更新日志（中文）
[1.3.71] - 2026-10-20
新增
- API 服务支持 ChatGPT 官方第三方授权：绑定原 Codex 账号，在官方授权与本地网关之间选择，保留原账号资料和额度。支持授权管理、自动续期和“不再提醒”；移出服务仍保留授权，测智与唤醒优先使用已绑定授权。官方通道受模型和工具权限限制，不支持生图。
- Codex 自动额度同步：活跃本地 OAuth 账号每 5–30 秒自动刷新额度，无需启动 API 服务；沿用账号禁用、套餐和分组筛选设置。
- OAuth 额度保护：自动执行安全垫、临界区并发限制和请求前后额度确认，展示额度新鲜度与保护状态。
- API 服务自动能力兜底：账号不支持所需能力时，在当前 Key 授权范围内寻找兼容账号；仅在明确未执行且历史可安全重放时切换。
- Codex 速度：恢复标准、快速、超高速三档，保留默认设置并保存用户选择；超高速费用按标准档 6 倍估算。
- 默认开启 Codex 客户端增强：纯 API Key 账号无需绑定 ChatGPT OAuth，即可使用官方速度菜单和麦克风录音、取消与重试；档位按模型能力显示，听写通过当前实例的转写 API 回填草稿，也可主动选择发送。原生控件在干净的新启动页面自动加载；启动加载、菜单交互和临时忙碌只延后加载，在安全的启动窗口内自动重试，保留正在使用的窗口和未发送草稿。混合模式下，绑定的 ChatGPT 订阅额度耗尽后仍可继续本地 API 对话。额度徽章在录音和窄窗口下自动让位。通用设置与 Codex 快捷设置共用配置，说明文字对齐，保留用户明确关闭的选择；实际能力与费用仍取决于供应商。
- 鹈鹕测智结果卡片支持快捷编辑和展示账号标签。
变更
- 改善安装包构建缓存，减少 macOS Universal 安装包的重复编译。
- API 服务新绑定 OAuth 账号默认保留 30% 的 5 小时额度和 10% 的周额度，保留已有设置；接近保留值时加快额度确认。
- 代理订阅自动更新并按需检查健康状态；沿用节点已声明的证书例外，普通 TLS 节点继续校验证书。
- OAuth 切号自动保留经本地记录验证且有效的外部桥接。
- 改善工具协议、流式响应和子智能体历史兼容，相关适配自动生效，无需额外开关或兼容模型名单。
修复
- 打包和验证后自动卸载临时 DMG 卷；卸载失败时保留临时文件并明确报错。
- 修复额度不足换号、重启或切换引擎后，部分宿主消息和旧工具历史无法继续对话的问题。
- 保留 TLS 故障详情并重建受影响的连接池，不强制重新授权或重放已部分输出的工具调用。
- 修复 Windows 代理进程退出后残留占用端口，以及从 SUBST 驱动器或链接目录构建时重复 React 导致的启动白屏。
- 修复模型配置保存与弹框菜单布局；统一 GPT-6 Astra、Sol、Luna 默认上下文为 272,000 tokens、最大 872,000 tokens，保留用户设置。
- 仅调整图片能力时保留 Responses 原生路径和各 API Key 独立配置。
- 启用客户端注入的受管 Codex 实例，其前端语音调用同样通过 Cockpit 代理。
- 修复更新下载重复启动、进度回退的问题。
Changelog (English)
[1.3.71] - 2026-10-20
Added
- Official third-party ChatGPT access for API Service: Link access to an existing Codex account and choose official access or the local gateway while retaining account details and quota displays. Includes access management, automatic renewal and reminder opt-out. Removing a service member retains access; Pelican and wakeups prefer linked official access. Official model and tool restrictions apply; image generation is unavailable.
- Automatic Codex quota sync: Active local OAuth accounts refresh quota every 5–30 seconds without API Service, respecting disabled accounts, plan filters and group settings.
- OAuth quota protection: Apply safety margins, near-reserve concurrency limits and quota confirmation before and after requests, with quota freshness and protection status visible.
- Automatic capability fallback in API Service: Find a compatible account within the current key’s authorized scope when needed, switching only after definite pre-execution rejection with safely replayable history.
- Codex speed: Restore Standard, Fast and Ultrafast, preserving defaults and saving user selections. Ultrafast cost estimates use 6× Standard rates.
- Codex desktop enhancements, enabled by default: Pure API-key sessions can use the native speed menu and microphone recording, cancel and retry controls without ChatGPT OAuth. Speed tiers follow model support; dictation uses the current instance’s transcription API and inserts text into the draft unless you explicitly choose to send. Native controls load automatically on a clean startup; startup loading, menu interaction and temporary busy states defer loading and are retried within the safe startup window, preserving windows already in use and unsent drafts. Mixed-mode local API conversations can continue when the bound ChatGPT subscription is exhausted. Quota badges make room during recording and in narrow windows. General Settings and Codex quick settings share configuration and aligned descriptions; explicit opt-outs are retained. Provider support and pricing still apply.
- Edit and view account tags directly on Pelican result cards.
Changed
- Improve build caches and reduce duplicate compilation for macOS Universal packages.
- New API Service OAuth bindings reserve 30% of the five-hour quota and 10% of the weekly quota by default, preserving existing settings. Quota confirmation accelerates near the reserve.
- Proxy subscriptions update automatically with on-demand health checks. Existing node certificate exceptions remain in effect; ordinary TLS nodes retain certificate validation.
- OAuth account switches automatically preserve active external bridges verified against local integration records.
- Improve tool protocols, streaming responses and subagent history compatibility, with automatic adaptation and no extra switches or compatibility-model lists.
Fixed
- Automatically unmount temporary DMG volumes after packaging and verification; retain temporary files and report errors when unmounting fails.
- Fix continuation failures involving host messages and legacy tool history after quota-triggered account switches, restarts or engine changes.
- Preserve TLS failure details and rebuild affected connection pools without forcing reauthorization or replaying partially delivered tool calls.
- Fix lingering Windows proxy processes holding ports, and startup blank screens caused by duplicate React runtimes in builds from SUBST drives or linked directories.
- Fix model-setting persistence and dialog menu layouts. Align GPT-6 Astra, Sol and Luna context defaults to 272,000 tokens with an 872,000-token maximum while retaining user settings.
- Preserve native Responses routes and independent API-key settings when only image capabilities change.
- Route frontend voice calls through Cockpit for managed Codex instances with client injection enabled.
- Fix duplicate update downloads and progress moving backward.