v0.1.0-rc.8 release notes 声称 DeepSeek 适配器支持原生图片请求，实测官方 API 全部 400 拒绝

[pigbiglong]

v0.1.0-rc.8 release notes 声称 "DeepSeek 模型适配器支持配置启用原生图片请求"，实测完全不可用。

按 release notes 说明配置 inputModalities: [text, image] 后，直接请求 DeepSeek 官方 API（api.deepseek.com），三个模型 ID 全部被 400 拒绝：

模型 ID	实测结果
deepseek-v4-flash	HTTP 400: unknown variant 'image_url', expected 'text'
deepseek-chat	同样的 400
deepseek-v4-flash-vision-exp	同样的 400

官方端点的 chat-completions schema 里根本没有 image_url 这个消息类型。适配器把图片序列化成 OpenAI 兼容格式没有意义——服务器端拒收。

你们自己的实现笔记（.agents/notes/implemented/feature/2026-08-19-direct-deepseek-vision-input.md）也写得很清楚：

"The shipped catalog does not advertise deepseek-v4-flash-vision-exp until its model endpoint is ready"

即官方视觉端点尚未就绪。那么 rc.8 的 release note 应该如实写明"适配器代码已支持、官方端点未上线、当前不可用"，而不是让人以为配置了 inputModalities 就能传图。

代码就绪 ≠ 功能可用。 请修正 release notes 或明确标注依赖官方端点上线，避免用户浪费时间排查。

[ylwl1997]

The 400 body (unknown variant 'image_url', expected 'text') is the official chat-completions schema rejecting the part type, so this is not a local serialization glitch: the rc.8 adapter can emit OpenAI-style image parts, and api.deepseek.com currently does not accept them. That matches the in-tree note that deepseek-v4-flash-vision-exp stays out of the catalog until the model endpoint is actually ready.

Until the catalog advertises a live vision model on deepseek-official:

• leave inputModalities at text-only for deepseek-chat / deepseek-v4-flash against the official API — enabling [text, image] will keep 400’ing;
• if you need images today, route the session through a provider whose /v1/chat/completions already accepts image_url (or wait for the vision endpoint to exist).

Adapter support in rc.8 is real; “configure it and official DeepSeek will take pictures” is not, and the release note should say that gap in one sentence so people stop debugging their own config.