I build and maintain practical, local-first tools and AI workflows—Windows-first when desktop integration matters. My public work spans desktop utilities, self-hosted assistants, browser tools, and agent governance tooling.
我打造並維護實用的本機優先工具與 AI 工作流程；需要桌面整合時以 Windows 優先。公開作品涵蓋桌面工具、自架助理、瀏覽器工具，以及 agent 治理工具。
Batch-export books you legally own in Google Play Books: DRM-free titles download as EPUB/PDF, protected ones keep the official .acsm handoff for Adobe Digital Editions.
批次匯出自己合法擁有的 Google Play 圖書：無 DRM 的直接下載 EPUB/PDF，受保護的保留官方 .acsm 交接流程。
Python Playwright httpx PySide6 Local-first
Self-hosted personal AI assistant inside LINE — chat, voice, vision, search, schedules, tasks, reminders, weather — running on your own OpenAI, LINE, and Supabase credentials. 可自架的 LINE 個人 AI 助理：聊天、語音、圖片、搜尋、行程、任務、提醒、天氣，全跑在自己的帳號上。
Documentation site ｜文件站 · Derived from memochou1993/gpt-ai-assistant, independently maintained ｜衍生自上游，獨立維護。
JavaScript Node.js 24 Vercel Supabase LINE Messaging API Google Calendar / Tasks
Chat co-pilot for Android: reads the conversation on screen through accessibility services, judges the other person's intent and urgency, then drafts a reply into the input box — sending stays with you. LINE support is original to this fork. Android 聊天副駕：透過無障礙服務讀取螢幕上的對話，先判斷對方意圖與急迫程度，再起草回覆填進輸入框，發不發由你。LINE 支援是本 fork 的原創功能。
Derived from jev-chat/jev-chat-jarvis, independently maintained ｜衍生自上游，獨立維護。
Kotlin Python Android Accessibility Service LINE
Local-first voice typing for Windows: hotkey recording, on-device Faster-Whisper transcription, optional LLM rewriting and translation, typed straight into the focused app. Windows 本機語音輸入：快捷鍵錄音、本機 Faster-Whisper 辨識、可選 LLM 潤飾與翻譯，直接輸入目前作用中的程式。
Derived from jfamily4tw/voicetype4tw-mac, independently maintained ｜衍生自上游，獨立維護。
Python PyQt6 Faster-Whisper CUDA Windows
Portable YouTube channel archiving with GUI and CLI: batch workflows, filters, ffmpeg integration, and content the signed-in user is already authorized to watch. 可攜式 YouTube 頻道保存工具，GUI 與 CLI 並具：批次工作流、篩選、ffmpeg 整合，也支援登入者原本就有權觀看的內容。
Python yt-dlp Tkinter ffmpeg Windows EXE｜macOS / Linux from source
Windows VRM desktop companion that turns an AI assistant's playback into lip sync, motion, states, and message bubbles; compatible agents drive it over a loopback-only MCP server. Windows VRM 桌面角色：把 AI 助理的聲音轉成口型、動作、狀態與訊息氣泡，相容 Agent 可經僅限本機的 MCP 控制。
Derived from xikhar/persona, independently maintained ｜衍生自上游，獨立維護。
TypeScript Electron Three.js VRM / VRMA MCP
Making agent behaviour constrainable and provable, not merely requested in a prompt. Two distinct scopes: the coding agent that writes on my behalf, and autonomous agents once they are deployed. 讓 agent 的行為變成可約束、可舉證的，而不是只用提示詞請它照做。兩個範圍不同：替我寫程式的 coding agent，以及上線後自主運行的 agent。
Coding agent — five layers, each usable on its own ｜Coding agent 治理，五層，每層都可以單獨使用
| Layer / 層 | Repo | |
|---|---|---|
| Supply chain / 供應鏈 | SkillSpector | Scan an agent skill before installing it: risk score plus a SAFE / CAUTION / DO_NOT_INSTALL call, runnable as a skill or an MCP install gate ｜安裝 agent skill 前先掃描，輸出風險分數與安裝建議，可當 skill 或 MCP 閘門 |
| Dispatch / 派工決策 | agent-advisor | Risk-gated routing across four agent runtimes ｜四種 agent runtime 的風險分流路由 |
| Execution / 動作攔截 | harness-guard | Runtime hooks blocking dangerous commands, unevidenced completion claims, commits over failing tests ｜攔截危險指令、無證據的完成宣稱、紅燈仍提交 |
| Output / 產出品質 | ai-quality-gates | Gherkin specs, coverage and mutation gates, architecture contracts, bounded agent-loop policy ｜可執行規格、覆蓋率與 mutation gate、架構契約、有界 loop policy |
| Delivery / 交付流程 | agent-cortex | Candidate, verification, independent review, completion evidence ｜候選、驗證、獨立審查與完成證據 |
Deployed agents ｜上線 agent — agent-governance-toolkit answers what an agent is allowed to do, which agent did it, and how you prove it afterwards: policy enforcement, zero-trust identity, sandboxing, tamper-evident audit records. Python fork of microsoft/agent-governance-toolkit.
