Skip to content

Claude Code Haha v0.7.9

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Nov 10:57
· 13 commits to main since this release
abdcff5

Claude Code Haha v0.7.9

本次把 Agent Teams 变成真正可控的团队:可以暂停和恢复整个团队与主任务,暂停时还能改成员模型;新增 Claude 官方订阅额度、应用图标角标、会话 Markdown 导出、深灰主题和可拖拽的工作区预览。图片请求会自动压缩,大截图不再因为体积失败;同时修复了供应商设置互相覆盖、后台任务通知混进会话、分支与 worktree 不同步、MCP 走代理失败、H5 控制面鉴权被路径变体绕过等一批问题。手机 App 升级到 0.1.5。

核心功能

  • 团队运行控制(#1890,#1951,#1954,#1860):团队工作台新增“团队运行控制”。可以暂停团队与主任务,暂停状态刷新页面、重启应用后仍保持;暂停时可批量修改成员模型,主 Agent 的模型保持独立;恢复时按需拉起成员,部分失败可以“重试恢复团队”,不会重复启动已恢复的成员;恢复后等待中的工具审批正常下发,成员名单刷新也不会打断自动重试。给运行中的团队新增成员时,主 Agent 会收到待提交的团队计划并继续推进,不再卡在“主 Agent 正在调整方案”。
  • Claude 官方用量额度(#1861,#1949):设置 → 供应商 → Claude Official 登录后显示当前订阅的用量,包括 5 小时、每周以及接口返回的 OAuth 应用 / Opus / Sonnet 额度,并附重置时间和刷新按钮;接口不可用时说明原因,不猜测数值。切换登录账号后,模型目录和 ChatGPT Official 的上下文窗口会按当前账号重新解析。
  • 应用图标角标(#1940):开启系统通知后,macOS Dock 和 Windows 任务栏会显示“待处理请求 + 未读回复”的会话数量。处理审批或读取回复后数字自动减少,关闭系统通知后归零。
  • 会话 Markdown(#1908):侧栏右键会话新增“复制 Markdown”和“导出 Markdown”。导出的文件使用会话标题命名,图片和附件以占位符保留;历史不完整时明确提示,不会生成残缺文件。
  • 新外观与可调工作区(#1898,#1889):外观新增“深灰”主题;工作区预览支持拖动分隔条调整宽度,记住上次宽度,宽窗口和窄窗口都能自适应,在 125% UI Zoom 下拖动也不会偏移。
  • 输出额度升级(#1897):设置 → 通用新增 “Max output tokens 升级” 三态开关(自动 / 启用 / 禁用)。启用后,支持的线路先用较低输出额度,截断时再以更高的受支持额度重试;显式配置的输出上限始终优先,重试次数有上限。
  • 图片上下文压缩(#1893):发请求前会把过大的图片压缩到合适体积,多张图片共享内联预算。4K 截图和用 Read 工具读取的大图不再因为体积失败,历史消息里的预览仍使用原图。

修复

  • 供应商与设置不再互相覆盖(#1895,#1927,#1882):保存一个供应商不会回滚另一个供应商的配置;worktree 根路径和代理绕过设置在保存后保留;代理托管的重启不再丢失默认推理强度和界面语言;子 Agent 显式设置的模型 env 不会被默认配置覆盖;从 cc-switch 导入供应商时保留原有的额度查询脚本(#1861)。
  • 团队与计划评审(#1925,#1950):计划模式评审中切换执行模型会真正生效;提交的运行时被拒绝时,选择回退到之前的模型,卡片显示原因并可再次提交;团队计划在 leader 写入 transcript 期间也能完成编辑和批准,不再因为一次历史冲突而失败。
  • 聊天可靠性(#1930,#1922,#1936,#1947,#1937,#1929):停止会立即生效并中止代理流,不再出现迟到的“幽灵输出”;后台任务通知移出会话正文,统一进入活动面板;聊天里的图片/视频卡片只解析已经确认存在的工作区文件;工具图片在活动组折叠后仍然内联可见;工具耗时显示自身执行时间,不再混入排队和流等待;计划修订不再被计为失败。
  • 会话与工作区(#1926,#1880,#1919,#1881):会话头部和输入框实时跟随 git 分支,回合进行中进入 worktree 会立即同步,重启后恢复仍停在该 worktree;显式添加的嵌套项目保持独立,不会被父项目合并;历史翻页和重放不再出现重复消息。
  • 连接与安全(#1871,#1914,#1904):MCP 连接器验证会走网络代理并保留 host 代理路由,失败时区分网络/代理问题和鉴权问题;H5 控制面鉴权统一了路径规范化,重复斜杠等变体不能再绕过限制;远程连接诊断导出会保留失败阶段、隧道和代理信息。
  • 插件与上下文(#1928,#1900):启用插件后,新会话第一回合就能发现插件的 slash 命令;上下文用量计量与压缩共用同一套统计,旧摘要不会再让读数回滚。
  • OAuth 与手机推送(#1949,#1923,#1948):模型目录刷新会跟随账号变化,并避免旧代次刷新覆盖新 token;手机端把 Expo 推送凭据错误和代理失败分开提示,切换投递通道后旧反馈会立即清除。
  • 打包与构建(#1941,#1955):修复打包版启动失败的问题;macOS 产物缺少文件时构建会直接失败,不再产出无法启动的发布包。
  • 其他:收紧输出截断和 Codex 工具输入停滞时的重试边界;Claude Code 兼容版本更新到 2.1.321;移除关于页的微信群二维码。感谢 @toBerlinWay、@Ericwyc-max、@yuki、@yuehua-meng、@lxcxjxhx 等贡献者。

