Claude Code Haha v0.7.5

Claude Code Haha v0.7.5

本次带来 跨会话模型分工、重做后的 Token 用量页 和 实验性的精简 Agent 模式，工作区也支持给 Word、PDF、表格加批注；同时修复 OpenCode Go 的 Gemini 模型不可用、macOS 全屏、缓存命中率统计、Windows 数据迁移权限预检等一批问题。

核心功能

• 跨会话模型分工（#1626，#1628，#1629，#1631）：一个会话可以把任务分给若干子会话，每个子会话单独指定供应商和模型；也可以先看一份方案，再决定要不要启动。分工列表移到了「活动」栏，展开就能看到每个会话用哪个模型、跑到哪一步。已启动的会话也能换模型，可以选「本轮结束后切换」或「立即暂停并切换」，切换没生效前会话照常用原来的模型把当前轮跑完；一次最多修改 8 个已启动会话。
• Token 用量页重做（#1612，#1618，#1627）：页面开头直接给出生命周期总量（中文按万/亿、英文按 k/M/B 显示），下面是按模型拆分的 90 天柱状流：每天一根柱子、零基线，模型颜色固定不变，排名变动也不会换色。指针指向某一天就能看到当天总量和每个模型各占多少，点一下会钉出一张当天卡片（含最忙的一天与活跃日均值对比、当天的会话数和工具调用数）。还可以在本地生成竖版、方形或横版的用量分享图，主题跟随设置。
• 精简 Agent 模式（实验性）（#1613，#1619）：桌面「设置 → 通用」可以选择标准或精简，新建会话在发出第一条消息之前也能临时选。精简模式使用更短的系统提示词，只保留读写文件、搜索、终端等核心工具，连接器（MCP）、子 Agent、智能体团队、计划模式等一律不可用，技能仍可按需调用。模式在会话创建后锁定，想换只能新建聊天。
• 工作区文档批注（#1608）：Word、PDF 和表格预览里可以直接选中文字或单元格加批注，批注会带上文件、页码或工作表位置、引用内容和你的修改要求一起发到对话里；同一个文件可以攒多条批注再一起发。

修复

• OpenCode Go 的 Gemini 模型不可用（#1630）：OpenCode Go 把每个模型绑定到一种协议，gemini-* 只在 Google 的 generateContent 接口上提供服务，此前一律落到 chat/completions 并报 400（ModelProtocolUnsupported）。现在按 generateContent 协议代理，请求、回复和分页续传都支持，函数调用的思考签名也能原样回传；桌面模型选择器里也能按协议分组显示。
• macOS 全屏（#1606）：主窗口恢复原生全屏，绿色按钮和「视图 → 进入全屏」都走系统全屏（会开独立 Space），不再只是放大窗口；隐藏窗口前会先退出全屏，不再留下黑色桌面。Windows 上按 F11 也能全屏。
• 缓存命中率统计（#1621）：中转站返回的 usage 里 prompt_tokens 已经算上缓存命中的部分时，不再把命中量重复加一遍，命中率回到正常值。
• Agent Teams 相关界面（#1621）：计划弹窗里模型写的长文本不再撑出横向滚动条；很短的消息下方「编辑」按钮不再点不到；队伍还在干活时，会话标签上的运行圆点会亮。
• Windows 数据迁移（#1620）：迁移前先在临时目录里试做同样的权限操作，不再等数据全部复制完才发现没法保留权限；失败时能看出是哪个操作、哪个条目失败，且不包含路径和账号信息。
• 项目选择菜单（#1609）：窗口较矮时，菜单按触发按钮上下的剩余空间限高并在内部滚动，「新建项目」「选择其他文件夹」不再被挤出窗口。
• IM 设置的「已保存」提示（#1610）：提示改为由页面状态驱动，离开设置页时定时器会一起清掉，不再留下残留的定时任务。

其他

• 桌面宠物的等待徽章（#1534）：宠物窗口新增一个独立的琥珀色徽章，标出有多少个对话在等你回复（等待输入或待批准），展开后可直接跳到对应会话。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

• macOS：Claude-Code-Haha-0.7.5-mac-arm64.dmg / Claude-Code-Haha-0.7.5-mac-x64.dmg
• Windows：Claude-Code-Haha-0.7.5-win-x64.exe / Claude-Code-Haha-0.7.5-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.7.5-linux-x86_64.AppImage、Claude-Code-Haha-0.7.5-linux-amd64.deb 或 Claude-Code-Haha-0.7.5-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.7.5-linux-arm64.AppImage、Claude-Code-Haha-0.7.5-linux-arm64.deb 或 Claude-Code-Haha-0.7.5-linux-aarch64.rpm

完整变更：v0.7.4...v0.7.5

English Version

Claude Code Haha v0.7.5

This release adds model delegation across sessions, a redesigned Token usage page and an experimental lean Agent profile, and lets you annotate Word, PDF and spreadsheet files in the workspace. It also fixes unusable Gemini models on OpenCode Go, macOS fullscreen, cache-hit accounting, Windows data-migration permission preflight and more.

Highlights

• Model delegation across sessions (#1626, #1628, #1629, #1631): a session can hand work to child sessions, each with its own provider and model, or present an editable proposal before anything launches. Delegated sessions now live in the Activity rail, where every row names the model it will run and how far it has got. Running sessions can switch models too — either after the current turn or by pausing immediately — and keep their old model until the switch takes effect. At most eight running sessions can be changed at once.
• Token usage page redesigned (#1612, #1618, #1627): the lifetime total now leads the page in your own units (万/亿 for Chinese, k/M/B for English), followed by a 90-day stream split by model. Each day is one zero-based column, and model colors are fixed so a model keeps its color when the ranking changes. Pointing at a day shows its total and every model's share; clicking pins a card with that day's figures against the busiest and average active day, plus its sessions and tool calls. You can also render a portrait, square or wide usage image locally, in the current or any chosen theme.
• Lean Agent profile (experimental) (#1613, #1619): pick standard or lean in Settings → General, or temporarily before sending the first message of a new session. The lean profile uses a shorter system prompt and keeps only the core file and terminal tools; connectors (MCP), sub-agents, agent teams and plan mode are unavailable, while skills can still be invoked on demand. The mode is locked once the session exists — switch back by starting a new chat.
• Workspace document annotations (#1608): select text in Word and PDF, or cells and ranges in spreadsheets, and attach an annotation. The annotation carries the file, its page or sheet location, the quoted content and your requested change into the chat, and several annotations on one file can be sent together.

Fixes

• Gemini models on OpenCode Go (#1630): OpenCode Go binds each model to one wire protocol and serves Gemini only at Google's generateContent endpoint, so gemini-* models fell through to chat/completions and every request failed with 400 ModelProtocolUnsupported. They are now served over generateContent, including streamed replies and continuations, and the thought signatures of function calls are replayed as-is; the desktop model picker also groups them by protocol.
• Fullscreen on macOS (#1606): the main window is fullscreenable again, so the green button and View > Toggle Full Screen use native fullscreen (with its own Space) instead of merely zooming the window, and hiding the window leaves fullscreen first, so no black desktop is left behind. F11 toggles fullscreen on Windows.
• Cache-hit accounting (#1621): when a relay reports OpenAI-shaped usage whose prompt_tokens already include the cache hit, the hit count is no longer added a second time, and the hit rate is correct again.
• Agent Teams surfaces (#1621): long model-written text in the plan dialog no longer forces a horizontal scrollbar, the edit button under a very short message is reachable again, and the session tab shows a running dot while a team is still working.
• Windows data migration (#1620): the same permission operation is now tried on empty probes in a private staging area before any data is copied, so an impossible permission transfer fails early instead of after copying everything. Failures report which operation and entry failed, without paths or account names.
• Project picker menu (#1609): in a short window the menu is capped at the room left on the chosen side and scrolls inside, so New project and Choose another folder stay reachable.
• IM settings "Saved" hint (#1610): the fade is driven by the save status, and leaving the page clears the timer instead of leaving one behind.

Other

• Waiting-session badge for the desktop pet (#1534): the pet window gains a separate amber badge counting the conversations waiting for you — an input or an approval — and expanding it takes you straight to that session.

Installation

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

• macOS: Claude-Code-Haha-0.7.5-mac-arm64.dmg / Claude-Code-Haha-0.7.5-mac-x64.dmg
• Windows: Claude-Code-Haha-0.7.5-win-x64.exe / Claude-Code-Haha-0.7.5-win-arm64.exe
• Linux x64: Claude-Code-Haha-0.7.5-linux-x86_64.AppImage, Claude-Code-Haha-0.7.5-linux-amd64.deb or Claude-Code-Haha-0.7.5-linux-x86_64.rpm
• Linux ARM64: Claude-Code-Haha-0.7.5-linux-arm64.AppImage, Claude-Code-Haha-0.7.5-linux-arm64.deb or Claude-Code-Haha-0.7.5-linux-aarch64.rpm

Full changelog: v0.7.4...v0.7.5

Claude Code Haha v0.7.4

Claude Code Haha v0.7.4

本次首要修复 Windows 上新建或恢复会话时反复弹出的黑色终端窗口，并把 Agents、工作流和斜杠命令提到 + 菜单一级；同时修复 Linux 包终端打不开、输入框上方浮层被裁、消息选中菜单跨会话串用 等问题。

紧急修复

• Windows 新建或恢复会话时反复弹出黑色终端窗口（#1593，#1583）：根因是桌面 sidecar 被打包成了 GUI 子系统，它自己和它拉起的控制台子进程（例如每次启动 CLI 都会跑的机器码探测）因此各开一个可见窗口。v0.7.3 只处理了其中一条路径，本次把 sidecar 恢复成控制台程序，并在构建时检查子系统类型，避免再次回退。

核心改进

• + 菜单一级直达 Agents、工作流和斜杠命令（#1601，#1597）：运行一个 Agent 或工作流不再需要先展开「更多工具」，一次悬停或点击就能到；「更多工具」这一行已移除，手机底部面板同样少一级。

修复

• Linux 包的终端面板（#1602，#1591）：此前的 Linux 包没有带上 node-pty 的原生模块，终端面板一律报「Failed to load native module: pty.node」，现在安装包会正确编译并校验该模块；同时把 Linux 模块改为按 glibc 2.28 编译，Ubuntu 20.04、Debian 11、RHEL 8 等环境也能正常使用终端。
• 输入框上方的浮层被页面裁掉（#1598，#1587）：新会话时输入框居中、或把窗口压矮、或打开底部终端把输入框顶上去时，斜杠命令、@ 引用、+ 菜单和 /help 面板都会按上方剩余空间限高并在内部滚动，不再切掉顶部，也不会盖住 + 按钮。
• 文本选中菜单跨会话串用（#1590）：当两个会话存在 id 与内容都相同的消息时（例如主会话和它的侧边对话），用快捷键切换会话后残留的「添加到对话 / 在侧边对话中提问」菜单不再生效，引用和侧边对话只作用于做选中的那个会话。
• 上游在流里报出的错误（#1600，#1592）：中转站返回 200、却在 SSE 体里报 api_error / overloaded_error 时，现在按 API 重试预算重试（默认 10 次、间隔递增），不再是 2 秒内重试两次就结束；而上下文超长、额度不足、请求非法这类重发也没用的错误会直接报出来，不再重试。

其他

• 文档（#1599）：中英文 README 增加「最全 API 中转站导航」推荐。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

• macOS：Claude-Code-Haha-0.7.4-mac-arm64.dmg / Claude-Code-Haha-0.7.4-mac-x64.dmg
• Windows：Claude-Code-Haha-0.7.4-win-x64.exe / Claude-Code-Haha-0.7.4-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.7.4-linux-x86_64.AppImage、Claude-Code-Haha-0.7.4-linux-amd64.deb 或 Claude-Code-Haha-0.7.4-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.7.4-linux-arm64.AppImage、Claude-Code-Haha-0.7.4-linux-arm64.deb 或 Claude-Code-Haha-0.7.4-linux-aarch64.rpm

完整变更：v0.7.3...v0.7.4

English Version

Claude Code Haha v0.7.4

This release first fixes the black console windows Windows users saw on every new or resumed session, puts agents, workflows and slash commands on the + menu's first level, and repairs the Linux terminal panel, clipped composer popovers, message selections leaking across sessions and more.

