使用自动下载约 160 MB 模型,之后可离线使用,不产生转写 API 费用。下载需要访问 GitHub。
语音待办独立保存,扫描对话不会覆盖。录音只临时用于识别,取消不保存任务。每次录音最长 2 分钟。模型、录音、确认和持久化通过独立接口连接,后续可替换语音模型;请保存前核对识别文字。
下载 ConversationTodo-Mainland-Setup-0.8.0.exe 安装或升级;个人待办和设置保留。模型 gzip 是应用自动下载的资源,用户无需手动解压,不是第二个安装程序。当前仍为中国大陆预览版,软件提供的公共分析 API 尚未开放。
Hold F2 to record an idea, release to transcribe locally with SenseVoiceSmall, then edit and confirm before saving a task. Supports Mandarin, Cantonese, English and automatic detection; the shortcut is configurable. First use downloads a ~160 MB model from GitHub. Later recognition works offline without transcription API charges. Audio is transient; captured tasks survive conversation analysis. Maximum recording: two minutes. A model adapter interface keeps recognition replaceable.
Install or upgrade with the EXE; existing tasks and settings remain. The gzip asset is downloaded by the app automatically and is not a separate installer. Mainland preview only; the public hosted analysis service remains unavailable.
Validation: verified model downloads, corrupt-cache recovery, offline reuse, PCM conversion, silence handling and bounded inference; WPF capture confirmation/cancellation, full-text persistence, hotkey/language settings, native key conflicts, task priority/completion after analysis replacement; existing subscription/API settings regression; public Mandarin/Cantonese/English audio inference. Sample tests are not a general accuracy benchmark. Live microphone input still requires testing with the user's own device.
Built and documented by AI from the author's requirements. Third-party speech models and libraries were created by their respective authors; license notices are bundled.
v0.7.5 · API supplier selection
API settings now list DeepSeek, 豆包 / Doubao, OpenAI, Claude and other providers separately. Presets configure endpoint and protocol automatically; custom providers retain manual connection settings. Changing endpoints requires the corresponding API key. Existing tasks and settings are retained.
API 设置按供应商分别列出,自动配置常见供应商地址和接口格式,其他供应商保留自定义连接。切换地址后需填写对应密钥。已有任务和设置保留。
Validation: supplier selection, automatic endpoints/protocols, custom controls, encrypted key persistence, existing settings and API/relay tests.
v0.7.4 · Reliable conversation matching
简体中文
修复会话 ID 不匹配导致历史分析中断的问题。模型输入采用本批临时短编号,输出 schema 限制编号取值,再通过明确映射恢复真实会话 ID。遗漏或重复的会话只修正一次,仍失败保留断点;有效会话继续处理,不按结果位置猜测来源。
English
Fixes conversation ID mismatches interrupting history analysis. Requests use batch-local short IDs constrained by the output schema, with explicit mapping back to real conversation IDs. Missing or duplicate conversations receive one targeted correction attempt; persistent failures retain their checkpoint while valid conversations continue. Results are never assigned by position.
Validation: constrained IDs, out-of-order reverse mapping, existing-task context preservation, wrong-ID repair, duplicate-ID isolation, failed checkpoint preservation, evidence validation, initial history/pause/resume and API/relay tests.
v0.7.3 · Reliable source evidence validation
简体中文
修复原文依据校验失败导致整批历史分析停止的问题。空白、换行和全角字符差异会匹配回真实原文。引用不匹配时,仅对出问题的对话修正一次;仍失败则保留断点,并继续其他对话。不会接收凭空生成或拼接的依据。引用修正调用会消耗当前分析服务的额度。
English
Fixes source-quote validation failures stopping the entire history batch. Whitespace, line breaks and full-width character differences map back to the original text. Invalid quotes receive at most one targeted correction call. Persistent failures keep their checkpoint while other conversations continue. Fabricated and cross-message quotations remain rejected. Correction calls consume the selected service's quota.
Validation: normalized original excerpts including Unicode; fabricated/cross-message quote rejection; single targeted repair; persistent failure isolation and checkpoint resume; initial history/pause/resume, incremental/API/relay, subscription arguments tests. Tests use mocked analysis.
v0.7.2 · Complete initial history scan
简体中文
首次历史分析不再每次只处理六个对话。现在会自动连续分批处理全部可读取、未排除的历史对话及长对话的后续片段,显示进度,支持暂停与断点续接。已有分析结果保留;成功处理的消息不重复分析。首次完成后仍按设置的频率检查变化,网页采集仍只读取已加载内容。
English
The initial history scan now automatically drains all readable, non-excluded conversation batches, including subsequent chunks of long conversations, without waiting for another scheduled check. Progress, pause and resumable checkpoints are supported. Existing results are retained and completed messages are not analyzed again. Subsequent checks follow the configured interval; web capture remains limited to loaded messages.