其他

  • Windows 安装包默认未签名:SignPath 签名申请仍在审批中,本版本 Windows 安装包未签名,运行时 SmartScreen 可能提示,见下方安装说明。
  • 手机 App 0.1.5:包含手机端推送错误分类和投递通道反馈修复;桌面端升级到 v0.7.9 后会自动构建并附上新的 APK,iOS 走 TestFlight。

安装

桌面端请在下方 Assets 中下载对应平台和架构的安装包:

  • macOS:Claude-Code-Haha-0.7.9-macOS-Apple-Silicon.dmg / Claude-Code-Haha-0.7.9-macOS-Intel.dmg
  • Windows:Claude-Code-Haha-0.7.9-Windows-x64-Setup.exe / Claude-Code-Haha-0.7.9-Windows-arm64-Setup.exe
  • Linux x64:Claude-Code-Haha-0.7.9-Linux-x64.AppImage、Claude-Code-Haha-0.7.9-Linux-x64.deb 或 Claude-Code-Haha-0.7.9-Linux-x64.rpm
  • Linux ARM64:Claude-Code-Haha-0.7.9-Linux-arm64.AppImage、Claude-Code-Haha-0.7.9-Linux-arm64.deb 或 Claude-Code-Haha-0.7.9-Linux-arm64.rpm

名字以 -update.zip 结尾的文件和 latest*.yml 是应用自动更新用的,不用手动下载。

Windows 安装包默认未签名:运行时若出现 SmartScreen 提示,确认文件来自本仓库的 Release 后,点「更多信息」→「仍要运行」。

手机 App 需要桌面端 v0.7.9 或更新版本:

  • Android:下载 Assets 中的 cc-haha-mobile-0.1.5-*.apk 安装,APK 会在桌面端安装包发布后自动附上。
  • iPhone / iPad:先在 App Store 安装 TestFlight,再打开 TestFlight 公开链接 加入测试。新构建需要苹果测试审核通过后,公开链接才会更新到 0.1.5。

完整变更:v0.7.8...v0.7.9

English Version

Claude Code Haha v0.7.9

This release makes Agent Teams genuinely controllable: you can pause and resume the whole team and its main task, and change member models while paused. It adds Claude plan usage, an app-icon badge, session Markdown export, a neutral dark theme and a resizable workspace preview. Image requests are compressed automatically, so large screenshots no longer fail on size. It also fixes providers overwriting each other's settings, background task notifications leaking into conversations, stale branches and worktrees, MCP through a proxy, and H5 control-plane auth bypasses through path variants. The phone app updates to 0.1.5.

Highlights

  • Team runtime controls (#1890, #1951, #1954, #1860): the team workbench adds a "Team runtime controls" section. You can pause the team and its main task; the paused state survives page refreshes and app restarts. While paused, member models can be changed in bulk while the main agent keeps its own model. Resume starts the members that need it, partial failures can be retried without starting already-resumed members again, pending tool approvals are delivered after resume, and roster refreshes do not interrupt automatic retries. Adding a member to a running team now leaves the lead with a draft plan to submit instead of getting stuck on "the main agent is adjusting the plan".
  • Claude plan usage (#1861, #1949): Settings → Providers → Claude Official shows the current subscription's usage after signing in, including the 5-hour window, the weekly window and any OAuth-app / Opus / Sonnet windows returned by the API, with reset times and a refresh button; when unavailable it explains why instead of guessing. After switching accounts, the model catalog and ChatGPT Official context windows are resolved for the current account.
  • App icon badge (#1940): with system notifications enabled, the macOS Dock and Windows taskbar show the number of sessions with pending requests or unread replies. The number drops as you handle approvals or read replies and returns to zero when notifications are off.
  • Session Markdown (#1908): the sidebar session context menu adds "Copy Markdown" and "Export Markdown". Exported files use the session title, images and attachments become placeholders, and incomplete history is reported instead of producing a partial file.
  • New appearance and adjustable workspace (#1898, #1889): Appearance adds a "Neutral dark" theme; the workspace preview divider can be dragged to resize, remembers its width, adapts to wide and narrow windows, and stays accurate at 125% UI zoom.
  • Output token escalation (#1897): Settings → General adds a three-state "Max output tokens escalation" toggle (Auto / Enabled / Disabled). When enabled, supported routes start with a lower output budget and retry with a higher supported budget on truncation; explicit output limits always win and retries are bounded.
  • Image context compression (#1893): oversized images are compressed before a request and share an inline budget. 4K screenshots and large images read with the Read tool no longer fail on size, while history previews still use the original attachments.