Critical fix

• Black console windows on every new or resumed session on Windows (#1593, #1583): the desktop sidecar was packaged as a GUI-subsystem binary, so it and every console child it spawned (such as the machine-id probe that runs at each CLI start) opened a visible window. v0.7.3 only covered one of those paths; the sidecar is a console program again, and the build now checks the subsystem so it cannot regress.

Highlights

• Agents, workflows and slash commands on the + menu's first level (#1601, #1597): running an agent or a workflow no longer needs the More tools submenu — one hover or click reaches it. The More tools row is gone, and the phone sheet loses the extra level too.

Fixes

• Terminal panel in Linux packages (#1602, #1591): Linux packages shipped without node-pty's native module, so the terminal panel failed with "Failed to load native module: pty.node". Packages now compile and verify that module, and the Linux build targets glibc 2.28 so Ubuntu 20.04, Debian 11 and RHEL 8 can use the terminal as well.
• Composer popovers clipped by the page (#1598, #1587): with the composer centred on a new session, in a short window, or pushed up by the bottom terminal, the slash menu, @ references, the + menu and the /help panel are now capped at the room above them and scroll inside; the top is no longer cut off and the + button stays uncovered.
• Text-selection menu leaking across sessions (#1590): when two sessions hold a message with the same id and content (a chat and its side chat, for example), the Add to chat / Ask in side chat menu no longer survives a keyboard session switch, and references and side chats only apply to the session where the selection was made.
• Upstream errors reported inside the stream (#1600, #1592): when a relay answers 200 but reports api_error / overloaded_error in the SSE body, the request is now retried on the API budget (10 attempts by default, with growing delays) instead of giving up after two retries in two seconds. Errors that a resend cannot fix — context length exceeded, insufficient quota, an invalid prompt — are surfaced immediately.

Other

• Docs (#1599): the bilingual READMEs recommend Zuiquan API Directory.

Installation

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

• macOS: Claude-Code-Haha-0.7.4-mac-arm64.dmg / Claude-Code-Haha-0.7.4-mac-x64.dmg
• Windows: Claude-Code-Haha-0.7.4-win-x64.exe / Claude-Code-Haha-0.7.4-win-arm64.exe
• Linux x64: Claude-Code-Haha-0.7.4-linux-x86_64.AppImage, Claude-Code-Haha-0.7.4-linux-amd64.deb or Claude-Code-Haha-0.7.4-linux-x86_64.rpm
• Linux ARM64: Claude-Code-Haha-0.7.4-linux-arm64.AppImage, Claude-Code-Haha-0.7.4-linux-arm64.deb or Claude-Code-Haha-0.7.4-linux-aarch64.rpm

Full changelog: v0.7.3...v0.7.4

Claude Code Haha v0.7.3

Claude Code Haha v0.7.3

紧急修复

• 修复 Windows 桌面端打开应用、新建或恢复会话时，可能反复弹出黑色终端窗口的问题（#1585）。

Claude Code Haha v0.7.2

Claude Code Haha v0.7.2

本次新增 DeepSeek 账号登录、会话标签数量上限 和 智能体团队提示，并集中修复长会话、上下文压缩、OpenAI Chat 代理和 Windows 平台上的一批问题。

核心功能

• DeepSeek 账号登录（#1548，#1578）：在「设置 → 提供商」中直接用 DeepSeek 账号登录，无需填写 API Key。模型列表按账号实际可用的目录拉取，可直接选择 DeepSeek V4.1 Flash 等模型；已保存的模型映射会自动迁移。
• 会话标签数量上限（#1550）：「设置 → 通用 → 标签页」可限制同时打开的会话标签数（5 / 8 / 12 / 20 或不限），可关闭启动时恢复标签，标签菜单新增「重新打开已关闭的标签」。正在运行、等待输入、有未发送草稿、团队队长和其他标签正在查看的会话不会被自动关闭；全部默认保持现有行为。
• 智能体团队（9fd74c66c）：界面统一为「智能体团队」，在设置或输入框开启前会说明 token 消耗和性能影响，输入框中的提示可选择不再提醒。
• 登录风险确认（#1526）：Claude、ChatGPT、Grok 官方登录前先确认第三方应用登录可能带来的账号风险，按提供商分别说明。
• 复制图片（#1551）：右键菜单可直接复制聊天中的图片。
• 模型更新（#1580，#1546）：新增 Modelflare 预设并刷新 AruHub 模型；第一方 Haiku 默认升级为 Haiku 5.5（1M 上下文），Claude 5 系列继续使用自适应思考。

稳定性与性能

• 长会话的回合改动预览（0894dc813）：不再因为记录过大而丢失每回合的改动卡片，也不再逐文件重复扫描，长会话打开改动预览明显更快。
• 自动压缩空转熔断（#1544）：压缩后上下文仍接近阈值时，连续三次后停止并提示，不再反复压缩直到用户新建会话；压缩确实释放了空间的任务不受影响。
• 缓存共享压缩超出窗口（#1527）：为摘要预留的输出额度不再挤爆上下文窗口，压缩继续走缓存路径，而不是回退重发整段上下文。
• 长工具输入不再被中断（#1536）：写入大文件时，工具输入的等待时间跟随请求超时，不再被固定的 120 秒预算打断。
• 恢复会话的上下文窗口（#1547）：原提供商已被删除或记录有误时，回退到当前或唯一匹配的提供商，不再把 1M 模型显示成「200,000 / 200,000，剩余 0%」。

修复

• 侧边会话恢复父链（#1529）：父会话停在还没有结果的工具调用上时，侧边会话也能正常创建。
• OpenCode Go 模型报 400（eb3d75aa6）：上游拒绝 thinking 字段时自动去掉该字段重发一次，只在重发成功后记住，不影响同一网关的其他模型。
• 图片生成（#1545）：中转地址和原生协议的图片请求都能正常发出。
• Cmd+N 新建会话（#1543）：快捷键恢复为直接新建会话；Windows / Linux 上 Ctrl+N 仍用于输入框换行。
• 路径含空格的输出文件（#1556）：以编码形式出现的链接能定位到磁盘上的真实文件，无法打开时给出提示。
• 分支选择器（#dee09f5ce）：打开时刷新分支列表；后台刷新失败不再清空已有列表、丢掉已选分支或冲掉刚创建的分支。
• Windows 控制台窗口（#1489）：补齐剩余子进程的隐藏窗口设置，打开诊断日志目录时仍能正常显示资源管理器。
• Windows 构建解压 ripgrep（#1491，9988877d1）：没有 unzip 时也能完成解压，不再因为缺少命令直接失败。
• 侧栏项目分组折叠（#1522）：折叠状态在重启后保留。
• Token 数量格式（#1572，#1465）：设置页与聊天头部使用同一个格式，十亿级数量显示为 1.5B 而不是 1500M。
• 已发送消息的标签（#1577）：重新显示团队、技能和命令胶囊。
• 等宽字体下的有序列表（9e1731309）：10. 之类的序号不再被裁掉。
• 选区操作菜单位置（#1541）：新增设置项，可让选中文本后的菜单浮在选择处上方。
• 运行时回归（#1570）：固定 Bun 1.4.3，修复历史加载、输出卡片路径和搜索索引 worker 的一批回归。

其他

• 许可证变更（#1575，#1576）：新增内容改用 cc-haha Community Source License 1.0，个人使用和公司内部使用不受影响；竞争性商业产品、付费二次开发和商业再分发需事先获得许可，再分发需保留来源署名。历史 MIT 授权与第三方声明保持不变。
• 赞助商（#1542，#1532）：新增 Modelflare 白金赞助，退役 ApiSmart 预设（已保存的提供商不受影响）。
• PR 门禁与流量回放（#1567）：CI 质量门禁更可靠，并新增回放录制的 DeepSeek Flash 流量来检查 agent 循环。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

• macOS：Claude-Code-Haha-0.7.2-mac-arm64.dmg / Claude-Code-Haha-0.7.2-mac-x64.dmg
• Windows：Claude-Code-Haha-0.7.2-win-x64.exe / Claude-Code-Haha-0.7.2-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.7.2-linux-x86_64.AppImage、Claude-Code-Haha-0.7.2-linux-amd64.deb 或 Claude-Code-Haha-0.7.2-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.7.2-linux-arm64.AppImage、Claude-Code-Haha-0.7.2-linux-arm64.deb 或 Claude-Code-Haha-0.7.2-linux-aarch64.rpm

完整变更：v0.7.1...v0.7.2

English Version

Claude Code Haha v0.7.2

This release adds DeepSeek account sign-in, session tab limits and an Agent Teams notice, and fixes a batch of issues around long sessions, context compaction, the OpenAI Chat proxy and Windows.

Highlights

• DeepSeek account sign-in (#1548, #1578): sign in with a DeepSeek account from Settings → Providers without an API key. Models are listed from the account's own catalog, including DeepSeek V4.1 Flash, and saved model mappings are migrated automatically.
• Session tab limits (#1550): Settings → General → Tabs limits how many session tabs stay open (5 / 8 / 12 / 20 or no limit), can stop restoring tabs on launch, and adds Reopen Closed Tab to the tab menu. Running chats, chats waiting for input, unsent drafts, team leads and chats viewed by another tab are never auto-closed. Every default keeps today's behavior.
• Agent Teams (9fd74c6): the feature is now named 智能体团队 across the interface, and enabling it from Settings or the composer explains the token cost and slowdown first. The composer notice can be dismissed permanently.
• Login risk confirmation (#1526): Claude, ChatGPT and Grok sign-in now confirm the account-ban risk of signing in through a third-party app, graded per provider.
• Copy Image (#1551): copy images from a chat directly from the context menu.
• Model updates (#1580, #1546): adds the Modelflare preset and refreshes AruHub models. First-party Haiku now defaults to Haiku 5.5 (1M context), and Claude 5 models keep adaptive thinking.