回答的是「這個動作准不准」「是哪一個 agent 做的」「事後怎麼舉證」：政策強制、零信任身分、沙箱執行與可稽核記錄。
| Category / 類別 | Highlights / 代表作 | |
|---|---|---|
| AI governance AI 治理 |
SkillSpector scan before install ｜安裝前掃描 · agent-advisor risk-gated dispatch ｜風險分流派工 · harness-guard runtime hooks ｜動作攔截 | all → |
| Agent skills Agent 技能包 |
agent-skills spec→ship flow ｜工程流程 · agent-astromech cross-model review ｜跨模型二審 · book-to-skill books as skills ｜技術書轉技能 | all → |
| Agent runtime & tooling Agent runtime 與工具鏈 |
opencodex swap the LLM ｜換掉背後的 LLM · agentdeck tray quota cockpit ｜系統匣額度控制台 · OpenSpec spec-driven development ｜規範驅動開發 | all → |
| Local-first desktop 本機桌面工具 |
openshelf export your books ｜圖書批次匯出 · channeldepot archive channels ｜頻道保存 · voxprose on-device dictation ｜本機語音輸入 | all → |
| Assistants & interfaces 助理與介面 |
gpt-ai-assistant assistant inside LINE ｜自架 LINE 助理 · chatgpt-sidebar browser side panel ｜瀏覽器側邊欄 · jev-chat-jarvis reply co-pilot on your phone ｜手機對話副駕 | all → |
| Content production 內容產製 |
html-anything agent writes the HTML ｜agent 寫 HTML · hyperframes render video from HTML ｜用 HTML 渲染影片 · video-autopilot-kit video pipeline ｜影片流程自動化 | all → |
| Other 其他 |
github-stars-organizer-playbook Stars Lists workflow ｜Stars Lists 整理 · public-apis free API directory ｜免費 API 目錄 · gpt-ai-assistant-docs documentation site ｜文件站 | all → |
Browse all public repositories ｜瀏覽全部公開 repo →
- Local-first by default ｜預設本機優先 — if it runs on the user's machine, it does not get a hosted backend first. 跑得動在使用者自己電腦上的東西，就不先架一個後端。
- Release-oriented ｜以交付為導向 — downloadable, runnable, verifiable beats a demo. 可下載、跑得起來、驗得出來，勝過一個 demo。
- Privacy-aware ｜重視隱私邊界 — user files, images, and tokens do not go to services that do not need them. 使用者的檔案、圖片與憑證，不送去不需要它們的服務。
- Explicit boundaries ｜寫明邊界 — licensing, platform rules, unsupported cases, and operational risks are stated, not implied. 授權、平台規則、不支援的情況與操作風險都寫出來，不靠讀者自己推測。
- AI-assisted, still verified ｜AI 協作但仍要驗證 — agents do a large share of the writing; what ships still has to pass tests, packaging, and stated limits. The governance repos above exist because "the agent said it was done" is not evidence. 程式有很大一部分是 agent 寫的，但要出貨仍得通過測試、打包與寫明的限制；上面那些治理 repo 存在的理由，就是「agent 說做完了」不算證據。
- Private work stays private ｜私人專案不公開 — this page shows public personal projects and selected open-source maintenance work. 本頁只展示公開的個人作品與精選的開源維護工作。
