lazyBoy/CHANGELOG.md

Changelog

All notable changes to LazyBoy are documented here.

[Unreleased]

Long computer-use runs no longer die when the prompt creeps a few dozen tokens over the model's context window. Stale browser snapshots are stubbed, oversized tool dumps are capped, and an exceed_context_size response compact-retries the same turn instead of failing the job.

Each model provider keeps its own API key, model, and endpoint in the database. Switching xAI → OpenCode Go no longer wipes the xAI key, and the settings field shows dots instead of the secret.

Sign in with your own account. The shared install token is gone, and so are the model keys in the environment.

• Register and log in with a username and password in the app. Passwords are stored as PBKDF2-HMAC-SHA256 (120,000 rounds, per-password salt) and sessions live in the database as token digests in an HttpOnly, SameSite=Strict cookie. Repeated failures cool a username down for five minutes.
• One workspace per account. Agents, chats, runs, schedules, credential vaults, and model keys belong to the account that created them; a second account on the same install sees none of it. Work created before this release is adopted by the first account that registers.
• Model keys come from Settings → Models only. LAZYBOY_APP_TOKEN, XAI_API_KEY, OPENCODE_GO_API_KEY, and OPENAI_API_KEY are no longer read. A workspace without a key does not quietly run on someone else's: the run fails immediately and points at the setting to fix.
• LAZYBOY_ALLOWED_HOSTS lists the domain names an install may be addressed by (DNS rebinding guard). IPs and localhost keep working with no config.
• Removed: the token login screen, LAZYBOY_APP_TOKEN, and the rule that a non-loopback bind required a 32-character token. SANDBOX_SUPERVISOR_TOKEN and LAZYBOY_VAULT_KEY are still required in .env.
• Context window fitting. Long computer-use runs stub stale browser snapshots, cap tool dumps, and compact-retry on exceed_context_size instead of failing the job 42 tokens over a 128k window. Tune with LAZYBOY_MODEL_CONTEXT_CHARS.
• Model keys stay put. Each provider keeps its own key, model, and base URL. Switching provider no longer clears the previous key; the form shows dots and never echoes the token.

繁體中文

長時間操作電腦時，舊的網頁快照不再把模型上下文塞爆：較舊的觀察會收成摘要，工具輸出有上限，模型回 exceed_context_size 時會壓縮後重試同一輪，而不是整份任務失敗。

每個供應商的 API 金鑰、模型與端點會各自記住。從 xAI 換到 OpenCode Go 不會清掉另一家的金鑰；設定畫面只顯示圓點，真正的 token 留在資料庫。

改成自己的帳號登入：共享的安裝 token 拿掉了，環境變數裡的模型金鑰也拿掉了。

• 在 App 裡註冊與登入：密碼以 PBKDF2-HMAC-SHA256（120,000 輪、每組密碼各自的 salt）存放，session 在資料庫只存雜湊，cookie 是 HttpOnly + SameSite=Strict。 同一帳號連續失敗會冷卻五分鐘。
• 一個帳號一個工作區：Agent、對話、執行紀錄、排程、憑證庫與模型金鑰都只屬於 建立它的帳號，同一套安裝的第二個帳號看不到任何資料；舊資料由第一個註冊的帳號接手。
• 模型金鑰只吃「設定 → 模型」：不再讀 LAZYBOY_APP_TOKEN、XAI_API_KEY、 OPENCODE_GO_API_KEY、OPENAI_API_KEY。沒設金鑰的工作區不會偷偷用別人的金鑰， 任務會立刻失敗並指出該去哪裡補。
• LAZYBOY_ALLOWED_HOSTS：要改用網域名稱連進來時列在這裡（DNS rebinding 防護）； 用 IP 或 localhost 不用設定。
• 移除：token 登入畫面、LAZYBOY_APP_TOKEN，以及「綁非 loopback 需 32 字元 token」 的規則。.env 仍需要 SANDBOX_SUPERVISOR_TOKEN 與 LAZYBOY_VAULT_KEY。
• 模型上下文配額。 長時間操作電腦時，較舊的網頁快照會收成摘要、工具輸出有上限， 模型回 exceed_context_size 時會壓縮後重試同一輪，而不是只超出幾十個 token 就 整份任務失敗。可用 LAZYBOY_MODEL_CONTEXT_CHARS 調整。
• 模型金鑰會記住。 每個供應商的金鑰、模型與端點各自存在資料庫。從 xAI 換到 OpenCode Go 不會清掉 xAI 的金鑰；畫面上只顯示圓點，真正的 token 不會再出現。

[v0.1.0-alpha] - 2026-09-09

First public alpha. Self-hosted AI agent workspace: each agent gets an isolated Linux desktop, you watch it live, and you can take over at any time.

This is a source release. There is no pre-built binary; make up builds the API, supervisor, and desktop images from this tree.