Stability and performance

• Turn change previews on long sessions (0894dc8): per-turn change cards no longer disappear when a transcript is large, and turns no longer rescan every tracked file, so opening a change preview on a long session is much faster.
• Auto-compaction thrashing (#1544): when a compaction leaves the context near the threshold, three consecutive compactions end the turn with a notice instead of compacting until a new conversation is started. Tasks where compaction does free space are unaffected.
• Cache-sharing compaction fork (#1527): the summary output allowance no longer overflows the context window, so compaction stays on the cached path instead of resending the whole context.
• Long tool inputs (#1536): the tool-input budget now follows the request timeout, so large file writes are no longer aborted by a fixed 120-second budget.
• Context window of resumed sessions (#1547): when the recorded provider is gone or wrong, the session falls back to the active or unique matching provider instead of showing "200,000 / 200,000, 0% left" for a 1M model.

Fixes

• Side-chat parent chain (#1529): a side chat can be created when the parent stopped on a tool call that has no result yet.
• OpenCode Go 400s (eb3d75a): when an upstream rejects the thinking field, the request is resent once without it and only remembered after that resend succeeds. Other models on the same gateway keep the toggle.
• Image generation (#1545): relay URLs and native provider protocols both work.
• Cmd+N (#1543): the shortcut creates a new session again; Ctrl+N stays the composer's next-line key on Windows / Linux.
• Output files with spaces in their paths (#1556): percent-encoded links resolve to the real file on disk, and a card that still cannot route its target says so.
• Branch picker (dee09f5): the branch list refreshes when opened, and a failed background refresh no longer blanks the list or drops a just-created branch.
• Windows console windows (#1489): the remaining child-process spawn sites hide their console, while opening the diagnostics log directory still shows Explorer.
• Extracting ripgrep on Windows (#1491, 9988877): builds without unzip on PATH now extract successfully.
• Sidebar project groups (#1522): collapsed groups stay collapsed after a restart.
• Token count formatting (#1572, #1465): settings and the chat header share one formatter, and billions render as 1.5B instead of 1500M.
• Pills on sent messages (#1577): team, skill and command pills show again on messages that were sent.
• Ordered lists in monospace (9e17313): numbers such as 10. are no longer clipped.
• Selection menu placement (#1541): a new setting floats the text-selection menu above the selection.
• Runtime regressions (#1570): pins Bun 1.4.3 and fixes history loading, output card paths and the search index worker.

Other

• License change (#1575, #1576): new content is covered by the cc-haha Community Source License 1.0. Personal and internal business use are unaffected; competing commercial products, paid customization and commercial redistribution require prior permission, and redistributions must keep source attribution. Historical MIT grants and third-party notices are unchanged.
• Sponsors (#1542, #1532): Modelflare joins as a platinum sponsor, and the ApiSmart preset is retired (saved providers keep working).
• PR gate and traffic replay (#1567): the CI quality gate is more reliable and now replays recorded DeepSeek Flash traffic to check the agent loop.

Installation

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

• macOS: Claude-Code-Haha-0.7.2-mac-arm64.dmg / Claude-Code-Haha-0.7.2-mac-x64.dmg
• Windows: Claude-Code-Haha-0.7.2-win-x64.exe / Claude-Code-Haha-0.7.2-win-arm64.exe
• Linux x64: Claude-Code-Haha-0.7.2-linux-x86_64.AppImage, Claude-Code-Haha-0.7.2-linux-amd64.deb or Claude-Code-Haha-0.7.2-linux-x86_64.rpm
• Linux ARM64: Claude-Code-Haha-0.7.2-linux-arm64.AppImage, Claude-Code-Haha-0.7.2-linux-arm64.deb or Claude-Code-Haha-0.7.2-linux-aarch64.rpm

Full changelog: v0.7.1...v0.7.2

Claude Code Haha v0.7.1

Claude Code Haha v0.7.1

本次新增 ChatGPT OAuth Fast 模式、普通 SubAgent 并发上限、语音听写快捷键 和 CSV / TSV 表格预览，让上下文压缩更快、进度更清楚。同时集中修复长任务卡住窗口、Bash 命令被误杀、会话关闭后进程不退出、回复重复显示和 Windows 升级失败等问题。

核心功能

• ChatGPT OAuth Fast 模式（#1502）：使用 ChatGPT OAuth 时，可在模型的推理强度面板中开启 Fast 模式；选中后模型选择器会显示 Fast 标记，重新打开会话也会保留选择。提速效果因模型而异，可能增加额度消耗。
• 普通 SubAgent 并发上限（#1506）：在「设置 → 通用」限制同一会话内普通 SubAgent 的并发数，前台、后台和恢复的子任务共用上限；达到上限时拒绝新启动，已有任务完成后可继续启动。默认不限，留空保存即可恢复不限；设置对运行中的会话生效，无需重启。主 Agent、Agent Teams 和 Workflow 不计入，也不是跨会话的全局配额。
• 语音听写快捷键（#1510）：应用窗口在前台时，按一下开始录音、再按一下结束，也支持按住说话、松开结束。macOS 默认右 Option，Windows / Linux 默认右 Ctrl；可在「设置 → 语音输入」开关或自定义。录音进入当前或最近聚焦的聊天输入框，在其他文本框、终端、弹窗和输入法组词期间不会触发；组合键不会误开麦克风。
• CSV / TSV 表格预览（#1505）：工作区里的 .csv 和 .tsv 可直接按表格查看，并通过「表格 / 源码」切换回原文，继续使用行评论和选区引用。支持 UTF-8、UTF-16 和 GB18030 解码，保留单元格原始值；回合改动中的 CSV 仍可打开差异审查。
• 更快的上下文压缩与实时进度（#1500）：压缩摘要减少可重新读取的文件正文，改用路径和必要信息，缩短摘要生成时间。压缩时显示已用时间和已生成的字符数；重试共用次数预算，避免上游失败时多层重试反复发送整段上下文。
• 自动更新开关（#1482）：在「设置 → 关于」控制是否自动检查和后台下载更新，关闭后仍可手动检查与安装。选择稍后安装后，不再在退出时替你安装。

稳定性与性能

• 长任务流式输出不再卡住整个窗口（#1503，#1480）：减少流式输出时工具列表、会话标签和侧栏的重复计算与渲染，让长会话运行期间的新建会话、切换标签和输入操作更流畅。
• 修复 Bash 命令被误杀（#1493）：修复进程守护管道与 Bun 文件描述符复用冲突，避免健康命令随机被终止、反复出现空的 Exit code 1，以及 Grep 明明有结果却显示「No files found」。
• 减少服务端内存增长与轨迹写入开销（#1493）：搜索索引复用后台 worker，避免每次写入新建线程；轨迹超过索引窗口后不再每追加一行就重扫整个窗口，减少长会话的内存累积和写入延迟。
• 回收关闭会话的 CLI 与 MCP 进程（#1501）：桌面、手机和平板不再长期保留无人查看且已空闲会话的连接；只有后台任务的会话也能按上限收尾。Windows 停止 CLI 时一并清理 MCP 子进程树；IM 聊天在无运行任务、无待审批且双向静默 30 分钟后释放连接，下一条消息仍回到原会话。
• Agent Teams 动画更流畅（#1504，#1492）：修复团队关系图流动连线的卡顿。

修复

• 思考和回复重复显示（#1478，#1476）：冷加载历史与实时输出合并时，同一段思考或回复不再同时显示实时版本和落盘版本；思考中的空白片段也会保留，保证回读一致。
• 完整回复被误判为流中断（#1496，#1481）：部分 Anthropic 兼容网关漏掉最后的 message_stop 帧时，只要已收到终止原因且所有内容块完整关闭，就正常接受回复；真正截断的内容仍会报错并重试。
• 项目路径大小写导致实时记录丢失（#1498）：项目目录仅大小写不同时，仍能找到正在运行的会话 transcript。
• 重启后无法批准 Agent Teams 计划（#1479，#1475）：批准计划时会先启动已断开的队长，启动失败则保留计划供再次审阅，避免只能新建会话重来。
• Windows 升级被安全软件拦截（#1495、#1499，#1486）：安装辅助脚本无法写入或运行时，改用进程检查后备路径；从受同类问题影响的旧版本升级时也能继续安装，同时保留真实写入错误和其他卸载失败的诊断。
• Windows 启动错误更清楚（#1479、#1495）：Bun 在执行代码前因内存提交失败崩溃时，给出系统内存提交上限的提示并保留原始输出；sidecar 无法执行时显示文件路径和重装建议。
• 轨迹视图可用高度（#1507）：打开轨迹视图时隐藏聊天输入框，让轨迹占满可用高度。
• 权限下拉菜单超出窗口（#1477）：设置里的权限菜单会保持在窗口内。
• AruHub 预设端点（#1484）：改为通用 Agent 接口地址。
• Unicode 文件名预览（#1511）：Windows / Linux 上，文件名看起来相同但 Unicode 编码形式不同时，也能找到工作区文件并预览。
• 只读目录的数据迁移（#1511）：修复 macOS / Linux 迁移只读目录时发布失败、回滚后暂存目录无法清理的问题。
• 停止后重新压缩（#1517）：停止任务后再执行 /compact，压缩进度和完成结果能正常恢复显示。
• 工作区文件更新遗漏（#1517）：补偿漏掉的文件监听事件，恢复目录后刷新相关缓存；搜索内容不变时，结果也会随文件更新刷新。
• 并发上限输入校验（#1517）：保留非法输入供保存时校验，避免被输入控件清空后误保存为「不限」。

其他

• PR 质量门禁补齐无真实模型的 Agent 端到端检查，改进测试发现、超时和产物诊断（#1511）。
• README 致谢区补充国内 AtomGit 托管地址。

安装

本次 Windows 应用和安装包未签名。

请在下方 Assets 中下载对应平台和架构的安装包：

• macOS：Claude-Code-Haha-0.7.1-mac-arm64.dmg / Claude-Code-Haha-0.7.1-mac-x64.dmg
• Windows：Claude-Code-Haha-0.7.1-win-x64.exe / Claude-Code-Haha-0.7.1-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.7.1-linux-x86_64.AppImage、Claude-Code-Haha-0.7.1-linux-amd64.deb 或 Claude-Code-Haha-0.7.1-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.7.1-linux-arm64.AppImage、Claude-Code-Haha-0.7.1-linux-arm64.deb 或 Claude-Code-Haha-0.7.1-linux-aarch64.rpm

完整变更：v0.7.0...v0.7.1

English Version

Claude Code Haha v0.7.1

This release adds Fast mode for ChatGPT OAuth, ordinary SubAgent concurrency limits, a dictation keyboard shortcut and CSV / TSV table previews, with faster context compaction and visible progress. It also fixes long runs freezing the window, healthy Bash commands being killed, processes surviving closed sessions, duplicate replies and Windows upgrade failures.

