v3.2.3

Highlights

• You can now search and fetch the web through a configurable, local-first toolchain. Web search supports ordered service/model fallback with visible attempts and usage, including built-in keyless Exa search. The new web_fetch tool can use local fetching, free keyless services, and configured Tavily, Exa, or DeepSeek providers. (#2891, #2892, #2901, #2957, #2976, #2982)
• Existing sessions can now switch agent backends without losing the conversation. Move a Claude Code, Codex, or OpenCode session to another enabled Agent with a handoff; if its current Agent is unavailable, Avibe can continue the held message on Vibey in one click. (#2916, #2934, #2946, #2950)
• Computer Use is faster and easier to set up. Desktop Settings now provides an actionable guide with live permission status, while vibe computer js becomes the primary agent interface. Faster window validation, cursor movement, and state checks reduce latency on common actions. (#2894, #2897, #2898, #2909, #2936)
• Agent routing is now consistent across the Workbench and IM. Slack, Discord, Telegram, and Feishu/Lark channels can route to any Agent, including user-created Agents, while preserving Vibey model and reasoning-effort choices. The Workbench picker now sits beside Send and shows catalog model names. (#2938, #2949, #2965, #2967, #2972)

Changes

Added

• Added link navigation, text selection, copying, and full-page scrolling to the phone PDF reader. (#3014, #3020, #3028)
• Added desktop updater status to connected browsers and phones, with update actions surfaced through the Workbench popover. (#2893, #2895, #3013)
• Added foreground/background visibility controls to the chat session menu, including Undo and a desktop session menu. (#2960)
• Added desktop-pet history above the input, top-edge placement with a flipped layout, and improved double-click behavior. (#2939, #2945)
• Added stalled Codex review detection so PR Watch can request another review instead of waiting indefinitely. (#2941)

Changed

• Moved the Agent/model picker into a two-row composer on chat, home, and new-session surfaces, giving message text more room on phones. (#2965, #2972)
• Prewarmed voice capture without opening the microphone, reducing the delay before the first spoken word. (#2943)
• Displayed update versions as plain text across Slack, Discord, Telegram, Feishu/Lark, and WeChat to avoid platform-specific formatting issues. (#2923)
• Signed and notarized every macOS release build with Developer ID. (#3005)
• Improved internal test, CI, dependency, and release-build performance and reliability.

Fixed

• Fixed cross-model history handling so one model’s private reasoning is never replayed as assistant text, malformed streamed tool calls recover safely, and raw citation markers degrade cleanly. (#2904, #2927, #2969)
• Fixed Anthropic, Gemini, Google RPC, and hosted-tool edge cases involving deferred server searches, thought signatures, overload replay, unknown stored tools, and concurrent proxied calls. (#2920, #2926, #3002, #3007, #3025, #3027)
• Fixed Model Hub retry and accounting behavior for managed response headers, Claude subscription reset times, prompt-cache TTL routes, source-specific opaque history, and keyless Exa diagnostics. (#2942, #2974, #2998, #3003, #3008)
• Fixed session handoffs to clear stale backend-specific model settings, preserve the Agent that authored each reply, settle unavailable-Agent turns, and keep held inputs recoverable. (#2930, #2946, #2950)
• Fixed Agent reasoning-effort selections being lost during catalog refreshes. (#2914)
• Fixed remote-access readiness checks so a replacement Cloudflare Tunnel is not considered ready before its edge connections are registered. (#2903)
• Fixed Computer Use screenshot-state detection, stale target-lock admission, retired-session cleanup, mixed-DPI pet hit testing, and Screen Recording relaunches that interrupted the Runtime. (#2929, #2940, #2947, #2951)
• Fixed dialog sizing and focus behavior so long names remain usable, skill dialogs stay within the viewport, and modal focus/Escape handling works consistently. (#2932, #2933, #2935)
• Fixed protected Vault flows in the desktop app by skipping incompatible sandbox warmup before browser handoff. (#2931)
• Fixed Run and Turn settlement under background-task completion or SQLite writer contention, and preserved hosted cards in background sessions. (#2944, #2952, #2975)
• Fixed transient Windows payload rename failures during desktop packaging. (#2953)

Security

• Upgraded bundled Node.js, CPython, Python packages, Rust TLS/QUIC libraries, cloudflared, Model Hub Engine, and avault to patched releases addressing published security advisories. (#2910, #2911, #2912, #2913, #2917, #2918, #2921, #3006, #3015)
• Updated vibe upgrade to move uv-managed Python installations to the newest available patch release so existing installs receive CPython and OpenSSL fixes. (#2919)

Breaking

• Unified the Agent-facing CLI around --session-id, caller-session defaults, compact records, concise help, and consistent output; removed legacy aliases, duplicate fields, and older command forms. (#2961, #2971, #2973, #2977, #2980, #2981, #2983)
• Scoped Session-owned list commands to the calling Agent’s Session by default instead of returning records from every Session. (#2961)

