更新日志（中文）
[1.3.50] - 2026-09-13
新增
- DeepSeek 支持绑定 GPT 账号进行生图：在启动预览里开启后选择 GPT 账号，Codex 内置生图工具发出的图片需求会交给所选账号完成，对话仍由 DeepSeek 处理。
- 模型供应商新增 Codex 启动预览：模型供应商的「启动」改为打开与账号总览一致的启动预览（账号、供应商、模型与用量、上下文管理等），确认后绑定实例并启动；DeepSeek 可在其中选择启动方式（网关列出 / CDP 注入 / 直连官方）。
- 上下文管理统一到启动预览：添加账号、编辑 API Key 与模型供应商编辑弹框不再提供逐模型上下文，改为在启动预览里按实例设置。
- DeepSeek 自动压缩兜底：切到 DeepSeek 时写入本地兜底压缩配置，切走时按原值还原，其他账号不受影响。
- 赞助商线路变更自动跟随：赞助商 Base URL 变化后自动替换已保存的供应商与账号地址（APIKEY.FUN 已切换到 apikey.fan）。
- Codex API 服务支持账号并发数：可限制同一账号同时运行的会话数，账号占满时优先切换空闲账号，等待超时返回可读原因。
变更
- 弹框尺寸与高度更稳定：弹框尺寸不再受 CSS 打包顺序影响，高度超出视口时由内容区滚动，标题与操作按钮不会被裁掉。
- 鹈鹕测智请求与本地 API 服务对齐：请求特征与 API 服务保持一致，降低被上游风控的概率。
修复
- 修复实例网关生图请求的模型改写：经 ChatGPT 账号执行的生图请求不再被改写成供应商模型名而被上游拒绝。
Changelog (English)
[1.3.50] - 2026-09-13
Added
- DeepSeek can bind GPT accounts for image generation: enable it in the launch preview and pick GPT accounts, and image requests from the built-in Codex image tool run on those accounts while the conversation stays on DeepSeek.
- Model providers gain the Codex launch preview: starting a model provider now opens the same launch preview as the account overview (account, provider, models and usage, context management, and more) and binds the instance on confirm. DeepSeek picks its start mode there (gateway list / CDP injection / official direct).
- Context management moves into the launch preview: the Add account, Edit API Key, and model provider edit dialogs no longer expose per-model context; it is configured per instance in the launch preview.
- DeepSeek auto-compaction fallback: switching to DeepSeek writes a local fallback compaction config and switching away restores the previous values, leaving other accounts untouched.
- Sponsor route changes are applied automatically: when a sponsor's base URL changes, saved providers and accounts are rewritten to the new address (APIKEY.FUN now serves apikey.fan).
- Per-account concurrency for the Codex API service: limit how many sessions one account can run at once; full accounts fail over to idle ones and a timed-out wait returns a readable reason.
Changed
- More stable dialog sizing: dialog sizes no longer depend on the CSS bundle order, and tall dialogs scroll inside the body so the title and action buttons stay reachable.
- Pelican testing matches the local API service: requests use the same client fingerprint as the local API service, reducing the chance of upstream risk-control flags.
Fixed
- Fixed image requests being rewritten to the provider model in the instance gateway: image requests executed by ChatGPT accounts are no longer renamed to the provider model and rejected upstream.