Skip to content

Claude Code Haha v0.6.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 06 Sep 03:58
· 2 commits to main since this release
v0.6.0
633320e

Claude Code Haha v0.6.0

We have completely rebuilt Computer Use on macOS, taking Codex's background-operation experience as the reference: the agent can click, type, scroll and drag in the background without taking over your mouse or keyboard, while you keep working in other apps. This is the headline of v0.6.0, alongside easier IM connections, a task-focused sidebar and a clearer view of what the model actually receives.

New & Improved

Computer Use, rebuilt for macOS

  • A native macOS engine delivers input to the target app and window, with a virtual cursor that shows the agent's actions independently of your own pointer.
  • Background window capture stays fresh when another app covers the target. Window movement, resizing, minimized targets and focus changes are handled more reliably, including repeated text entry during long tasks.
  • App selection shows native icons and resolves the exact app bundle. Accessibility controls with duplicate groups or descriptive labels remain actionable.
  • Requires macOS 14.4 or later and system Accessibility / Screen Recording permissions. Repeated per-app approval prompts are replaced by confirmation when enabling Computer Use. Hidden or minimized targets may be restored to a visible window, and background helpers shut down with the desktop app.
  • Windows Computer Use also gains more reliable input handling, target validation and a consistent virtual cursor. Windows still shares the system mouse and keyboard; the background-input experience described above is specific to macOS.

Connect more chat apps, with less setup

  • Create a Feishu bot by scanning a QR code in desktop settings: authorize on your phone and the new bot's credentials are saved automatically. Existing bot configurations stay intact, and international Lark tenants are supported.
  • WeCom, QQ and Slack adapters join the desktop session workflow. WeCom and QQ support scan-to-bind setup; Slack provides an app manifest with the required permissions and Socket Mode configuration.
  • The three new adapters use persistent connections without a public callback URL, and support private chats with pairing, session recovery, tool permissions and attachments.

Find tasks and understand model requests

  • Task view in the sidebar: switch from project grouping to a flat list with running sessions first, then calendar-day groups. Each task shows its workspace, and tasks waiting for permission have a distinct status indicator.
  • Expanded Trace view: inspect the opening system prompt and tool catalog from the session overview, distinguish injected context from your messages, and scan tool inputs and token counts. OpenAI Responses requests and truncated request previews are parsed more faithfully.
  • Conversation-only rewind (#1273, #962): roll back pure-chat or failed turns without restoring files, including when no file checkpoint is available. Replacement turns now use the correct checkpoint scope.

Fixes

  • GPT relay reasoning effort (#1238): preserve the selected reasoning effort when sending requests through compatible providers.
  • DeepSeek reasoning round trips (#1269): preserve reasoning_content for DeepSeek models behind OpenAI-compatible relays, fixing missing-reasoning 400 errors.
  • GLM 5.3 / GLM-5.3-Flash compatibility (#1283): recognize the models' reasoning capabilities and send compatible thinking parameters, fixing the reported 1210 errors.
  • Computer Use screenshots through compatible providers (#1277): preserve tool-result image blocks in OpenAI Chat proxy transformations so vision-capable models receive the screenshots.
  • Tool input interrupted after 120 seconds (#1271, #1272, #1288, #1295): use stalled-progress detection instead of aborting while valid tool parameters are still arriving. Stalled streams and the overall response limit remain bounded.
  • Long rate-limit retry messages (#1266): explain usage or rate limits explicitly and show readable wait times instead of a large raw seconds count.
  • Incomplete Telegram replies (#1264): retain complete streamed replies and recover delivery correctly after rate limiting.
  • Images moved out of the reply body (#1240): preserve local images at their original Markdown positions instead of collecting them at the end.
  • Preserve the selected project directory after creating a session, count unique edited files in the activity summary, and clarify cached-token accounting.
  • Improve macOS dependency installation and application shutdown, including cleanup failures and background helper termination.
  • Retire expired sponsor placements while preserving providers users have already saved.

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-0.6.0-mac-arm64.dmg / Claude-Code-Haha-0.6.0-mac-x64.dmg
  • Windows: Claude-Code-Haha-0.6.0-win-x64.exe / Claude-Code-Haha-0.6.0-win-arm64.exe
  • Linux x64: Claude-Code-Haha-0.6.0-linux-x86_64.AppImage, Claude-Code-Haha-0.6.0-linux-amd64.deb or Claude-Code-Haha-0.6.0-linux-x86_64.rpm
  • Linux ARM64: Claude-Code-Haha-0.6.0-linux-arm64.AppImage, Claude-Code-Haha-0.6.0-linux-arm64.deb or Claude-Code-Haha-0.6.0-linux-aarch64.rpm

