更新日志（中文）
[1.3.51] - 2026-09-13
新增
- 启动预览新增 OAuth 绑定行：API Key 账号的启动预览里会多出一行「OAuth 绑定」，位于「强制刷新 Token」一行上方、风格与其他行一致，可直接绑定或更换 OAuth 账号，绑定账号需要重新授权时同一行显示「授权异常」并提供「重新授权」。说明文案明确绑定后启动与普通账号没有任何差异，可使用 OAuth 的全部能力（远端压缩，浏览器操作，各类插件等），对话仍由当前 API Key 供应商处理。模型供应商的启动预览同样可用，并与账号总览共用同一个绑定弹框。
修复
- DeepSeek 模型元数据与官方声明完全对齐：直连与桌面注入模式下，客户端此前会把部分 GPT 模型的能力当成 DeepSeek 的能力显示（例如并不存在的极速档位与套餐限制），多 agent 能力、客户端最低版本等也与官方声明不一致。现在实例里的 DeepSeek 模型直接使用官方
models.json的完整条目。 - 切到 DeepSeek 时禁用客户端内置联网搜索并清理冲突配置：此前切换到 DeepSeek 后，客户端仍可能发出上游不支持的联网搜索请求，config.toml 里遗留的思考档位、压缩提示词、服务等级等设置也会被一起发给上游并可能报错。现在切号时按官方要求禁用内置联网搜索并临时移除这些冲突项，切走时按原值精确还原。
- 修复跨供应商会话切到官方账号后请求被拒：同一会话在 DeepSeek 与官方 GPT 账号之间来回切换时，历史里第三方产生的推理项会让官方后端报
input[i].content校验失败（array_above_max_length），普通请求和自动压缩都无法继续。网关在转发给官方账号前会清空这类字段，并丢弃只带第三方思考文本、没有可复用推理状态的推理项。 - 修复从 DeepSeek 网关切到普通账号后旧会话报错：第三方用过的会话会在历史里留下带可见思考文本的 reasoning 项，而官方直连账号要求这类项必须为空，切换后整段请求（普通回合与自动压缩）都会被拒。现在切号与启动前会自动清理这类历史项，写入前先生成备份，且只作用于直连官方的普通账号（走网关的账号仍由请求级清洗处理）。
Changelog (English)
[1.3.51] - 2026-09-13
Added
- OAuth binding row in the launch preview: API Key accounts get an OAuth binding row right above the force-refresh-token row, styled like the other rows, that shows the current binding and lets you bind or change it, plus an authorization-error badge and a Re-authorize action when the bound account needs it. The description states that after binding, launching is identical to a normal account and every OAuth capability is available (remote compaction, browser control, and more), while conversations still run on the current API Key provider. The model provider launch preview offers the same row and shares one binding dialog with the account overview.
Fixed
- DeepSeek model metadata now matches the official declaration exactly: in official-direct and desktop-injection modes the client used to display capabilities borrowed from GPT models as if they belonged to DeepSeek (a fast tier and plan gating that do not exist), and the multi-agent capability and minimum client version differed from what DeepSeek declares. Instance catalogs now use the complete official
models.jsonentries. - Built-in web search is disabled and conflicting settings are cleaned up when switching to DeepSeek: previously the client could still issue web-search requests that the upstream does not support, and leftover reasoning-effort, compaction-prompt and service-tier settings in config.toml were forwarded to the upstream and could fail. Switching now disables the built-in web search per the official setup and temporarily removes those conflicting keys, restoring the original values on switch-away.
- Fixed requests being rejected after switching a cross-provider thread to an official account: when a thread alternates between DeepSeek and an official GPT account, reasoning items produced by the third party in its history made the official backend reject the request with an
input[i].contentvalidation error (array_above_max_length), breaking both normal turns and automatic compaction. Before forwarding to an official account the gateway now clears that field and drops reasoning items that only carry third-party reasoning text with no reusable reasoning state. - Fixed old threads failing after switching from the DeepSeek gateway to a plain account: a thread used on a third-party provider keeps reasoning items with visible thinking text in its history, while an official direct account requires those items to be empty, so the whole request (normal turns and automatic compaction alike) was rejected after switching. Switching and launching now clean those history items up front, with a backup written first, and only for plain accounts that talk to the official backend (gateway-bound accounts keep the request-level cleanup).