Skip to content

Releases: showjiangnan/zcode-feiyu

ZCode Feiyu 3.14.3 — macOS arm64 / Windows x64

Choose a tag to compare

@showjiangnan showjiangnan released this 27 Sep 12:45

ZCode Feiyu / 飞鱼版 v3.14.3-feiyu.1

首个桌面预发布,同时提供 macOS Apple Silicon arm64 和 Windows x64。两个目标均在 Apple Silicon Mac 上由公开生产源码完成构建。

  • 构建源码 / Build source: 629ae0cad6d862ddb62b9dfedc10182a83fb6bc2
  • 上游基线 / Upstream baseline: ZCode v3.14.3 (29628c9acdb81b703bbd4080c207a0e7ce5e276e)
  • 应用 / App: ZCode Preview 3.14.3; bundled Agent: 0.16.9
  • 标签保留实际构建的生产源码提交;main 分支随后仅更新两份 README 的下载说明。

下载 / Downloads

安装包 / Package 大小 / Size
ZCode-Feiyu-3.14.3-mac-arm64.dmg 176.7 MiB
ZCode-Feiyu-3.14.3-mac-arm64.zip 168.3 MiB
ZCode-Feiyu-3.14.3-win-x64.exe 142.8 MiB

另附 SHA256SUMS.txt。macOS 用 shasum -a 256 <文件>、Windows PowerShell 用 Get-FileHash <文件> -Algorithm SHA256 计算摘要后比对。

飞鱼版增加的能力

  • 同一工作区本地顶层任务之间的发现、创建、读取、发送、回执查询、等待和生命周期控制;目标忙时自动发送使用 guide/steer,UI 持久显示“来自zcode其他任务自动发送”。
  • 统一遥测开关控制 ARMS、ZCode 事件及桌面/Agent OTLP;无明确启用配置时关闭。
  • 动态工作流、本地 Session Mailbox 默认开放;工作区文件搜索范围设置可见。
  • fal.ai 图片生成设置、模型选择、连接验证、生成与编辑、作业加载/结果显示及恢复/取消;支持按模型约束传递尺寸、数量、格式。Subagents 默认无权调用付费图片服务,可在设置中开启。

这是 ZCode 自有协议和共享运行时上的应用服务能力,不承诺 OpenAI Codex App Server 协议直接兼容。

已完成验证与边界

  • 冻结锁文件安装、pnpm typecheck、pnpm lint(0 errors / 70 existing warnings)通过。两个平台完整生产打包、运行时依赖闭包、原生资源策略和包体积审计通过。
  • Mac:DMG 和 ZIP 内应用的 117 个文件/符号链接条目与签名后的构建结果一致,严格签名验证通过;包内 Agent 启动通过。按现有 TerminalService 的 helper 权限准备逻辑完成终端原生模块 spawn/read/exit 冒烟。
  • Windows:完整 NSIS EXE 已生成;安装器的 85 个文件与打包目录逐一 SHA-256 一致,x64 主程序、PTY、Agent、图片技能和搜索资产已核对。未在 Windows 实机完成安装或运行,不宣称所有功能与 Mac 已验证一致。
  • Windows 交叉构建中的 SSH 可选原生加速模块无法使用,走库自带的 JS/Node crypto 实现;该回退路径在本机的握手与命令收发验证通过。这不是 Windows SSH 实机测试。
  • 本轮没有重复完整 GUI 回归或发起 fal 付费出图;前期已验证 macOS 安装及 fal 无付费连接。Windows 的终端、任务协作、图片生成、权限及升级需实机验收。

安装、签名与更新

  • Mac:打开 DMG 后拖入“应用程序”,或解压 ZIP 后复制 .app,再从“应用程序”启动。此包在生成归档前已做 ad-hoc 签名,没有 Apple Developer ID,也没有 Apple 公证。系统拦截时,在核验来源及摘要后按“隐私与安全性”提示处理。
  • Windows:运行 EXE 并按向导安装。安装器没有 Authenticode 发布者签名,可能触发未知发布者/SmartScreen 提示。
  • 后续从本仓库 Releases 下载新包,退出应用后覆盖安装并保留数据目录。没有飞鱼版自动更新服务;官方包不应覆盖本版。Preview 身份不保证数据隔离,需要隔离时使用 ZCODE_DATA_BASE_DIR。
  • 仅发布 Mac arm64 和 Windows x64,未发布 Intel Mac、Windows arm64 或 Linux 安装包。
  • fal.ai 需要用户在设置中自行提供 API Key 和可用额度;安装包不包含用户凭据或会话数据。设置方法与对比表见 中文 README。

English

This first desktop prerelease provides macOS arm64 DMG/ZIP packages and a Windows x64 NSIS installer, both built on Apple Silicon from the source commit above. The tag identifies the actual build source; subsequent changes on main only add bilingual download documentation.

Feiyu adds local top-level task coordination within one workspace (including automatic guide/steer delivery to busy tasks and visible message provenance), unified telemetry controls, enabled dynamic workflows/Session Mailbox/search settings, and built-in fal.ai image generation. Image access for subagents is disabled by default. This uses ZCode's own service protocol and runtime; it is not a claim of Codex protocol compatibility.

Frozen dependency installation, typecheck, lint, production packaging, runtime closure, native-resource policy, and size checks passed for both targets. Lint reported 70 existing warnings and no errors. Mac archive contents and strict signatures were checked, and bundled Agent and prepared PTY smoke checks passed. All 85 Windows installer payload files matched the packaged directory by SHA-256. Windows has not been installed or run on a real Windows machine, so terminal, collaboration, image generation, permissions, and upgrade behavior remain unverified there. The optional SSH native accelerator is unavailable in this cross-build; its JS/Node crypto fallback passed a local handshake and command-exchange check. No full GUI regression or paid fal generation was performed in this release run.

Mac packages are ad-hoc signed before archiving, without Developer ID or notarization. Windows installers have no Authenticode publisher signature. Verify SHA-256 digests; platform trust prompts can appear. On Mac, copy the app to Applications before launching. On Windows, run the installer wizard. For upgrades, quit the app and install the newer Feiyu package while keeping your data. There is no Feiyu automatic-update feed, and Preview does not by itself isolate user data. Configure your own fal key and credit in Settings. See the English README for setup, feature comparisons, and reproducible build commands.