Highlights

• Fast mode for ChatGPT OAuth (#1502): enable Fast mode in the model's reasoning-effort panel when using ChatGPT OAuth. The model selector shows a Fast indicator, and reopening the session preserves the choice. Speed gains depend on the model and may consume more quota.
• Ordinary SubAgent concurrency limits (#1506): Settings → General can limit ordinary subagents within each session. Foreground, background and resumed workers share the limit; new starts are refused at capacity and can proceed after existing work finishes. The default is unlimited, restored by saving an empty value. Changes apply to running sessions without a restart. The main agent, Agent Teams and Workflow are excluded; this is not a global quota across sessions.
• Dictation keyboard shortcut (#1510): with the app in front, tap to start recording and tap again to stop, or hold to talk and release to stop. Defaults to right Option on macOS and right Ctrl on Windows / Linux; configure or disable it in Settings → Voice input. Dictation targets the focused or last-focused composer and does not trigger in other text fields, terminals, dialogs or during IME composition. Key combinations do not accidentally open the microphone.
• CSV / TSV table previews (#1505): open .csv and .tsv files as tables in the workspace, with a Table / Source switch that keeps line comments and selection references available. UTF-8, UTF-16 and GB18030 are supported, and cell values stay as written. Changed CSV files still open in the turn's diff review.
• Faster context compaction with progress (#1500): summaries use file paths and essential information instead of copying file contents that can be read again, reducing generation time. Compaction displays elapsed time and generated characters. A shared retry budget prevents nested retries from repeatedly resending the full context when the upstream fails.
• Automatic update control (#1482): Settings → About controls automatic checks and background downloads. Manual checks and installation remain available when disabled. Deferring installation no longer causes the update to install when you quit.

Stability and performance

• Long streaming runs no longer freeze the whole window (#1503, #1480): reduces repeated computation and rendering in tool lists, session tabs and the sidebar, keeping new-session creation, tab switching and typing responsive during long runs.
• Healthy Bash commands no longer killed (#1493): fixes a conflict between the process guardian pipe and Bun's file-descriptor reuse that caused random command termination, empty Exit code 1 failures and Grep reporting "No files found" despite matches.
• Lower server memory growth and trace-writing overhead (#1493): search indexing reuses a worker instead of creating one per write, and appending past the trace index window no longer rescans the entire window on every line.
• Closed sessions release CLI and MCP processes (#1501): desktop, phone and tablet clients release connections for idle sessions no longer being viewed; background-only sessions also reach their cleanup ceiling. Windows CLI shutdown reaps the MCP process tree. IM chats release their connection after 30 quiet minutes with no active turn or pending approval, and the next message reconnects to the same session.
• Smoother Agent Teams animation (#1504, #1492): fixes stuttering flow lines in the team graph.

Fixes

• Duplicate thinking and replies (#1478, #1476): merging cold history with live output no longer shows both the streamed and durable copies of the same content. Whitespace-only thinking fragments are preserved for consistent replay.
• Complete replies mistaken for interrupted streams (#1496, #1481): an Anthropic-compatible gateway that drops the final message_stop frame can still complete a reply when a terminal stop reason was received and all content blocks are closed. Truly truncated content still fails and retries.
• Project path casing losing the live transcript (#1498): a project directory that differs only in case still resolves the running session's transcript.
• Team plans could not be approved after restart (#1479, #1475): approval starts a disconnected lead first. If startup fails, the plan stays reviewable instead of forcing a new session.
• Windows upgrades blocked by security software (#1495, #1499, #1486): when the installer helper cannot be written or executed, setup uses a fallback process check. Upgrades from older installers affected by the same block can continue, while real write errors and other uninstall failures retain their diagnostics.
• Clearer Windows startup errors (#1479, #1495): Bun crashes caused by memory-commit failure before code runs include guidance about the system commit limit and preserve the original output. An unexecutable sidecar reports its path and reinstall guidance.
• Full-height trajectory view (#1507): hides the chat composer while viewing the trajectory so the view uses all available height.
• Permission menu outside the window (#1477): the settings permission menu stays within the window.
• AruHub preset endpoint (#1484): points the preset to its general agent endpoint.
• Unicode filename previews (#1511): workspace files on Windows / Linux can be found and previewed when visually identical names use different Unicode normalization forms.
• Migrating read-only directories (#1511): fixes publication failures when migrating read-only directories on macOS / Linux and staging directories that could not be cleaned up after rollback.
• Compaction after Stop (#1517): running /compact after stopping a task restores progress and completion updates correctly.
• Missed workspace file updates (#1517): reconciles missed watch events and refreshes affected caches when directories are restored. Search results refresh after file updates even when the query stays unchanged.
• Concurrency limit input validation (#1517): keeps invalid drafts for validation on save instead of letting the input control clear them and accidentally save an unlimited value.

Other

• The PR quality gate gains an end-to-end agent lane witho...

Claude Code Haha v0.7.0

Claude Code Haha v0.7.0

这是一次大版本改版：桌面端整体换上全新的「素 Porcelain」视觉系统，手机和平板 H5 按触屏重做，输入框「+」菜单重新组织为技能 / 连接器 / 运行方式三组，新增「编辑已发送的提示词并重跑」和会话「轨迹」视图。Agent Teams 的可靠性和后台任务的收尾也做了系统性修复——停止后保持停止、成员失败自动唤醒、跨会话的后台任务不再永久转圈。

核心功能

• 全新桌面视觉系统「素 Porcelain」（#1452）：所有页面重做在同一个视觉系统上——浅色主题、工具调用读成时间线、用户消息用气泡、输入框是悬浮卡片；六个主题统一圆角（4/6/8/12/16/20）和字号（11/12/13/14/15/18/22/26）阶梯，字重不超过 semibold；状态色按语义取（运行中 = info、等待 = warning、完成 = success、失败 = error）；图标全部来自 lucide，去掉 Material Symbols 字体。designSystem.test.ts 会拦下越级的字号、过重的字重、斜体、Material 图标和散落的自定义圆角。
• 手机 / 平板 H5 按触屏重做（#1460、#1461）：抽屉式布局换成单层导航栈，首页是按状态分组的会话列表（「等你处理」置顶），系统返回手势回上一层；输入框只留一行（+ / 模型 / 发送），其余收进「+」底部面板；审批、提问和计划直接占用输入框位置，不再弹窗遮挡；消息长按出菜单；每轮工具调用折成一张卡，点开看完整时间线；首页右下角悬浮「新任务」，点开从底部升起全屏卡片，项目在顶部。平板保留左列表右详情。
• 输入框「+」菜单重组（#1462）：一级分为「给这次任务用」（技能 / 连接器 / 文件）和「运行方式」（Agent Teams、操作电脑两个开关），其余进「更多工具」；桌面端二级从右侧展开；连接器按「已连接 / 需要处理 / 推荐」列出并显示各自 logo，不再单列「插件」；技能列出最近用过、全部技能和两个可一键安装的市场热门。从菜单打开「操作电脑」与设置里走同一套授权确认。
• 编辑已发送的提示词并重跑（#1432，#1343）：悬停已发送的提示词出现「编辑并重发」，气泡变成内联编辑器，发送前会先试算回滚、在影响后续轮次或文件时确认，然后原地重跑；回滚失败不改动任何东西并保留草稿。当前无法编辑时（轮次运行中、有后台任务、检查点加载中）按钮保持可见并给出禁用原因，不再直接消失。
• 会话「轨迹」视图（#1443）：会话头部新增聊天 / 轨迹切换。轨迹是每行一个事件的高密度账本（系统提示、用户消息、注入的上下文、助手回复、工具调用），配三轨小地图、按轮折叠、搜索和逐行详情面板，也能钻进子 Agent。设置里原来的 Trace 列表、trace 标签页和独立窗口取消，已保存的 trace 标签会迁移成会话标签。
• 安全的数据目录迁移（#1433、#1473）：在「设置 → 通用」更改数据目录不再直接搬文件，而是走暂存 → 校验 → 迁移凭据 → 重启的流程，写迁移日志以便中断后恢复；同时会先让 IM sidecar 停稳再动盘。Windows 上被系统扫描程序短暂占用的文件会重试发布（最多六次、约 3 秒退避），仍失败则原样回滚。
• 语音听写（#1457、#1427）：录音时工具栏整行变成录音条——取消、麦克风听到内容的滚动痕迹、倒计时时钟（最后 15 秒）、停止和发送；发送会在草稿带上文字后走正常发送路径，被判定为中途改动而扣下的文字不会发送。语音输入默认开启（已保存过 false 的保持不变），Windows 上的识别 worker 也能正常启动。

Agent Teams

• 长任务可恢复（#1426）：成员回合被截断的流会自动重发；服务端托管成员，停下的成员收到消息后从自己的 transcript 重启，瞬时失败按 15s/45s/2m/5m/10m 自动续跑，只有重试耗尽才上报队长；停止现在是暂停整组，队长的下一条用户消息会带上停止成员的清单和未完成任务；队长重启（切模型、崩溃）不连带成员，服务端重启会重新接管团队。邮箱改为原子写入、按身份标记已读，并对齐官方 CLI 2.1.284 的回合间投递语义。
• 停止后保持停止（#1446）：按下停止后队长不再继续轮询邮箱，队员在停止前发出的汇报不会偷偷把团队唤醒，只有用户的下一条消息能解除；因用量超限、计费、网络或崩溃停下的成员，无论是否按过停止，都会在用户下一条消息里逐个列出其失败原因和未完成任务。
• 强制执行任务依赖（#1459）：依赖未完成的成员只在依赖就绪时才收到指令，之前收到的消息在收件箱里等待；TaskUpdate 拒绝队员在依赖未完成时开始或完成自己的任务（队长不受此限）；每个成员会被告知谁在等它的结果，并在完成前把结果发过去。
• 成员任务状态（#1446）：成员停下或失败后，它的任务不再显示「进行中」的动画条；成员抽屉把任务分成「当前 / 下一个 / 已完成」，标出被阻塞任务缺哪个依赖，时长测量不到时不再打印 +0:00。

后台任务与运行时

• 停止与运行时丢失时收敛后台任务（#1441）：按停止会一并收掉本会话的后台 shell 任务（run_in_background、Dream、workflow），不再只断前台轮；没有客户端、只靠后台任务续命的会话有 31 分钟的硬上限，到期停掉运行时并写终态；运行时已经没了但任务记录还在时直接写终态，不再让重连把「运行中」灌回来。
• 桌面没运行过的会话也能收敛（#1471）：别的 CLI 跑完的会话里，后台 shell 任务不再永久「运行中」——任务终态在 transcript 里有四种落盘形态，现在全部读取；没有 CLI 进程时，停止一个只存在于历史里的任务会收敛为已停止而不是报错。

模型与供应商

• OpenAI Chat 代理默认发图（#1458）：不再按主机名和模型名猜图片能力，图片一律先发给模型；上游明确以「不支持图片输入」拒绝时，去掉图片重发一次，只有这次成功才会把该端点和模型记为纯文本 30 分钟。这样 OpenCode Go 上每个新的视觉模型都能直接用，也覆盖了 vLLM 和 OpenAI 的措辞。
• 新会话默认推理强度改为 low（#1439）：服务端和桌面 store 的全局默认都从 max 调到 low，已保存过偏好的人不变；不在 provider 列表里的 ChatGPT Official、Grok Official 也会按模型目录解析强度，不再出现选择器显示全局值、实际跑模型默认值的情况。