Highlights

• A real computer per agent — Debian + XFCE + Chromium in Docker. Browser, terminal, files, and GUI, with a live noVNC view.
• Cua desktop driver — clicks, typing, windows, clipboard, and the visible terminal go through Cua (cua-driver-rs v0.23.2). The old AT-SPI / CDP path is gone.
• Chat that carries work — text, images, and file attachments; the agent opens them on its own desktop. Inbox copies expire on their own.
• Take over, then hand back — sign in, pass a check, or nudge the same desktop by hand. Saved logins live in an encrypted vault and never pass through the model.
• Several agents and groups — shared Team computers or private desktops. @name wakes one agent; otherwise a short router picks who should answer.
• Teach by demo, then schedule — turn a walkthrough into a skill; cron runs it again.
• Your models and tools — xAI, OpenCode Go, OpenAI-compatible endpoints, MCP, and file skills.
• Voice calls — talk to the agent after you enable a voice provider.
• Desktop and phone UI — English and Traditional Chinese. On a phone: collapsible chat, keyboard, trackpad, right-click, drag, and scroll.
• Runs that finish the job — no hard turn quota on real work. The agent has to verify against the current screen before it stops; looping is paused for you instead of being cut off. Hover the thinking avatar for the live trace; a failed run can be retried.
• Memory and hygiene — optional long-term memory (pgvector), room-scoped recall, and retention jobs for traces, recordings, and checkpoints.

Install

You need Docker and Compose, Git, Make, Python 3, and an API key for a supported model.

git clone https://github.com/igs170911/LazyBoy.git
cd LazyBoy
git checkout v0.1.0-alpha
make env

Edit .env and set one of XAI_API_KEY, OPENCODE_GO_API_KEY, or OPENAI_API_KEY. Then:

make up
make health

Open http://127.0.0.1:3101 and sign in with LAZYBOY_APP_TOKEN. The first run builds images from source and takes a while.

Alpha caveats

• Early software. Expect rough edges in the desktop driver, group routing, and long-running tasks.
• The API binds 127.0.0.1 by default. For a phone on your network set LAZYBOY_BIND_IP=0.0.0.0 (token must be at least 32 characters). Put it behind HTTPS before you expose it.
• Do not publish supervisor port 7091 or mount the host Docker socket into an agent computer.
• You bring the model API key. Screenshots and page contents can still reach the model; keep passwords in the vault, not in the prompt.

See Operations for resource limits, environment variables, and the security model.

繁體中文

第一個公開 alpha。自架 AI Agent 工作空間：每個 Agent 有自己的隔離 Linux 桌面，你可以即時觀看，也可以隨時接手。

這是原始碼發行版，沒有預先編好的執行檔；make up 會從此樹狀目錄建出 API、supervisor 與桌面映像。

重點

• 每個 Agent 一台電腦 — Docker 裡的 Debian + XFCE + Chromium。瀏覽器、終端、檔案與圖形介面，畫面用 noVNC 即時看。
• Cua 桌面驅動 — 點擊、打字、視窗、剪貼簿與可見終端都走 Cua（cua-driver-rs v0.23.2）。舊的 AT-SPI / CDP 路徑已移除。
• 對話裡就能交工作 — 文字、圖片、檔案附件；Agent 在自己的桌面開啟。收件匣複本會到期清除。
• 隨時接管再交回 — 同一桌面登入、過驗證或手動調整。帳密放加密保險庫，不會經過模型。
• 多 Agent 與群組 — Team 共用電腦或 Private 獨立桌面。@名字 只叫醒一個人；沒點名時由短路由決定誰回。
• 示範教學與排程 — 把操作示範收成技能，再用 cron 重複執行。
• 自選模型與工具 — xAI、OpenCode Go、OpenAI 相容端點、MCP、檔案技能。
• 語音通話 — 設定語音服務後可以直接通話。
• 桌面與手機介面 — 英文與繁體中文。手機可收合聊天側欄，遠端桌面有鍵盤、觸控板、右鍵、拖曳與捲動。
• 任務以做完為準 — 正式工作不再用固定輪數腰斬。Agent 要對著當下畫面自我驗證才准停；鬼打牆會暫停等你，而不是直接切斷。滑過思考頭像可看即時紀錄，失敗的 run 可以重試。
• 記憶與清理 — 可選的長期記憶（pgvector）、依房間範圍回想，以及紀錄、錄影、checkpoint 的保留週期。

安裝

需要 Docker 與 Compose、Git、Make、Python 3，以及一組支援的模型 API 金鑰。

git clone https://github.com/igs170911/LazyBoy.git
cd LazyBoy
git checkout v0.1.0-alpha
make env

編輯 .env，設定 XAI_API_KEY、OPENCODE_GO_API_KEY 或 OPENAI_API_KEY 其中之一，然後 make up 與 make health。打開 http://127.0.0.1:3101，用 LAZYBOY_APP_TOKEN 登入。第一次會從原始碼建映像，需要一段時間。

Alpha 注意事項

• 早期軟體。桌面驅動、群組路由、長任務都還可能不穩。
• API 預設只綁 127.0.0.1。要給區網手機用請設 LAZYBOY_BIND_IP=0.0.0.0（token 至少 32 字）。對外之前請放在 HTTPS 後面。
• 不要把 supervisor :7091 對外，也不要把主機 Docker socket 掛進 Agent 電腦。
• 模型金鑰自己準備。截圖與頁面內容仍可能進模型；密碼放保險庫，不要寫在提示詞裡。

資源上限、環境變數與安全模型見 部署指南。