OpenCode v1.14.42 - 更新日志

## 核心 ### 改进 - 为大型非流式响应添加了 HTTP API 响应压缩。 - 新增了 Scout 代理，用于仓库研究、文档查找和依赖源检查。 - 新增了工作区同步功能，以便适配器支持的工作区可以自动发现和注册。 - 为 `opencode run` 添加了交互式分割底部模式。 - 将 TUI 按键绑定配置简化为扁平的键绑定格式。 - 重复的工作树名称现在会回退到父目录名称，使标签更清晰。 ### Bug 修复 - 修复了 HTTP API 认证响应，使其在空授权结果和共享错误时与之前的传输格式一致。 - 从 HTTP API 返回结构化的验证错误。 - 在 HTTP API 中拒绝无效的权限和问题 ID。 - 在类型化的 HTTP API `401` 响应中包含认证挑战信息。 - 修复了 HTTP API OpenAPI 文档路由。 - 修复了分离的会话，使其能够重新被源项目认领。 - 修正了通过 npm shim 运行时 `SIGINT`、`SIGTERM` 和 `SIGHUP` 信号的转发。(@chubes4) - 允许没有描述的技能正常加载。 - 在 effect HTTP API 根路由上要求认证。(@RajvardhanPatil07) - 保持工具顺序稳定，使工具列表在不同运行之间不再重新排列。 - 使重试对话框更具体地针对服务提供商和失败原因。 - 修复了 Gemini 推理控制，使支持的推理力度级别和小模式默认值与模型匹配。 - 修复了 Anthropic Opus 4.5 的推理力度选项。 - 将 OpenAI 深度研究模型限制在其实际支持的推理级别。 - 修复了 GPT-5 推理变体，使暴露的力度选项与各模型系列匹配。 ## TUI ### 改进 - 在项目侧边栏中将重试中的会话显示为活动状态。(@edemaine) ### Bug 修复 - 修复了侧边栏中语言服务器状态的提示消息。(@Polo123456789) - 将会话选择器的排序方式从按天分组改为按完整的最后更新时间排序。(@Sleepful) - 在草稿历史中保留已清除的较长提示，以便可以恢复。 ## 桌面版 ### 改进 - 将桌面版更新切换为静默的逐用户安装流程。 **感谢 5 位社区贡献者：** - @edemaine： - 功能(桌面版)：项目侧边栏上的活动指示器 (#26223) - @RajvardhanPatil07： - 修复(服务器)：要求 effect 根路由进行认证 (#26361) - @chubes4： - 修复(CLI)：从 npm shim 转发信号 (#26259) - @Sleepful： - 修复(TUI)：按完整更新时间戳排序会话选择器 (#24725) - @Polo123456789： - 修复(侧边栏)：修复逻辑和误导性消息 #26469 (#26470)

## Core ### Improvements - Added HTTP API response compression for large non-streaming responses. - Added the Scout agent for repo research, docs lookup, and dependency-source inspection. - Added workspace sync so adapter-backed workspaces can be discovered and registered automatically. - Added an interactive split-footer mode for `opencode run`. - Simplified TUI keybinding config into a flat keybind format. - Made duplicate worktree names fall back to the parent directory for clearer labels. ### Bugfixes - Fixed HTTP API auth responses to match the previous wire format for empty authorize results and share errors. - Returned structured validation errors from the HTTP API. - Rejected invalid permission and question IDs in the HTTP API. - Included auth challenges on typed HTTP API `401` responses. - Fixed the HTTP API OpenAPI document route. - Fixed detached sessions so they are claimed by the source project again. - Forwarded `SIGINT`, `SIGTERM`, and `SIGHUP` correctly when running through the npm shim. (@chubes4) - Allowed skills without descriptions to load correctly. - Required auth on effect HTTP API root routes. (@RajvardhanPatil07) - Kept tool ordering stable so tool lists stop reshuffling between runs. - Made retry dialogs more specific to the provider and failure reason. - Fixed Gemini reasoning controls so supported effort levels and small-mode defaults match the model. - Fixed Anthropic Opus 4.5 reasoning effort options. - Limited OpenAI deep research models to the reasoning level they actually support. - Fixed GPT-5 reasoning variants so the exposed effort options match each model family. ## TUI ### Improvements - Show retrying sessions as active in the project sidebar. (@edemaine) ### Bugfixes - Fixed the sidebar message shown for language server state. (@Polo123456789) - Sorted the session picker by full last-updated time instead of day buckets. (@Sleepful) - Kept longer cleared prompts in draft history so they can be restored. ## Desktop ### Improvements - Switched desktop updates to silent per-user install flow. **Thank you to 5 community contributors:** - @edemaine: - feat(desktop): working indicator on project sidebar (#26223) - @RajvardhanPatil07: - fix(server): require auth for effect root routes (#26361) - @chubes4: - fix(cli): forward signals from npm shim (#26259) - @Sleepful: - fix(tui): sort session picker by full updated timestamp (#24725) - @Polo123456789: - fix(sidebar): fix logic and missleading message #26469 (#26470)