Cockpit Tools v1.3.8

更新日志（中文）

[1.3.8] - 2026-07-17

新增

• 已有 Codex 账号可直接加入 Codex API 服务（#1628）：已导入 Cockpit 的符合条件账号，现在可在卡片、列表或表格视图中直接加入 API 服务，无需离开当前页面；操作复用增量加入账号池流程，并继续遵守 Free 账号、待授权账号和不兼容 API Key 的限制。感谢 @Ac-spider。
• Kiro 支持 AWS IAM Identity Center 登录：添加账号弹框新增 AWS Builder ID 与企业账号设备授权；企业账号可填写 AWS Region 和 IAM Identity Center Start URL，授权成功后会保留客户端注册上下文并写入官方 AWS SSO 缓存文件，保证 Token 刷新与 Kiro 真实切号继续可用。

修复

• 修复第三方纯文本模型因图片工具结果中断 Codex 对话的问题：Provider Gateway 现在按模型真实能力声明图片输入；模型不支持图片且没有有效视觉路由时，会将 view_image 结果及历史图片替换为明确的省略说明并继续使用当前模型，配置有效视觉路由时仍会自动切换处理图片，同时图片路由配置在应用重启后不再丢失。
• 修复 macOS Codex 受管实例在窗口尚未就绪时被误报为已启动的问题：实例现在通过 LaunchServices 的 open -n -a 启动，同时保留 CODEX_HOME、CODEX_ELECTRON_USER_DATA_PATH 和独立的 --user-data-dir；Cockpit 会匹配真实 ChatGPT 主进程 PID，不再保存临时 open launcher PID 或僵尸进程。
• 修复 macOS 官方客户端迁移到 /Applications/ChatGPT.app 后，Codex 启动仍停留在旧 /Applications/Codex.app 路径的问题：解析应用根目录前，会通过受保护的配置更新迁移精确匹配的官方旧路径；自定义安装位置以及未检测到 ChatGPT 应用的环境保持不变。（#1631）感谢 @jackychanisnotme。
• 修复 Responses 流式响应提前输出半段 JSON 或合并多个事件的问题：网络分片会先缓存到完整事件边界，符合严格条件的连续 JSON 事件会拆成独立 SSE 帧，避免客户端解析失败和后续事件丢失。（#1632）感谢 @Ac-spider。
• 修复关闭历史存储时，长对话回放可能因 Item with id not found 失败的问题：回放前清理缺少有效加密内容的孤立 reasoning ID，保留有效签名和明确启用 store=true 的请求，并将大型历史改为单次重建。（#1634）感谢 @Ac-spider。
• 修复兼容 Chat Completions 供应商遗漏工具调用 ID 时 Codex 无法执行工具的问题：流式与非流式 Responses 事件会生成并复用稳定的备用 ID，不覆盖规范供应商已经返回的 ID。（#1633）感谢 @Ac-spider。
• 修复图片专用 gpt-image-* 模型被错误调度到 Chat Completions 的问题：无效请求会在占用账号池并发前返回明确的 400，Responses 与图片专用端点的行为保持不变。（#1630）感谢 @Ac-spider。

Changelog (English)

[1.3.8] - 2026-07-17

Added

• Existing Codex accounts can be added directly to Codex API Service (#1628): eligible accounts already imported into Cockpit can now be added from the card, list, or table view without leaving the current page; the action reuses the incremental account-pool flow and keeps the existing restrictions for Free accounts, pending authorization, and incompatible API keys. Thanks @Ac-spider.
• Kiro supports AWS IAM Identity Center sign-in: the add-account dialog now supports AWS Builder ID and Enterprise device authorization; Enterprise sign-in accepts an AWS Region and IAM Identity Center Start URL, while successful accounts preserve their client-registration context and write the official AWS SSO cache files so token refresh and real Kiro account switching continue to work.

Fixed

• Fixed Codex conversations stopping when image tool results reached text-only third-party models: Provider Gateway now advertises image input according to each model's actual capabilities; when the selected model has neither image support nor a valid vision route, view_image results and historical images are replaced with an explicit omission notice and the conversation continues on the current model, while configured vision routes still switch automatically and remain preserved across app restarts.
• Fixed managed Codex instances on macOS sometimes being reported as started before their GUI was ready: instances now launch through LaunchServices with open -n -a, while CODEX_HOME, CODEX_ELECTRON_USER_DATA_PATH, and the isolated --user-data-dir are preserved; Cockpit resolves the real ChatGPT process PID and no longer stores the temporary open launcher PID or zombie processes.
• Fixed macOS Codex launches remaining pinned to the legacy /Applications/Codex.app after the official client moved to /Applications/ChatGPT.app: exact legacy official paths are migrated through the guarded config update before the app root is resolved, while custom locations and systems without the ChatGPT app remain unchanged. (#1631) Thanks @jackychanisnotme.
• Fixed Responses streams emitting incomplete or concatenated JSON events: transport fragments are buffered until they form a complete event, while narrowly valid concatenated event documents are separated into independent SSE frames, preventing client parse failures and lost follow-up events. (#1632) Thanks @Ac-spider.
• Fixed long Responses conversations failing with Item with id not found while history storage is disabled: orphan reasoning IDs without usable encrypted content are removed before replay, valid reasoning signatures and explicit store=true requests remain unchanged, and large histories are rebuilt in one pass. (#1634) Thanks @Ac-spider.
• Fixed compatible Chat Completions providers omitting tool-call IDs and breaking Codex tool execution: deterministic fallback IDs are preserved across streaming and non-streaming Responses events without replacing IDs supplied by conforming providers. (#1633) Thanks @Ac-spider.
• Fixed image-only gpt-image-* models being dispatched through Chat Completions: invalid requests now return a clear 400 before occupying account-pool concurrency, while Responses and dedicated image endpoints keep their existing behavior. (#1630) Thanks @Ac-spider.