Full changelog: v0.5.5...v0.6.0

中文版本 / Chinese Version

Claude Code Haha v0.6.0

我们彻底重构了 Computer Use 在 macOS 上的表现，对标 Codex 的后台操作体验：Agent 可以在后台点击、输入、滚动和拖拽，不占用你的鼠标、不占用你的键盘，你可以继续使用其他应用。这是 v0.6.0 最重要的更新，同时带来更便捷的 IM 接入、面向任务的侧边栏，以及更清晰的模型请求追踪。

新增与改进

彻底重构 macOS Computer Use

  • 使用原生 macOS 引擎，向目标应用和窗口投递操作；独立的虚拟光标展示 Agent 的动作，不影响你自己的鼠标指针。
  • 目标窗口被其他应用遮挡时，仍能持续获取新画面。窗口移动、缩放、最小化恢复、焦点切换和长任务中的连续输入更加可靠。
  • 应用选择显示原生图标并精确定位应用路径；修复辅助功能树中的重复分组和带描述标签控件无法操作的问题。
  • 需要 macOS 14.4 及以上版本，以及系统辅助功能与屏幕录制权限。逐应用授权弹窗改为启用 Computer Use 时统一确认；隐藏或最小化的目标可能恢复为可见窗口，退出桌面端时后台辅助进程一并结束。
  • Windows Computer Use 同步改进输入可靠性、目标校验和虚拟光标。Windows 仍会使用系统鼠标和键盘，上述后台输入体验仅适用于 macOS。

更多 IM 平台，更简单的接入

  • 飞书扫码创建机器人：在桌面端设置中扫码、手机确认，即可创建新机器人并自动保存凭据；保留已有机器人的配置，同时支持国际版 Lark。
  • 新增企业微信、QQ 和 Slack 适配器，接入统一的桌面会话流程。企业微信和 QQ 支持扫码绑定；Slack 提供预填权限和 Socket Mode 配置的应用清单。
  • 三个新适配器均使用长连接，无需公网回调地址；当前支持私聊，包含配对、会话恢复、工具权限和附件处理。

更容易找到任务、看懂模型请求

  • 侧边栏任务视图：从按项目分组切换为平铺任务列表，运行中的任务优先展示，其余按自然日归组。每项显示所属工作区，等待授权的任务有独立状态标识。
  • 增强 Trace 追踪视图：在会话概览查看起始系统提示词和工具列表，区分注入上下文与用户消息，直接浏览工具输入和 Token 用量；完善 OpenAI Responses 请求和截断请求预览的解析。
  • 仅回退对话（#1273、#962）：纯聊天或失败的回合也能回退，不恢复磁盘文件，没有文件检查点时同样可用；替换回合后的检查点范围也得到修正。

修复

  • GPT 中转推理强度丢失（#1238）：修复通过兼容 Provider 请求时未正确传递所选推理强度的问题。
  • DeepSeek 推理内容回传报错（#1269）：为 OpenAI 兼容中转后的 DeepSeek 模型保留 reasoning_content，修复缺少推理内容导致的 400 错误。
  • GLM 5.3 / GLM-5.3-Flash 兼容（#1283）：补齐模型推理能力识别，发送匹配的思考参数，修复报告中的 1210 错误。
  • 兼容 Provider 丢失 Computer Use 截图（#1277）：修复 OpenAI Chat 代理转换丢弃工具结果图片块的问题，让支持视觉的模型收到截图。
  • 工具参数生成超过 120 秒被误中断（#1271、#1272、#1288、#1295）：改为检测流是否停滞，持续收到有效参数时不再按总耗时误终止；仍保留停滞超时和整段响应时长上限。
  • 限流重试提示不明确（#1266）：明确显示用量或速率限制原因，并用易读时长替代大量秒数。
  • Telegram 回复不完整（#1264）：保留完整流式回复，并修复遇到限流后的消息投递恢复。
  • 图片被移到回复末尾（#1240）：保留本地图片在 Markdown 正文中的原始位置，不再统一提取到末尾。
  • 创建会话后保留所选项目目录；活动摘要按唯一文件统计修改数量；缓存 Token 统计说明更加清晰。
  • 改进 macOS 构建依赖安装和应用退出流程，处理清理异常与后台辅助进程残留。
  • 下线过期赞助展示，保留用户已经保存的服务商配置。

安装

请在本页面底部 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-0.6.0-mac-arm64.dmg / Claude-Code-Haha-0.6.0-mac-x64.dmg
  • Windows：Claude-Code-Haha-0.6.0-win-x64.exe / Claude-Code-Haha-0.6.0-win-arm64.exe
  • Linux x64：Claude-Code-Haha-0.6.0-linux-x86_64.AppImage、Claude-Code-Haha-0.6.0-linux-amd64.deb 或 Claude-Code-Haha-0.6.0-linux-x86_64.rpm
  • Linux ARM64：Claude-Code-Haha-0.6.0-linux-arm64.AppImage、Claude-Code-Haha-0.6.0-linux-arm64.deb 或 Claude-Code-Haha-0.6.0-linux-aarch64.rpm

完整变更：v0.5.5...v0.6.0

Assets 25
Loading