修复

• 中文文件名被截断（#1429、#1431，#1423）：正文和内联代码里的中文文件名按整个 token 读取，测试文档1.docx 不再变成 1.docx；文案无法自行界定的名字（报告v2.docx、无空格粘连的名字、全角括号、纯中文加扩展名）交给磁盘确认，优先取最长的已存在文件。
• 不存在的图片报红卡（#1450）：文件路由返回 400/403/404/413 时视为图片本就不存在，内联卡片静默消失、Markdown 图片退回 alt；只有服务出错、连接断开、凭据被拒或字节无法解码才显示可重试的错误。
• 硬编码文案没走翻译（#1450）：把组件里直接写死的界面文字接入 t()（回复里的图片标题、Mermaid 的错误与加载文案、侧栏「未命名」、Computer Use 的状态检查、/context 的 token 数、轨迹视图标签和读屏标签等），补齐五个语种；untranslatedText.test.ts 会扫描所有组件防止回退。
• provider 流中断的报错（#1449）：上游在 message_stop 之前关闭 200 流且所有重试都失败时，错误会指明是上游模型供应商的问题、明确不是上下文超限，并记录这次收到的证据（事件数、最后一个事件、stop_reason、是否缺 message_stop、未闭合块、耗时和重试次数）；桌面端五个语种都有本地化说明。
• 文件卡片和链接乱弹（#1444）：回复里只是「提到」的路径不再当成产出的文件；输出卡片需要本回合确实产出过该文件（或有改动记录），链接要等磁盘确认文件存在且在本回合之后被改过才渲染，没有依据的名字保持为普通代码。
• 会话查找每次全量解析（#1430）：只有一个候选文件时不再解析整个 transcript 去比对内容，多文件时也在第一条会话记录处停止，长会话分页重开的开销不再随文件大小平方增长。
• 工作区浏览器改用 <webview>（#1437）：不再用永远盖在 DOM 之上的原生 WebContentsView，网页放进会话面板内的一层 <webview>，其他页面、菜单和弹窗能正常盖在上面，选择元素的截图不再拍到已关闭的气泡或闪白。
• 虚拟占位符在分数缩放下抖动（#1428）：非整数缩放时消息列表的占位高度稳定，不再跳动。
• 技能市场首页像两页拼接（#1434）：标题、免责声明、分类和搜索与卡片网格回到同一个页面表面，切换首页和详情不再闪色。
• Agent Teams 头部在窄栏被挤扁（#1472）：控件改为换行排布，按钮和时长不再被压成竖排文字。
• 空白会话的侧边聊天（#1455）：还没开始对话的会话不再显示侧边聊天入口，避免每次点击只弹一句「先开始一段对话」。
• 自定义 provider 预设的语言（#1472）：自定义预设不再固定显示英文「Custom」，按当前语言显示并在预填名称里一致。
• 其他：token 用量页按用量、活动、模型排序；token 用量表移到活动热力图下方。

其他

• README 改为中文默认、英文移到 README.en.md，并换上产品宣传片封面。
• 官网首页新增点击播放的宣传片区块。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

SignPath Foundation 尚未接入完成，本次 Windows 应用和安装包暂未签名；macOS 安装包已签名并完成公证。

• macOS：Claude-Code-Haha-0.7.0-mac-arm64.dmg / Claude-Code-Haha-0.7.0-mac-x64.dmg
• Windows：Claude-Code-Haha-0.7.0-win-x64.exe / Claude-Code-Haha-0.7.0-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.7.0-linux-x86_64.AppImage、Claude-Code-Haha-0.7.0-linux-amd64.deb 或 Claude-Code-Haha-0.7.0-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.7.0-linux-arm64.AppImage、Claude-Code-Haha-0.7.0-linux-arm64.deb 或 Claude-Code-Haha-0.7.0-linux-aarch64.rpm

完整变更：v0.6.8...v0.7.0

English Version

Claude Code Haha v0.7.0

This is a major release: the desktop app moves to a new visual system, 「素 Porcelain」; the phone and tablet H5 shell is rebuilt around touch; the composer's + menu is reorganized into skills, connectors and run modes; and "edit and resend a prompt" plus an in-session trajectory view are new. Agent Teams and background tasks also get systematic reliability fixes — a stopped team stays stopped, failed members are woken automatically, and background tasks from other sessions no longer spin forever.

Highlights

• New desktop visual system, 「素 Porcelain」 (#1452): every page is rebuilt on one visual system — a quiet light theme, tool calls read as a timeline, user turns as bubbles, and the composer as a floating card. All six themes share one radius scale (4/6/8/12/16/20) and one type scale (11/12/13/14/15/18/22/26), weights stop at semibold, and status colours carry meaning (running = info, waiting = warning, done = success, failed = error). Every icon comes from lucide and the Material Symbols font is gone. designSystem.test.ts fails on off-scale font sizes, weights above semibold, italics, Material icons and ad-hoc radii.
• Phone and tablet H5 rebuilt around touch (#1460, #1461): the drawer layout becomes a single navigation stack, with the home being a session list grouped by status ("waiting for you" first) and the system back gesture moving up a level. The composer is one row (+, model, send) and everything else moves into a + bottom sheet; approvals, questions and plans take the composer's place instead of covering the chat; messages use a long-press menu; each tool run folds into one card that opens the full timeline; a floating "New task" button on the home opens a full-height sheet with the project at the top. Tablets keep the list beside the page.
• Composer + menu reorganized (#1462): the root splits into "for this task" (skills, connectors, files) and "how it runs" (Agent Teams and Computer Use switches), with the rest under More tools; on desktop the second level opens in a side panel. Connectors list connected services, ones that need sign-in and suggestions, each with its own logo, and the separate Plugins entry is gone. Skills show recently used ones, all skills, and two market skills that install in place. Turning Computer Use on from the menu goes through the same consent dialog as Settings.
• Edit and resend a sent prompt (#1432, #1343): hovering a sent prompt offers "Edit and resend" — the bubble becomes an inline editor, sending dry-runs the rewind, confirms when later turns or restorable files are at stake, and reruns from there; a failed rewind changes nothing and keeps the draft. When an edit cannot start (a turn running, a background task still going, checkpoints loading) the action stays visible, disabled, with the reason instead of disappearing.
• In-session trajectory view (#1443): sessions gain a Chat / Trajectory switch. The trajectory is a dense one-line-per-event ledger (system prompt, user, injected context, assistant responses, tool calls) with a three-lane minimap, turn folding, search and a per-row detail panel, plus subagent drill-in. The Settings trace list, trace tabs and standalone window are removed; saved trace tabs migrate to session tabs.
• Safe data directory migration (#1433, #1473): changing the data directory in Settings → General now stages, verifies, migrates credentials and restarts instead of moving files directly, with a journal for recovery after an interruption, and IM sidecars are quiesced first. On Windows a file briefly held by a system scanner is retried (up to six attempts with about 3 seconds of backoff) and still rolls back unchanged if it keeps failing.
• Voice dictation (#1457, #1427): while recording, the toolbar row becomes a recording bar — cancel, a scrolling trace of what the microphone heard, a countdown clock (last 15 s), stop and send; sending writes the text and then submits through the normal path, and text held back because the draft changed is never sent. Voice input is on by default (a saved false is kept), and the compiled recognizer worker now starts on Windows.

Agent Teams

• Long-running teams recoverable (#1426): a member turn whose stream was truncated resends it; the server supervises members, restarting a stopped member from its own transcript when messaged and continuing transient failures on a 15s/45s/2m/5m/10m backoff, with only exhausted retries reaching the lead. Stop now pauses the whole team, and the lead's next user message lists the stopped members and their open tasks; lead restarts (model switch, crash) keep members and a server restart re-owns the team. The mailbox gets atomic writes, identity-based read marking and the official CLI 2.1.284 between-turn delivery semantics.
• A stopped team stays stopped (#1446): after Stop the lead stops polling its mailbox, reports sent before the Stop no longer wake the team, and only the use...

Claude Code Haha v0.6.8

Claude Code Haha v0.6.8

本次新增 本地语音听写、工作区文档预览 和 精选技能目录，加入 等待处理的会话高亮 和 按格式的文件图标；同时修复了「停止并关闭」没停干净、请求过大后每条消息都失败、后台任务回复被吞、Grok 报 426 等一批稳定性问题。

核心功能

• 本地语音听写：输入框旁新增麦克风按钮，点击录音、再点结束。音频重采样为 16 kHz 单声道 PCM16 WAV 后交给本机识别进程转写，文字进入输入框草稿、不会自动发送；草稿已被改动或正在输入中文时，结果会挂在「插入」按钮后面等你决定。「设置 → 语音输入」可开关、选语言和麦克风，并试听查看波形。识别引擎和模型运行时下载到本地（校验 sha256/sha512，支持断点续传、中断自动重试），下载源可选官方或国内镜像，取消后已下载的部分会保留。仅 Electron 桌面端，Windows / Linux 与签名公证包尚未实机验证。
• 工作区文档预览：Agent 写出的 PDF、Word、Excel 直接在工作区侧边打开，不再跳到别的应用。支持缩放与适应窗口、按文件记住滚动位置、缩放和当前工作表，文件被改写后自动刷新；聊天里 Agent 提到的本地图片（工作目录之外、~、C:\、file:// 路径）直接显示，Read 等工具返回的图片以缩略图出现在调用下方。工作区之外的文档仍交给系统程序打开。@olo12119（#1397）
• 精选技能目录：技能市场默认打开内置精选目录（398 个 ClawHub / SkillHub 技能，13 个分类），不再每次实时请求两个注册表；「搜索全部市场」保留为显式范围，结果会标注「未精选」。详情页补齐统计、安全报告、更新日志、能力与触发条件；未审计或被标记有风险的技能需要确认后才能安装。ClawHub 的技能固定按卡片上的作者安装，避免同名技能被别人顶替。
• 等待处理的会话高亮：会话产出需要你决策的卡片（AskUserQuestion、权限请求）时，顶部选项卡和侧边栏会显示醒目标记，不用逐个点开找。@kizemo（#1391）
• 按格式显示文件图标：生成的文件卡片、回合改动卡片和消息附件按扩展名显示带颜色和角标的文件图标（PDF 红、Word 与 Markdown 蓝、Excel 绿、PowerPoint 橙、压缩包 琥珀、代码 紫），不再一律灰色。

模型与供应商

• Claude Sonnet 5.5：新增 claude-sonnet-5-5（1M 上下文，low 到 max 推理强度），并刷新模型清单与提示词中的「最新模型」信息（知识截止日期对齐 2026 年 6 月）。
• GPT-6.1-Sol：桌面端与 Codex OAuth 都支持 gpt-6.1-sol（258,400 上下文，low 到 max，默认 low）。
• Fable 模型：配置好的 Fable 模型现在会出现在模型选择器里。
• Grok 报 HTTP 426：CLI 代理身份对齐 1.0.46 之后不再被要求升级。

