Skip to content

Claude Code Haha v0.6.2

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 11 Sep 11:02
· 7 commits to main since this release
v0.6.2
85e7f3a

Claude Code Haha v0.6.2

本次以修复为主：Computer Use 在正式安装包里的点击反馈与 JavaScript 单元定时器、Provider 请求的输出预算与协议收尾、计划模式批准后的权限模式，以及长回复的总时长上限。同时新增了服务商级的「请求兼容」设置。

Computer Use

打包应用中的点击反馈

  • 点击动画资源改为从运行中的原生 helper 目录解析，不再依赖 SwiftPM 生成的资源访问器；后者在正式安装包里会回退到构建机路径，导致资源加载失败并让后台进程退出。
  • 动画帧缺失或不可读时保留程序化点击反馈，可见点击仍然正常返回，不再影响后续操作。
  • 加固原生回归与打包门禁：纯原生改动纳入 macOS 检查，在最终发布包中验证重定位后的资源确实可加载，并拒绝解析到应用目录之外的资源。

JavaScript 单元内的定时器

  • setTimeout / clearTimeout / setInterval / clearInterval 现在可以在 Computer Use 的 JavaScript 单元里使用（数字句柄）。定时器归属当前单元，单元结束时一并清理，不会跨单元泄漏。
  • 改写持久化调用时保留原有语句边界，不再因为自动分号插入（ASI）改变执行语义。

新增与改进

Provider 请求兼容设置（#1290）

  • 服务商编辑中新增「回复输出预算」与「高级兼容设置」：可设置普通回复的最大输出 token、已知端点输出上限、输出 token 参数字段，以及采样参数、推理参数、并行工具调用、结构化输出四项能力开关。仅对 OpenAI 兼容（Chat / Responses）格式显示，不影响 Anthropic 格式。
  • 配置的预算不再被模型名推断出的上限压回默认值：普通回复按该预算请求，后台短请求仍保持较小预算。
  • 兼容项随服务商一起保存并同步到托管环境变量；对已保存服务商的协议探测与线上请求共用同一套解析逻辑。

计划模式审批携带权限模式（#1254）

  • 计划审批弹窗新增「批准并自动接受编辑」与「批准并跳过权限」，与官方终端行为一致，批准时显式固定进入实施阶段后的权限模式。
  • 修复批准计划后实施阶段对每个工具都询问一次权限的问题；同时补齐了通过权限更新切换模式时被跳过的退出计划模式清理（退出标志、退出与重入提醒、自动模式回收、进入时摘除的规则）。

其他修复

  • 流式回复总时长上限（#1307、#1275）：单次流式回复的总时长上限改为随「AI 请求超时」同步放宽，且不低于原有的 600 秒，慢速本地模型的长思考不再被中途切断。已运行中的会话在修改超时后立即生效，无需重开会话。
  • OpenAI 协议收尾与输出预算契约：Chat 与 Responses 两条流式路径按事件边界完整解析 SSE，正确处理结束事件与不完整流，避免回复被提前结束或截断；请求本身不合法时返回 invalid_request_error 而不是 502。
  • 会话内「本轮更改」卡片（#1305）：不再把上一轮的文件改动带到之后只聊天的回合；卡片默认折叠，展开状态在滚动和列表重建后保持，键盘焦点正确回到展开按钮。
  • 提交与 PR 不再自动附加署名（#1274）：移除提交信息与 PR 描述中的 Co-Authored-By 和 "Generated with Claude Code"。attribution / includeCoAuthoredBy 作为兼容字段保留，旧配置不会导致设置文件校验失败。

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-0.6.2-mac-arm64.dmg / Claude-Code-Haha-0.6.2-mac-x64.dmg
  • Windows：Claude-Code-Haha-0.6.2-win-x64.exe / Claude-Code-Haha-0.6.2-win-arm64.exe
  • Linux x64：Claude-Code-Haha-0.6.2-linux-x86_64.AppImage、Claude-Code-Haha-0.6.2-linux-amd64.deb 或 Claude-Code-Haha-0.6.2-linux-x86_64.rpm
  • Linux ARM64：Claude-Code-Haha-0.6.2-linux-arm64.AppImage、Claude-Code-Haha-0.6.2-linux-arm64.deb 或 Claude-Code-Haha-0.6.2-linux-aarch64.rpm

完整变更：v0.6.1...v0.6.2

English Version

Claude Code Haha v0.6.2

This is a fixes-focused release: click feedback and JavaScript-cell timers for Computer Use inside packaged builds, output budgets and protocol termination for provider requests, the permission mode applied after approving a plan, and the overall stream duration cap for long responses. It also adds provider-level request compatibility settings.

