fix(core): preserve image_url content parts for DeepSeek vision model

[nachoeverywhere]

What this PR does

The DeepSeek provider was flattening every message into a plain string for all models, which turned image content into a [Unsupported content type: image_url] placeholder. When a user configures the vision model deepseek-v4-flash-vision-exp (with modalities.image), an image it reads via read_file never reaches the model — the request silently degrades to text. This PR stops flattening content when the config declares image support, so image_url parts are forwarded to the vision model as-is, while text-only models keep the existing placeholder behavior.

Why it's needed

DeepSeek's text models require message.content to be a plain string, which is why the provider flattens arrays and substitutes non-text parts. But the dedicated vision model accepts OpenAI content parts with image_url. Because flattening was unconditional, images read via read_file were replaced with [Unsupported content type: image_url] and the vision model never saw them, making vision-capable configs silently text-only. Users who set deepseek-v4-flash-vision-exp as their active model (a documented vision model) get a broken experience with no error.

Reviewer Test Plan

How to verify

Configure deepseek-v4-flash-vision-exp as the active model with modalities.image: true, then ask the agent to look at an image (or read_file a .png). Before this PR the request body contains [Unsupported content type: image_url] in place of the image; after this PR the image_url content part is preserved and sent to the vision model. The regression test asserts that image_url parts survive when modalities.image is set.

Evidence (Before & After)

N/A — non-UI change (internal request shaping, no user-facing difference). See the regression test for the before (placeholder) vs after (preserved image_url) of the request body.

Tested on

OS	Status
🍏 macOS	✅
🪟 Windows	N/A
🐧 Linux	N/A

Environment (optional)

Unit tests via npm run test --workspaces (core package, vitest run); 28 tests including the new regression case.

Risk & Scope

• Main risk or tradeoff: vision-capable models switch from flattened-string to content-part messages. Only triggers when modalities.image === true, so text-only configs are unaffected.
• Not validated / out of scope: a live round-trip against the DeepSeek vision API; the DeepSeek text behavior is covered by existing tests.
• Breaking changes / migration notes: none — no config or CLI surface changed.

Linked Issues

Fixes #10027

中文说明

这个 PR 做了什么

DeepSeek 提供商此前将所有消息扁平化为纯字符串，这会把图片内容转换成 [Unsupported content type: image_url] 占位符。当用户配置视觉模型 deepseek-v4-flash-vision-exp（并设置 modalities.image）时，它通过 read_file 读取的图片永远不会到达模型——请求会静默退化为纯文本。此 PR 在配置声明支持图片时停止扁平化内容，从而将 image_url 部分原样转发给视觉模型，而纯文本模型则保留原有的占位符行为。

为什么需要

DeepSeek 的文本模型要求 message.content 是纯字符串，这也是提供商扁平化数组并用占位符替代非文本部分的原因。但专用的视觉模型接受带有 image_url 的 OpenAI 内容部分。由于扁平化是无条件的，通过 read_file 读取的图片会被替换为 [Unsupported content type: image_url]，视觉模型从未真正看到它们，使得配置了视觉能力的模型静默地变成纯文本。用户将 deepseek-v4-flash-vision-exp 设为主模型（一个文档化的视觉模型）时会得到一个没有报错的破损体验。

Reviewer 测试计划

如何验证

将 deepseek-v4-flash-vision-exp 配置为主模型并设置 modalities.image: true，然后让智能体查看一张图片（或 read_file 一个 .png）。在此 PR 之前，请求体中图片位置是 [Unsupported content type: image_url]；在此 PR 之后，image_url 内容部分被保留并发送给视觉模型。回归测试断言当 modalities.image 被设置时 image_url 部分能够存活。

前后证据（Before & After）

不适用——非 UI 变更（内部请求整形，无用户可见差异）。请参阅回归测试中的请求体前后对比（占位符 vs 保留的 image_url）。

测试环境

OS	状态
🍏 macOS	✅
🪟 Windows	N/A
🐧 Linux	N/A

环境（可选）

通过 npm run test --workspaces（core 包，vitest run）运行单元测试；28 个测试，包括新增的回归用例。

风险与范围

• 主要风险或权衡：视觉模型从扁平化字符串切换为内容部分组成。仅在 modalities.image === true 时触发，因此纯文本配置不受影响。
• 未验证 / 超出范围：针对 DeepSeek 视觉 API 的在线往返；DeepSeek 文本行为由现有测试覆盖。
• 破坏性变更 / 迁移说明：无——未改变任何配置或 CLI 表面。

关联 Issue

Fixes #10027

[qwen-code-ci-bot]

✅ Qwen Triage finished — view run. See the stage comments in this thread for the result.

✅ Qwen Triage 已完成 —— 查看运行。结果见本线程中的各阶段评论。

[qwen-code-ci-bot]

Thanks for the PR, @nachoeverywhere! Before we can review the change itself, the PR description needs to follow the repository's pull request template.

The current body uses free-form headings (Summary, Changes, Test plan, Screenshots / Demo), but none of the template's required sections are present:

• ## What this PR does
• ## Why it's needed
• ## Reviewer Test Plan (with ### How to verify, ### Evidence (Before & After), and the ### Tested on OS table)
• ## Risk & Scope
• ## Linked Issues
• the bilingual <details> Chinese summary the template asks for

Please edit the PR body to follow the template: https://github.com/QwenLM/qwen-code/blob/main/.github/pull_request_template.md

The substance you've already written maps cleanly onto those sections — this is a reformat, not new work. Once the body follows the template, re-trigger triage (@qwen-code /triage) and we'll continue with the review.

中文说明

感谢提交 PR，@nachoeverywhere！在进入代码审查之前，PR 描述需要按照仓库的 pull request 模板来填写。

当前正文使用的是自由格式的小节（Summary、Changes、Test plan、Screenshots / Demo），模板要求的小节一个都没有：

• ## What this PR does
• ## Why it's needed
• ## Reviewer Test Plan（包含 ### How to verify、### Evidence (Before & After) 以及 ### Tested on 操作系统表格）
• ## Risk & Scope
• ## Linked Issues
• 模板要求的双语 <details> 中文说明

请按照模板编辑 PR 正文：https://github.com/QwenLM/qwen-code/blob/main/.github/pull_request_template.md

你已经写好的内容可以很自然地对应到这些小节——这只是重新排版，不需要新增内容。正文符合模板后，重新触发 triage（@qwen-code /triage），我们会继续审查。

— Qwen Code · qwen3.8-max

[nachoeverywhere]

@qwen-code /triage

[qwen-code-ci-bot]

Sandboxed verification: ⚠️ incomplete — cancelled - workflow run

The verification job was cancelled before producing a report.

中文 — 判定：⚠️ 未完成 · 已取消

验证作业在生成报告前被取消。

— Qwen Code · sandboxed verification

[qwen-code-ci-bot]

tmux real-user testing: cancelled - workflow run

The testing job was cancelled before producing a verdict. See the workflow run for details.

— Qwen Code · tmux real-user testing

[nachoeverywhere]

Thanks for the review and for the related work in #9854 — it targets the same root cause with the same modalities.image gating, and the team's PR (tied to #9832) already covers both the official endpoint and custom gateways. I wasn't aware of the existing report/PR when I opened this; apologies for the overlap.

Closing this in favor of #9854, which is the right home for the change — earlier, maintained by the team, and with broader regression coverage. Happy to help with anything else if a piece of this is useful there; feel free to tag me.

— @nachoeverywhere