修复

• 「停止并关闭」没停干净：后台 shell 任务和它的 CLI 会话继续在跑，重开会话还在转圈、再关又弹「会话运行中」。现在关闭前会一并停掉所有运行中的后台任务，「是否运行中」和「停哪些任务」共用同一套判定。@zyr017（#1398）
• 请求过大后每条消息都失败：413 报错改为报出实际发送的体积、其中图片和文档的占比，并指明是供应商还是中转站拒绝；出错后自动把这次请求携带的图片和文档（包括工具结果里嵌套的、@ 附件）替换为占位符，下一条消息即可继续，压缩逻辑复用同一套处理。@hemppp（#1399）
• 请求过大文案误导：不再归咎于你选的模型，也不再让你手动删大文件。
• 后台任务完成后的回复不显示：模型在后台命令结束后给出的回复、以及其后的工具调用，实时和回读都不再被吞掉。@CyL-Cly（#1389）
• 重载会话后整个窗口变成错误页：AskUserQuestion 的选项描述被服务端判为非法时，每次启动都会复现；现在只渲染能显示的字段，每条记录有独立的错误边界。@kizemo（#1400）
• 侧栏冷启动变慢、会话列表退回全量扫描：单条过大的 transcript（超过 8 MiB）不再让整个索引降级，会话列表继续走本地索引，不再每次冷启动都全量扫描历史文件。
• 工作区打开文件与图片加载：通过软链接路径写出的文件现在能识别为工作区文件在工作区内打开；网页端和 H5 的本地图片加载失败会自动改走带凭据的通道重试一次（此前被浏览器直接拦截、显示「无法加载图片」）；表格预览的数字列不再被截断成 12,000...。
• 定时任务全部 500：scheduled_tasks.json 带 UTF-8 BOM（PowerShell 5.x 写出）或写入中断留下空文件时，列表和创建不再报错；有内容但解析失败仍然报错，且永远不会被覆盖。
• PDF 读取状态：先读后写 PDF 的中间状态不再丢失。
• 权限模式：切换权限模式改用实时模块绑定，不再读到过期值。
• @ 提及：内置技能（如 imagegen）现在也会出现在 @ 列表里，并按当前会话的 provider 运行时环境判断是否可用。@NanmiCoder（#1420）
• 其他：Markdown 图片路径支持 ~/ 相对父目录；超大会话保留工作区改动记录；Markdown 图片加载失败给出可重试的错误提示；预览的「适应」按钮改为独立图标且已适应时禁用，不再是点了没反应的常亮按钮。

其他

• 技能市场文档与截图更新。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

SignPath 尚未开通可用，本次 Windows 应用和安装包暂未签名；macOS 安装包已签名并完成公证。

• macOS：Claude-Code-Haha-0.6.8-mac-arm64.dmg / Claude-Code-Haha-0.6.8-mac-x64.dmg
• Windows：Claude-Code-Haha-0.6.8-win-x64.exe / Claude-Code-Haha-0.6.8-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.6.8-linux-x86_64.AppImage、Claude-Code-Haha-0.6.8-linux-amd64.deb 或 Claude-Code-Haha-0.6.8-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.6.8-linux-arm64.AppImage、Claude-Code-Haha-0.6.8-linux-arm64.deb 或 Claude-Code-Haha-0.6.8-linux-aarch64.rpm

完整变更：v0.6.7...v0.6.8

English Version

Claude Code Haha v0.6.8

This release adds local voice dictation, document preview in the workspace and a curated skills catalog, plus highlights for sessions waiting on you and per-format file icons. It also fixes "Stop & Close" leaving background tasks alive, every later message failing after a request-too-large rejection, replies to background task notifications being swallowed, and Grok rejecting the CLI with HTTP 426.

Highlights

• Local voice dictation: a microphone button sits beside the composer — click to record, click again to stop. The audio is resampled to 16 kHz mono PCM16 WAV and transcribed by a local worker; the text lands in the draft and is not sent. If the draft changed or an IME is composing, the result waits behind an insert button. Settings → Voice input controls the switch, language and microphone, and includes a test transcription with a live waveform. The engine and model are downloaded at runtime with pinned sha256/sha512 checks, resumable transfers and retry after an interruption; you can pin the download source to the official host or a domestic mirror, and a cancelled download keeps what was already fetched. Electron desktop only; not yet verified on Windows or Linux, with a real microphone, or in a signed and notarized build.
• Document preview in the workspace: PDF, Word and Excel files the agent writes open in the workspace panel instead of another application, with zoom and fit, per-file memory of scroll position, zoom and sheet, and a refresh when the file is rewritten. Local images the agent mentions (outside the working directory, ~, C:\, file:// paths) render in the conversation, and images returned by tools such as Read appear as thumbnails under the call. Documents outside the workspace still open in the system application. @olo12119 (#1397)
• Curated skills catalog: the skills market opens on a bundled catalog of 398 curated ClawHub and SkillHub skills in 13 categories instead of querying both registries live. "Search all markets" stays available as an explicit scope and its results are marked as not curated. The detail page gains stats, a security report, a changelog, a capability panel and triggers, and installing an unaudited or flagged skill requires an acknowledgement. ClawHub skills are pinned to the card's owner so a same-slug copy by another author is never shown or installed in its place.
• Highlighting sessions waiting on you: when a session produces a card that needs your decision (AskUserQuestion, a permission request), its tab and sidebar row are marked so you do not have to open each one to find it. @kizemo (#1391)
• Per-format file icons: generated-file cards, the turn change card and message attachments render a colored, labelled file icon by extension (PDF red, Word and Markdown blue, Excel green, PowerPoint orange, archives amber, code purple) instead of one grey glyph.

Models and providers

• Claude Sonnet 5.5: adds claude-sonnet-5-5 (1M context, low through max reasoning effort) and refreshes the model catalog and the "latest models" line of the environment prompt (knowledge cutoff aligned to June 2026).
• GPT-6.1-Sol: gpt-6.1-sol is supported in the desktop app and through Codex OAuth (258,400 context, low through max, low by default).
• Fable models: configured Fable models now appear in the model selector.
• Grok HTTP 426: aligning the CLI proxy identity with 1.0.46 clears the upgrade demand.

Fixes

• "Stop & Close" left tasks running: a background shell command and its CLI session kept running, so the reopened session spun again and closing asked the same question once more. Stop & Close now stops every running background task before the socket closes, and "is the session running" and "which tasks to stop" share one predicate. @zyr017 (#1398)
• Every later message failed after a request-too-large rejection: the error now reports the size actually sent, how much of it is images or documents, and whether the provider or the relay rejected it; afterwards the images and documents that request carried (including media nested in tool results and @-mentioned attachments) are replaced with placeholders so the next message goes through. Compaction shares the same placeholder logic. @hemppp (#1399)
• Misleading copy for oversized requests: the message no longer blames the selected model or tells you to delete large files by hand.
• Replies to background task notifications were hidden: what the model said once a background command finished, and any tool work after it, now reaches the chat both live and on reload. @CyL-Cly (#1389)
• The whole window became the error page on session reload: an AskUserQuestion option description the server rejected as invalid reproduced on every launch; the card now renders only the fields it can show, and each transcript row has its own error boundary. @kizemo (#1400)
• Slower cold start and session lists falling back to full scans: one oversized transcript (over 8 MiB) no longer degrades the whole index, so session lists keep being served from the local index instead of scanning history files on every cold start.
• Opening workspace files and loading images: a file written through a symlinked path is now recognized as a workspace file and opens in the workspace; when a local image fails to load in the web UI or H5, it is retried once through the credentialed channel (the browser used to block it and the chat showed "unable to load image"); numbers in spreadsheet previews are no longer clipped to 12,000....
• Scheduled tasks returning 500: a scheduled_tasks.json written with a UTF-8 BOM (PowerShell 5.x) or left blank by an interrupted write no longer breaks listing and creation; content that has data but does not parse still throws and is never overwritten.
• PDF read state: the read-before-write state of a PDF is no longer lost.
• Permission modes: switching a permission mode now uses the live module binding instead of a stale value.
• @ mentions: bundled skills such as imagegen now show up in the @ list, evaluated against the current session's provider runtime environment. @NanmiCoder (#1420)
• Also: Markdown image paths resolve ~/-relative parents, oversized conversations keep their workspace change record, a failed Markdown image shows a retryable error, and the preview's fit control gets its own icon and is disabled while already fitted instead of staying lit and doing nothing.

Other

• The skills market documentation and screenshots were refreshed.

Installation

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

Windows applications and ins...

Claude Code Haha v0.6.7

Claude Code Haha v0.6.7

本次新增 工作区侧边对话 和 Agent Teams 启动前人工审核，补上 超时自动回答 和 聊天外观设置；同时修复了超大会话打不开、并发写丢配置、IM 重连串会话等一批稳定性问题。

核心功能

• 侧边对话：在输入框敲 /btw 问题（或选中文字后点「在侧边聊天中提问」），会在工作区右侧开一个临时对话，继承父会话的上下文和模型但独立运行，追问不会打断主任务。关闭标签或退出应用即清空。
• Agent Teams 启动前人工审核：团队方案就绪后先出「团队执行前确认」卡片，逐个成员改 Agent 预设、服务商、模型、任务归属与前置任务，可批量设置或「跟随主 Agent」，确认后才真正启动成员。启动前的成本由你把关。
• 超时自动回答：「设置 → 通用 → 自动回答问题」开启后，待回答的问题超过设定时长（1/5/10/30 分钟，默认 5）会由该会话的小模型基于对话上下文自行决策；判断不了就继续等，你任何点击或输入都会取消计时。
• 聊天外观：「设置 → 通用 → 聊天外观」可调聊天字体（跟随系统/无衬线/衬线/等宽）、字号（12–24px）和对话宽度（标准/宽/铺满），带预览、独立于界面缩放、按设备保存。

模型与供应商

• Claude Opus 5.5：新增 claude-opus-5-5（1M 上下文，支持 low 到 max 推理强度），并设为官方账号的默认模型；Claude OAuth 兼容版本对齐 2.1.281。
• OpenAI 官方账号：新增 GPT-6 Astra、GPT-6-Sol、GPT-6-Luna，默认模型改为 GPT-6-Sol。
• OAuth 模型映射：「设置 → 服务商」里每个官方账号新增「OAuth 模型映射」，可为 main/sonnet/haiku/opus/fable 五个槽位指定该账号实际可用的模型或手填别名。
• 直连 Anthropic 供应商输出预算：不再把你配置的输出上限覆盖成固定值 @omazili-guga（#1392）。
• Grok Fable 槽位：Fable 槽位补齐到模型映射体系，cc-switch 导入和运行时环境都能正确写入 @MrWangBigData（#1369）。