Computer Use

Click feedback in packaged apps

  • Click animation assets are now resolved from the running native helper instead of SwiftPM's generated resource accessor, which fell back to a build-machine path inside the signed app, failed to load, and took the daemon down with it.
  • When animation frames are missing or unreadable, procedural click feedback is kept, so visible clicks still return successfully and later actions keep working.
  • Native regressions and package gates are tightened: native-only changes run through the macOS checks, relocated resources are verified inside the final package, and resources resolving outside the app are rejected.

Timers inside a JavaScript cell

  • setTimeout / clearTimeout / setInterval / clearInterval are now available inside a Computer Use JavaScript cell with numeric handles. Timers belong to the current cell and are cleared when it ends, with no leakage across cells.
  • Rewriting persistent calls preserves the original statement boundaries, so automatic semicolon insertion no longer changes execution semantics.

New & Improved

Provider request compatibility settings (#1290)

  • The provider editor gains a "reply output budget" field and advanced compatibility options: a known endpoint output limit, the output token field, and switches for sampling parameters, reasoning parameters, parallel tool calls and structured output. They appear only for OpenAI-compatible (Chat / Responses) formats and do not affect the Anthropic format.
  • A configured budget is no longer clamped back to the model-family default: ordinary replies request that budget while short background requests keep their smaller one.
  • Compatibility options are saved with the provider and mirrored into managed environment variables; protocol probes for saved providers and live requests share the same resolution logic.

Plan approval carries a permission mode (#1254)

  • The plan approval dialog adds "Approve & auto-accept edits" and "Approve & bypass permissions", mirroring the official terminal and pinning the permission mode used during implementation.
  • This fixes implementation asking for permission on every tool after a plan was approved, and completes the plan-exit cleanup that was skipped when a mode arrived through a permission update (exit flags, exit and re-entry reminders, auto-mode teardown, and rules stripped on entry).

Other fixes

  • Overall stream duration cap (#1307, #1275): the wall-clock cap for a single streamed response now scales with the configured AI request timeout instead of staying at 600s, and never drops below it. Slow local models can keep thinking and generating without being cut off, and a session that is already running picks up a raised timeout immediately without being restarted.
  • OpenAI protocol termination and output budget contracts: both the Chat and Responses streaming paths parse SSE by event boundary, handle terminal events and incomplete streams correctly, and no longer end or truncate a response early; genuinely invalid requests now return invalid_request_error instead of a 502.
  • Per-turn file change card (#1305): changes from an earlier turn are no longer carried into later chat-only turns; cards default to collapsed, the expanded state survives scrolling and list remounts, and keyboard focus returns to the disclosure button.
  • No automatic attribution in commits and PRs (#1274): Co-Authored-By and "Generated with Claude Code" are no longer injected into commit messages or PR bodies. attribution and includeCoAuthoredBy remain as accepted compatibility fields, so existing settings files stay valid.

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-0.6.2-mac-arm64.dmg / Claude-Code-Haha-0.6.2-mac-x64.dmg
  • Windows: Claude-Code-Haha-0.6.2-win-x64.exe / Claude-Code-Haha-0.6.2-win-arm64.exe
  • Linux x64: Claude-Code-Haha-0.6.2-linux-x86_64.AppImage, Claude-Code-Haha-0.6.2-linux-amd64.deb or Claude-Code-Haha-0.6.2-linux-x86_64.rpm
  • Linux ARM64: Claude-Code-Haha-0.6.2-linux-arm64.AppImage, Claude-Code-Haha-0.6.2-linux-arm64.deb or Claude-Code-Haha-0.6.2-linux-aarch64.rpm

Full changelog: v0.6.1...v0.6.2

Assets 25
Loading