Full Changelog: v3.2.2...v3.2.3

---

亮点

• 现在可以通过可配置、优先本地的工具链完成网页搜索和抓取。 网页搜索支持服务与模型混排回退，可查看每次尝试和用量，并内置无需密钥的 Exa 搜索。新的 web_fetch 可使用本地抓取、免费免密服务，以及已配置的 Tavily、Exa 或 DeepSeek。 (#2891, #2892, #2901, #2957, #2976, #2982)
• 已有会话现在可以切换 Agent 后端，并接着原对话继续。 Claude Code、Codex 或 OpenCode 会话可携带上下文切换到其他已启用 Agent；当前 Agent 不可用时，也能一键转到 Vibey，继续处理刚才未执行的消息。 (#2916, #2934, #2946, #2950)
• 电脑操作的配置更直观，执行也更快。 桌面端设置新增带实时权限状态的操作向导，vibe computer js 成为 Agent 使用电脑操作的主要入口；窗口校验、光标移动和状态检查的延迟也明显降低。 (#2894, #2897, #2898, #2909, #2936)
• Workbench 与 IM 的 Agent 路由方式现已统一。 Slack、Discord、Telegram 和飞书/Lark 频道可以选择任意 Agent，包括自定义 Agent，同时保留 Vibey 的模型与推理强度设置。Workbench 的选择器移到发送按钮旁，并显示模型目录中的名称。 (#2938, #2949, #2965, #2967, #2972)

变更

新增

• 新增手机 PDF 阅读器的链接跳转、文字选择与复制，并支持滚动浏览所有页面。 (#3014, #3020, #3028)
• 新增桌面端更新状态同步，连接到同一 Runtime 的浏览器和手机可查看结果，并从 Workbench 弹窗执行更新操作。 (#2893, #2895, #3013)
• 新增会话前台/后台切换、撤销操作和桌面端会话菜单。 (#2960)
• 新增桌面 Vibey 的输入框上方历史记录、靠近顶部时向下展开的布局，并优化双击交互。 (#2939, #2945)
• 新增 Codex Review 卡住检测，让 PR Watch 能重新请求审查，不再无限等待。 (#2941)

调整

• 将 Agent/模型选择器移入聊天、首页和新建会话的双行输入框，手机端输入区域更宽。 (#2965, #2972)
• 预热语音采集组件但不提前打开麦克风，缩短首次录音的启动等待。 (#2943)
• 将 Slack、Discord、Telegram、飞书/Lark 和微信中的更新版本号改为普通文本，避免各平台格式解析异常。 (#2923)
• 为所有 macOS 发布构建加入 Developer ID 签名和公证。 (#3005)
• 优化测试、CI、依赖维护和发布构建流程的速度与稳定性。

修复

• 修复跨模型历史处理，不再把其他模型的私有推理当作回复文本；同时完善畸形流式工具调用恢复和无注解引用标记降级。 (#2904, #2927, #2969)
• 修复 Anthropic、Gemini、Google RPC 和托管工具的多项边界情况，包括延后返回的服务端搜索、思考签名、过载重放、未知历史工具及代理调用并发。 (#2920, #2926, #3002, #3007, #3025, #3027)
• 修复 Model Hub 对托管响应头、Claude 订阅额度重置时间、提示缓存 TTL、Source 专属原生历史及免密 Exa 诊断信息的处理。 (#2942, #2974, #2998, #3003, #3008)
• 修复会话切换后残留旧后端模型设置、历史回复 Agent 标注错误、不可用 Agent 的 Turn 未结算，以及待处理输入无法恢复的问题。 (#2930, #2946, #2950)
• 修复刷新模型目录时推理强度选择被吞掉的问题。 (#2914)
• 修复远程访问就绪判断，Cloudflare Tunnel 的替代连接尚未在边缘注册完成时不会提前标记为可用。 (#2903)
• 修复电脑操作的截图状态判断、过期目标锁、已结束会话清理，以及混合 DPI 下桌面 Vibey 点击失效和屏幕录制重启中断 Runtime 的问题。 (#2929, #2940, #2947, #2951)
• 修复弹窗尺寸与焦点管理，长名称不再挤出操作按钮，技能弹窗不会超出视口，Esc 和焦点回收也恢复正常。 (#2932, #2933, #2935)
• 修复桌面端受保护 Vault 流程，在跳转系统浏览器前不再启动不兼容的沙箱预热。 (#2931)
• 修复后台任务结束或 SQLite 写入竞争时的 Run/Turn 结算，并保留后台会话中的托管工具卡片。 (#2944, #2952, #2975)
• 修复 Windows 桌面打包过程中偶发的 payload 重命名失败。 (#2953)

安全

• 升级内置 Node.js、CPython、Python 依赖、Rust TLS/QUIC 库、cloudflared、Model Hub Engine 和 avault，纳入已公布漏洞的修复版本。 (#2910, #2911, #2912, #2913, #2917, #2918, #2921, #3006, #3015)
• 调整 vibe upgrade，升级时会把 uv 管理的 Python 更新到最新补丁版本，让已有安装也能获得 CPython 与 OpenSSL 安全修复。 (#2919)

破坏性变更

• 统一面向 Agent 的 CLI：使用 --session-id、默认采用调用方会话、输出紧凑记录与精简帮助；旧别名、重复字段和旧命令形式已移除。 (#2961, #2971, #2973, #2977, #2980, #2981, #2983)
• 将会话所属资源的列表命令默认限定到当前 Agent 的会话，不再返回所有会话的数据。 (#2961)

Full Changelog: v3.2.2...v3.2.3