Validation: initial queue beyond six conversations, long-conversation chunks, unchanged-history no-call, failure checkpoint resume, pause; existing incremental/API/relay and subscription argument tests. Tests use mocked analysis and incur no model charges.
v0.7.1 · Subscription model choices
简体中文
- Codex 模型改为下拉选择,优先读取本机官方模型目录,并保留自定义型号。
- Codex 推理强度可选,列表随模型支持范围变化;后台调用使用保存的选择。
- Claude 提供 Haiku、Sonnet、Opus、Fable 及具体版本下拉选项,保留自定义型号。
- 默认设置和已有任务保留。账号权限和官方客户端版本决定实际可用性;Fable 可能使用额外付费额度。
- 大陆内置服务仍锁定原有低价 DeepSeek 模型,内置服务尚未向公众开放。
English
- Codex model dropdown uses the local official model catalog when available, with custom IDs supported.
- Select reasoning effort from the selected model's supported levels. Analysis uses the saved choice.
- Claude model dropdown includes Haiku, Sonnet, Opus, Fable and specific versions, with custom IDs supported.
- Existing tasks and settings are retained. Availability depends on account permissions and official client versions. Fable may bill additional usage credits.
- Mainland app-funded service remains pinned to the existing low-cost DeepSeek model and is not publicly available yet.
Validation: model selection and persistence UI tests; model-dependent effort test; CLI argument/default/validation tests; existing incremental/API/relay tests; Windows installer build. No paid model calls were made for these checks.
v0.7.0 · 中国大陆预览版 / Mainland China Preview
v0.7.0 · 中国大陆预览版 / Mainland China Preview
简体中文
下载 ConversationTodo-Mainland-Setup-0.7.exe 即可安装。支持 Windows 10 / 11 x64;内置 Node.js,无需单独装运行环境或手动运行脚本。
本项目的应用代码、UI、打包配置和全部文档(包括本发布说明)均由 AI 生成并迭代,作者提出需求和反馈;第三方依赖由各自作者开发。
- 被动读取本机 Codex / Claude Code 对话,提取下一步并按优先级显示。
- Chrome 网页采集支持已打开、已加载的 ChatGPT / Claude 消息;扩展需按设置中的引导确认加载一次。
- 用户承诺自动进入待办,AI 建议待确认;支持完成、取消、忽略、暂停、排除、原文依据和托盘。
- 默认每 30 分钟检查,只有变化才分析;历史分批处理,避免反复上传全部对话。
- 当前只发布大陆版。作者付费的内置服务锁定
deepseek/deepseek-v4-flash-0731,客户端列表和后台权限共同限制型号,并限制供应商路由单价。
公共内置 API 尚未开放。 下载者请使用自己已安装并登录的 Codex / Claude Code,或自己的 API Key。必要文字会发送至选定云端服务,初次历史分析会消耗额度。
已检查增量逻辑、密钥加密、设置保存、型号限制、鉴权、配额、价格参数、私有后台拒绝高价型号、一次真实合成推理,以及安装/卸载。安装检查使用隔离模式,未替换现有 Chrome 注册和快捷方式。Chrome 真实页面和更多电脑仍待验证。安装包未签名。
详细要求、隐私及限制:简体中文说明。校验文件为 SHA256SUMS.txt。
English
Install ConversationTodo-Mainland-Setup-0.7.exe on Windows 10 / 11 x64. Node.js is bundled; no separate runtime installation or manual scripts are needed.
All project-specific application code, UI, packaging configuration, and documentation—including these release notes—were generated and iterated by AI. The author supplied requirements and feedback. Third-party dependencies retain their own authorship.
This mainland-only preview passively extracts prioritized actions from retained local Codex / Claude Code conversations, with optional Chrome capture of opened ChatGPT / Claude chats. Explicit commitments become tasks; AI suggestions require confirmation. It supports evidence, completion/cancellation, pause, exclusion, tray behavior, and incremental checks every 30 minutes by default.
The author's hosted service is pinned to deepseek/deepseek-v4-flash-0731, enforced by both client and server, with provider price caps. The public hosted API is not open. Use your own authenticated official CLI or API key. Analysis sends necessary text to the selected cloud service and consumes its quota. Chrome extension loading requires a one-time user action.
Checks covered incremental processing, settings, encryption, exact model restrictions, relay authentication/limits, price parameters, live rejection of a higher-priced model, one live synthetic inference, and isolated installation/uninstallation. Real Chrome pages and broader device compatibility need more testing. The installer is unsigned.
English README · Verify downloads with SHA256SUMS.txt.