Fixes

  • Providers and settings no longer overwrite each other (#1895, #1927, #1882): saving one provider does not roll back another; worktree roots and proxy-bypass settings survive saves; proxy-managed restarts keep the default effort and UI language; explicit subagent model env values are preserved; cc-switch imports keep their quota query script (#1861).
  • Teams and plan review (#1925, #1950): changing the execution model during plan review now takes effect; when a submitted runtime is rejected, the previous model is restored, the card explains why and the plan can be submitted again; team plans can be edited and approved while the leader is rewriting its transcript.
  • Chat reliability (#1930, #1922, #1936, #1947, #1937, #1929): Stop takes effect immediately and aborts proxy streams, with no late "ghost" output; background task notifications move out of the transcript into the activity panel; image/video cards resolve only confirmed workspace files; tool images stay inline when the activity group folds; tool duration reports its own execution time instead of queue and stream waits; plan revisions are no longer counted as failures.
  • Sessions and workspaces (#1926, #1880, #1919, #1881): the session header and composer follow the live git branch; entering a worktree mid-turn syncs immediately and resumes there after a restart; explicitly created nested projects stay independent; paging and replay no longer duplicate messages.
  • Connectivity and security (#1871, #1914, #1904): MCP connector checks go through the network proxy and keep host proxy routes, and failures distinguish network/proxy problems from authentication problems; H5 control-plane authorization normalizes paths, so slash variants can no longer bypass restrictions; remote diagnostics exports keep the failure stage, tunnel and proxy details.
  • Plugins and context (#1928, #1900): plugin slash commands are discovered on the first turn of a new session; the context meter shares compaction accounting, so older summaries no longer roll readings back.
  • OAuth and phone push (#1949, #1923, #1948): model catalog refreshes follow account changes and stale generations no longer overwrite newer tokens; the phone distinguishes Expo push credential errors from proxy failures and clears old feedback when the delivery channel changes.
  • Packaging (#1941, #1955): packaged app startup is restored, and an incomplete macOS artifact now fails the build instead of shipping an app that cannot start.
  • Other: tighter retry bounds for truncated output and stalled Codex tool input; Claude Code compatibility updated to 2.1.321; the WeChat QR code is removed from About. Thanks to @toBerlinWay, @Ericwyc-max, @yuki, @yuehua-meng, @lxcxjxhx and other contributors.

Other

  • Windows installers unsigned by default: the SignPath signing application is still pending, so this release's Windows installers are unsigned and SmartScreen may warn; see the installation notes below.
  • Phone app 0.1.5: includes the push-error classification and delivery-channel feedback fixes; after the desktop upgrades to v0.7.9 the new APK is built and attached automatically, and iOS ships through TestFlight.

Installation

Download the desktop installer for your platform and architecture from the Assets section below:

  • macOS: Claude-Code-Haha-0.7.9-macOS-Apple-Silicon.dmg / Claude-Code-Haha-0.7.9-macOS-Intel.dmg
  • Windows: Claude-Code-Haha-0.7.9-Windows-x64-Setup.exe / Claude-Code-Haha-0.7.9-Windows-arm64-Setup.exe
  • Linux x64: Claude-Code-Haha-0.7.9-Linux-x64.AppImage, Claude-Code-Haha-0.7.9-Linux-x64.deb or Claude-Code-Haha-0.7.9-Linux-x64.rpm
  • Linux ARM64: Claude-Code-Haha-0.7.9-Linux-arm64.AppImage, Claude-Code-Haha-0.7.9-Linux-arm64.deb or Claude-Code-Haha-0.7.9-Linux-arm64.rpm

Files ending in -update.zip and latest*.yml are used by in-app auto-update and do not need to be downloaded manually.

The Windows installers are unsigned by default: if SmartScreen warns, confirm the file came from this repository's Release, then choose More info → Run anyway.

The phone app requires desktop v0.7.9 or later:

  • Android: install cc-haha-mobile-0.1.5-*.apk from the Assets; the APK is attached automatically after the desktop installers are published.
  • iPhone / iPad: install TestFlight from the App Store, then open the TestFlight public link to join. The public link updates to 0.1.5 after the new build passes Apple's beta review.

Full changelog: v0.7.8...v0.7.9