修复

• #1373 超大会话打不开：对话记录超过 20MB 后报「提示词过长」且无法继续的问题修复。压缩时按供应商报错里的真实 token 上限计算丢弃量并为回复预留空间；超大消息不再让整条元数据被跳过，会话标题、工作目录、模型和权限模式保持完整。
• IM 临时断线后串成新会话：重连失败时不再悄悄换成新会话导致上下文丢失、项目目录回默认，改为保留原会话绑定并提示稍后重试 @yuehua-meng（#1381）。
• 定时任务并发覆盖：并发执行时不再互相覆盖执行时间，已删除的任务不会「复活」 @yuehua-meng（#1371）。
• 记忆编辑保存冲突：保存期间继续输入，编辑器不再退出编辑态、隐藏未保存内容，也不会被自己上一次的版本号挡住 @yuehua-meng（#1372）。
• 团队配置读到半截 JSON 会让团队在界面里消失：所有写路径纳入文件锁并改为原子写 @yuehua-meng（#1387）。
• 思考型模型自动命名静默失败：后台命名请求的 token 预算不再被思考过程耗尽，会话标题不再停留在首条消息的占位文本 @realLoganLuo（#1340 / #1382）。
• Windows 上垂直滚动方向相反 @MikhailTail（#1378）。
• CORS 预检导致 Failed to fetch：本机来源的 OPTIONS 预检提前放行，远程来源仍严格拒绝 @MrWangBigData（#1369）。
• 工具参数损坏后卡住：模型输出了非法 JSON 工具参数时不再静默退化成空参数执行，改为把明确错误回灌给模型让它自己重试。
• 会话模型选择丢失：重开会话后模型不再回退默认；删除供应商后旧模型和推理强度不再串到新供应商。
• 提示词重复发送：生成中追加引导内容后不再看起来发了两遍。
• 切换标签 / 停止生成后聊天内容消失：从终端或工作台切回会话保留聊天视图和滚动位置；停止生成后历史重载不再把已收到的消息删掉或重复。
• 工作区侧边对话无法滚动、工作区文件树操作和预览控制恢复、拖拽标签时的滚动区域判定修正、标签页之间来回切换保留聊天。
• 图片转发：OpenCode Go 下的 Space Bunny 形象图和 Kimi K3 图片不再丢失。
• 本地 Markdown 链接点击后仍指向原目标；生成的输出文件按声明的项目根目录解析路径。
• 其他：活动档案名过长时自动换行；/clear 恢复为可搜索；检查点预览不再占用聊天里的预算显示；自动回答改为纯模型决策；回退后的回合审查状态正确清理；会话发现有边界且会报告历史分页上限。

其他

• Remotion 技能包：扩展页新增 Remotion 技能包（官方 12 项技能，不含运行库和渲染器），用于用 React 制作程序化视频。
• 站点与文档改版：官网落地页整体重做，新增四步工作流与实战案例；文档站新增 5 篇实战文档（读懂陌生项目、修 Bug、实现功能、每日代码审查、手机接力）；首次访问按浏览器语言选择站点语言。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

SignPath 尚未开通可用，本次 Windows 应用和安装包暂未签名；macOS 安装包已签名并完成公证。

• macOS：Claude-Code-Haha-0.6.7-mac-arm64.dmg / Claude-Code-Haha-0.6.7-mac-x64.dmg
• Windows：Claude-Code-Haha-0.6.7-win-x64.exe / Claude-Code-Haha-0.6.7-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.6.7-linux-x86_64.AppImage、Claude-Code-Haha-0.6.7-linux-amd64.deb 或 Claude-Code-Haha-0.6.7-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.6.7-linux-arm64.AppImage、Claude-Code-Haha-0.6.7-linux-arm64.deb 或 Claude-Code-Haha-0.6.7-linux-aarch64.rpm

完整变更：v0.6.6...v0.6.7

English Version

Claude Code Haha v0.6.7

This release adds contextual side chats in the workspace and human review before an Agent Teams run, plus auto-answer after a timeout and chat appearance settings. It also fixes oversized conversations that would not open, configuration lost to concurrent writes, and IM reconnects that silently started a new session.

Highlights

• Side chats: type /btw question in the composer (or select text and pick "Ask in a side chat") to open a temporary conversation on the right of the workspace. It inherits the parent session's context and model but runs independently, so follow-up questions never interrupt the main task. Closing the tab or quitting the app clears it.
• Human review before an Agent Teams run: once a team plan is ready, a "Confirm before the team runs" card lets you change each member's agent preset, provider, model, task ownership and dependencies, with batch assignment and "inherit from lead". Members only start after you confirm, so cost is under your control before launch.
• Auto-answer after a timeout: enable it under Settings → General → Auto-answer questions and a pending question that goes unanswered past your chosen delay (1/5/10/30 minutes, 5 by default) is answered by a small model on that session using the conversation context. If it cannot tell, it keeps waiting; any click or input from you cancels the timer.
• Chat appearance: Settings → General → Chat appearance controls chat font (system / sans-serif / serif / monospace), font size (12–24px) and conversation width (standard / wide / full), with a preview. It applies independently of the interface zoom scale and is saved per device.

Models and providers

• Claude Opus 5.5: adds claude-opus-5-5 (1M context, low through max reasoning effort) and makes it the default for official accounts; Claude OAuth compatibility is aligned with 2.1.281.
• OpenAI official accounts: adds GPT-6 Astra, GPT-6-Sol and GPT-6-Luna, with GPT-6-Sol as the new default.
• OAuth model mapping: every official account card in Settings → Providers gains an "OAuth model mapping" panel where the main / sonnet / haiku / opus / fable slots can be pointed at a model that account actually has, or given a hand-written alias.
• Output budget for direct Anthropic providers: your configured output ceiling is no longer overwritten with a fixed value @omazili-guga (#1392).
• Grok Fable slot: the Fable slot is wired into the model mapping system, including cc-switch import and the runtime environment @MrWangBigData (#1369).

Fixes

• #1373 Oversized conversations would not open: conversations whose history exceeds 20MB no longer fail with "prompt too long". Compaction now computes the discard amount from the real token limit reported by the provider and reserves room for the reply; oversized messages no longer cause the whole metadata record to be skipped, so the title, working directory, model and permission mode stay intact.
• A temporary IM disconnect started a new session: a failed reconnect no longer silently swaps in a fresh session and loses context or resets the project directory; the original binding is kept and you are asked to retry @yuehua-meng (#1381).
• Concurrent scheduled-task writes: concurrent runs no longer overwrite each other's schedule, and deleted tasks no longer come back @yuehua-meng (#1371).
• Memory edit save conflicts: typing while a save is in flight no longer drops you out of edit mode, hides unsaved text, or gets blocked by your own previous revision @yuehua-meng (#1372).
• Reading half-written team config made teams vanish from the UI: every write path now takes the file lock and writes atomically @yuehua-meng (#1387).
• Silent failure of auto-naming on thinking models: the background title request no longer spends its whole token budget on thinking and leaves the title as the first message's placeholder @realLoganLuo (#1340 / #1382).
• Vertical scrolling ran the wrong way on Windows @MikhailTail (#1378).
• CORS preflight caused Failed to fetch: OPTIONS preflights from local origins are answered before the H5 gate, while remote origins are still rejected @MrWangBigData (#1369).
• Malformed tool arguments could stall the run: when a model emits invalid JSON for a tool call, the arguments are no longer silently degraded to {} and executed; the explicit error is fed back so the model can retry.
• Lost session model selection: reopening a session no longer resets the model to the default, and removing a provider no longer leaks its old model and effort onto a different provider.
• Duplicate prompt: appending a steering message mid-generation no longer makes it look like the same prompt was sent twice.
• Chat content vanishing after a tab switch or a stop: switching to the terminal or workbench and back keeps the chat view and scroll position; reloading history after stopping no longer deletes or duplicates messages that already arrived.
• Side chats would not scroll, workspace file tree actions and independent preview controls are restored, tab drag exclusions are confined to the viewport, and switching between tabs preserves the chat.
• Image forwarding: Space Bunny images and Kimi K3 images survive the OpenCode Go path.
• Local Markdown links keep their target, and generated output files resolve against the declared project roots.
• Also: long activity profile names wrap, /clear is searchable again, the checkpoint preview no longer shows the budget, auto-answering is now purely a model decision, review state after a rewind is cleared, and session discovery is bounded and reports history page limits.

Other

• Remotion skill pack: the extensions page gains a Remotion skill pack (the official 12 skills, without the runtime or renderer) for building procedural video with React.
• Site and documentation redesign: the landing page is rebuilt with a four-step workflow and real usage cases; the docs site gains five new walkthroughs (reading an unfamiliar project, fixing a bug, shipping a feature, daily code review, handing off from a phone); first-time visitors get the site language matching their browser.

Installation

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

Windows applications and installers in this release are unsigned because SignPath is not yet available. macOS builds are signed and notarized.

• macOS: Claude-Code-Haha-0.6.7-mac-arm64.dmg / Claude-Code-Haha-0.6.7-mac-x64.dmg
• Windows: Claude-Code-Haha-0.6.7-win-x64.exe / Claude-Code-Haha-0.6.7-win-arm64.exe
• Linux x64: Claude-Code-Haha-0.6.7-linux-x86_64.AppImage, Claude-Code-Haha-0.6.7-linux-amd64.deb or `Claude-Code-Haha-0.6.7-linu...

Claude Code Haha v0.6.6

Claude Code Haha v0.6.6

本次新增 引用历史会话 和 跨会话协作（让 Claude 自己开独立会话并行干活），并对会话加载、历史滚动和搜索做了一轮性能改造。

核心功能

• 引用之前的对话：输入 @ 可以搜索并引用历史会话，例如「参考 @登录方案 的结论，补充退出登录流程」。引用不会把整段历史复制进上下文，Claude 按需读取被引用会话的内容；点击消息里的引用可直接打开那段对话，原会话不会被唤醒。
• 跨会话聊天与控制：可以让 Claude 新建独立会话并行推进任务，并在会话之间收发消息、等待状态变化、汇报结果（ListSessions / ReadSession / CreateSession / SendSessionMessage / WaitSessions）。
  • 新会话会出现在会话列表里，可以单独打开继续聊；其他会话发来的内容以带来源标注的普通消息显示，工具卡片上可以一键打开相关会话。
  • 每组协作最多同时运行 3 个工作会话，其余排队；Git 项目里的工作会话使用独立 worktree。

性能

• 大会话加载与滚动重做：历史改为一次字节上限内整体加载、整体挂载，不再分页拼接，消除滚动时的跳动、闪烁和丢行。
• @ 搜索不再卡住其他请求：提及搜索和会话清理由请求线程挪走，中文 @ 搜索不再让其他接口排队；会话建议改为直接查本地索引。
• 会话索引状态稳定：索引重建期间不再在本地索引和全量 JSONL 扫描之间来回切换，侧边栏不再反复闪「使用标准历史」。
• 引用会话的读取有界：读取被引用会话不再把会话列表轮询拖到超时报错。
• 新增接口耗时监控：慢请求和大响应会记入诊断面板，卡顿可以直接定位到具体调用。

修复

• #1322 文件树的「添加到对话」和右键键盘菜单恢复可用。
• #1327 OpenAI 兼容供应商的嵌套缓存命中量不再被上报为 0。
• #1335 electron:dev 下渲染进程的 CORS 预检不再被拒绝。
• #1339 已清除的 goal 不会被旧 transcript 重新激活。
• #1345 / #1352 后台任务的完成通知会立即推送，任务不再一直显示「运行中」（这两个 issue 中与后台任务无关的部分不在本次修复内）。
• #1346 「打开方式」新增 Zed。
• #1349 auto-dream 中断后可以自动恢复，不再推迟 24 小时。

其他

• AGENTS.md 项目指令：默认优先读 CLAUDE.md，缺失时回退到 AGENTS.md，可以在 /config 的 Project instructions 里切换。
• 供应商预设可以切换协议：双协议端点不再被预设地址锁死，协议与地址不匹配时会提示。
• Grok 4.7：支持 grok-4.7 及快速档，推理强度按官方在线目录解析，不再被静默降级。
• SubAgent / 团队成员的运行卡片会显示实际使用的模型。
• 上下文用量环按已用比例填充（此前一直是空环）；「关于」面板新增企业微信群二维码；新增模型时不再把当前会话的模型写进该供应商的配置。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

SignPath 官方审批尚未完成，本次 Windows 应用和安装包暂未签名；macOS 安装包已签名并完成公证。

• macOS：Claude-Code-Haha-0.6.6-mac-arm64.dmg / Claude-Code-Haha-0.6.6-mac-x64.dmg
• Windows：Claude-Code-Haha-0.6.6-win-x64.exe / Claude-Code-Haha-0.6.6-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.6.6-linux-x86_64.AppImage、Claude-Code-Haha-0.6.6-linux-amd64.deb 或 Claude-Code-Haha-0.6.6-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.6.6-linux-arm64.AppImage、Claude-Code-Haha-0.6.6-linux-arm64.deb 或 Claude-Code-Haha-0.6.6-linux-aarch64.rpm

完整变更：v0.6.5...v0.6.6

English Version

Claude Code Haha v0.6.6

This release adds referencing past conversations and cross-session collaboration (Claude can spawn independent sessions and work in parallel), plus a performance overhaul of session loading, history scrolling and search.

Highlights

• Reference a past conversation: type @ to search and reference an earlier session, e.g. "follow the conclusion in @login-plan and add the logout flow". The reference does not copy the whole history into context — Claude reads the referenced session on demand — and clicking it opens that conversation without waking it up.
• Cross-session chat and control: Claude can create independent sessions to work in parallel, then message them, wait for state changes and report back (ListSessions / ReadSession / CreateSession / SendSessionMessage / WaitSessions).
  • New sessions show up in the session list and can be opened and continued on their own; deliveries from other sessions appear as ordinary messages labeled with their source, and tool cards link to the sessions they touched.
  • Each collaboration runs at most 3 working sessions at a time; the rest queue. Working sessions in a Git project use an independent worktree.

Performance

• Large-session loading and scrolling redone: history loads in one byte-bounded response and stays mounted as a single array instead of being stitched from pages, so scrolling no longer jumps, flickers or drops rows.
• @ search no longer blocks other requests: mention search and session clearing moved off the request thread, so a CJK @ query no longer queues every other API call, and session suggestions are read straight from the local index.
• Stable index state: while the index reconciles it no longer flips between the local index and a full JSONL scan, so the sidebar stops flashing "using standard history".
• Bounded reads for referenced sessions: reading a referenced session no longer starves the session list poll into a timeout.
• New API timing monitor: slow requests and large responses are recorded in the diagnostics panel, so stalls can be traced to a specific call.

Fixes

• #1322 Restores file-tree "Add to chat" and the keyboard context menu.
• #1327 Nested OpenAI cache hits are no longer reported as 0.
• #1335 CORS preflights from the renderer under electron:dev are no longer rejected.
• #1339 A cleared goal is no longer restored from a stale transcript.
• #1345 / #1352 Background task completion is published immediately instead of being dropped, so tasks stop showing as running (the parts of these issues unrelated to background tasks are not covered by this fix).
• #1346 Adds Zed to the "open with" menu.
• #1349 An interrupted auto-dream recovers instead of being postponed by 24 hours.

Other

• AGENTS.md project instructions: CLAUDE.md is still preferred, falling back to AGENTS.md; switchable under Project instructions in /config.
• Provider presets can switch protocol: a dual-protocol endpoint is no longer locked to the preset URL, and a mismatch is called out.
• Grok 4.7: supports grok-4.7 and its fast variant, with reasoning effort resolved against the live catalog instead of being silently dropped.
• SubAgent and team member run cards show the model that actually served the run.
• The context meter fills by used share (it used to render an empty track); the About panel gains the WeCom group QR; adding a model no longer writes the current session model into that provider's settings.

Installation

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

Windows applications and installers in this release are unsigned while SignPath approval is pending. macOS builds are signed and notarized.

• macOS: Claude-Code-Haha-0.6.6-mac-arm64.dmg / Claude-Code-Haha-0.6.6-mac-x64.dmg
• Windows: Claude-Code-Haha-0.6.6-win-x64.exe / Claude-Code-Haha-0.6.6-win-arm64.exe
• Linux x64: Claude-Code-Haha-0.6.6-linux-x86_64.AppImage, Claude-Code-Haha-0.6.6-linux-amd64.deb or Claude-Code-Haha-0.6.6-linux-x86_64.rpm
• Linux ARM64: Claude-Code-Haha-0.6.6-linux-arm64.AppImage, Claude-Code-Haha-0.6.6-linux-arm64.deb or Claude-Code-Haha-0.6.6-linux-aarch64.rpm

Full changelog: v0.6.5...v0.6.6

Claude Code Haha v0.6.5

Claude Code Haha v0.6.5

本次新增 OpenCode Go 订阅供应商、可搜索的 Composer 能力菜单、批准计划时切换执行模型，并修复了大文件会话打不开等一系列性能问题。

新功能

• OpenCode Go 订阅供应商：填入订阅的 API Key，即可使用 GLM、Kimi、MiniMax、Qwen、Grok、GPT 等模型。不同模型的协议差异由本地代理自动分流，无需手动配置。
• Composer 能力菜单：输入框的「+」菜单现在可搜索，Skills、连接器、Agents、工作流、Agent Teams 和 Computer Use 都从这里进入。
• 批准计划时切换执行模型：在执行计划确认条上直接选模型——同供应商立即切换，跨供应商自动重启并继续执行，不用重新描述需求。
• 上下文用量面板重做：以剩余窗口为主，速度、缓存命中率、成本并排展示，不再只给一份分类占比。
• 项目选择器：Composer 的目录选择支持搜索全部项目，并可直接新建项目（使用系统文件夹对话框）。
• Claude Opus 5：新增 Opus 5 作为官方默认模型。

性能与稳定性

• 大会话可以正常打开了：subagent 记录不再内联进会话历史，之前 542 MB 的响应降到 40 MB。部分会话此前会直接报 Unexpected end of JSON input。
• 历史与 Trace 走本地索引：打开或轮询 Trace 不再整份读取会话 JSONL，长会话的历史滚动恢复连续，搜索索引不再无上限增长。
• 修复大记录和部分 transcript 无法查看的问题。

修复

• 模型选择：会话侧栏刷新后不再把当前模型回退成旧值。
• AskUserQuestion：切换标签或滚动不再丢失已填写的答案；等待回答期间输入框会锁定，避免消息被误认为已经回复。
• Token 统计：改为新输入 + 输出，不再把缓存命中算成消耗（此前 270k 窗口的 DeepSeek 会话会被显示成 4000 万 tokens）。
• 新建会话目录：从项目根目录创建，不再落到过期的 worktree 目录（曾导致 Branch "main" is already checked out）。
• 任务停止：已结束的任务再点停止不再报错，Activity 面板不会一直停在「运行中」。
• 连接器安装：修复 enabledPlugins 残留记录导致后续所有 Skill 安装失败的问题。
• Composer 中的超长连续字符串不再撑破面板宽度。

安装

请在下方 Assets 中下载对应平台和架构的安装包：

SignPath 官方审批尚未完成，本次 Windows 应用和安装包暂未签名；macOS 安装包已签名并完成公证。

• macOS：Claude-Code-Haha-0.6.5-mac-arm64.dmg / Claude-Code-Haha-0.6.5-mac-x64.dmg
• Windows：Claude-Code-Haha-0.6.5-win-x64.exe / Claude-Code-Haha-0.6.5-win-arm64.exe
• Linux x64：Claude-Code-Haha-0.6.5-linux-x86_64.AppImage、Claude-Code-Haha-0.6.5-linux-amd64.deb 或 Claude-Code-Haha-0.6.5-linux-x86_64.rpm
• Linux ARM64：Claude-Code-Haha-0.6.5-linux-arm64.AppImage、Claude-Code-Haha-0.6.5-linux-arm64.deb 或 Claude-Code-Haha-0.6.5-linux-aarch64.rpm

完整变更：v0.6.4...v0.6.5

English Version

Claude Code Haha v0.6.5

This release adds the OpenCode Go subscription provider, a searchable composer capability menu, model switching at plan approval, and fixes for a batch of large-session and history performance problems.

New features

• OpenCode Go subscription provider: enter your subscription API key to use GLM, Kimi, MiniMax, Qwen, Grok and GPT models. The local proxy routes each model to the wire format it expects, so no manual setup is needed.
• Composer capability menu: the "+" menu is now searchable, and is where Skills, connectors, agents, workflows, Agent Teams and Computer Use are reached.
• Switch the execution model at plan approval: pick a model right on the plan approval bar — same provider switches in place, cross-provider restarts the CLI with the new settings and continues, without restating the task.
• Context usage redone: the panel now leads with the remaining window and shows speed, cache hit rate and cost side by side.
• Project picker: the composer's directory picker searches all projects and can create a new one through the native folder dialog.
• Claude Opus 5: added as the official default model.

Performance and stability

• Large sessions open again: subagent transcripts are no longer inlined into session history, dropping a 542 MB response to 40 MB. Some sessions previously failed outright with Unexpected end of JSON input.
• History and traces served from the local index: opening or polling a trace no longer reads the whole session JSONL, scrolling back through long sessions stays continuous, and the search index no longer grows unbounded.
• Fixed inspection of large records and partial transcripts.

Fixes

• Model selection: a sidebar refresh no longer reverts the model you selected.
• AskUserQuestion: switching tabs or scrolling no longer loses answers already entered; the composer locks while a question is pending so a message is not mistaken for an answer.
• Token accounting: counts new input + output instead of treating cache hits as spend (a 270k-window DeepSeek session used to read as 40M tokens).
• New session directory: sessions are created from the project root instead of a stale worktree (which caused Branch "main" is already checked out).
• Task stop: stopping an already-finished task no longer errors, and the Activity panel stops showing it as running.
• Connector installs: fixed the leftover enabledPlugins entry that made every later Skill install fail.
• Very long unbreakable strings in the composer no longer widen past the panel.

Installation

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

Windows applications and installers in this release are unsigned while SignPath approval is pending. macOS builds are signed and notarized.

• macOS: Claude-Code-Haha-0.6.5-mac-arm64.dmg / Claude-Code-Haha-0.6.5-mac-x64.dmg
• Windows: Claude-Code-Haha-0.6.5-win-x64.exe / Claude-Code-Haha-0.6.5-win-arm64.exe
• Linux x64: Claude-Code-Haha-0.6.5-linux-x86_64.AppImage, Claude-Code-Haha-0.6.5-linux-amd64.deb or Claude-Code-Haha-0.6.5-linux-x86_64.rpm
• Linux ARM64: Claude-Code-Haha-0.6.5-linux-arm64.AppImage, Claude-Code-Haha-0.6.5-linux-arm64.deb or Claude-Code-Haha-0.6.5-linux-aarch64.rpm

Full changelog: v0.6.4...v0.6.5