fix(core): recognize DeepSeek V4 Flash Vision image input

[sunrioa]

Generated-by: OpenAI Codex

Summary

DeepSeek's first-party inventory returns deepseek-v4-flash-vision-exp as a bare ID. Without a catalog entry, Maka classifies it as text-only, filters image attachments, and displays only the raw model ID.

Add explicit static metadata from DeepSeek's Vision guide and model details, including text-and-image input, friendly display metadata, documented context/output limits, structured output, tool calling, and reasoning controls. Keep this entry local to the first-party DeepSeek provider until the generated models.dev catalog catches up.

Add focused coverage for metadata and modality resolution, catalog projection from a bare provider ID, and the adjacent non-vision deepseek-v4-flash model.

Fixes #3417

Root cause

The first-party /models response supplies no modality facts for the new model, and Maka's generated metadata snapshot predates the 2026-08-21 release. resolveModelVisionSupport() therefore finds no provider or static vision signal and resolves to false.

Verification

• npm --workspace @maka/core test — 578 passed, 0 failed
• npm run lint — passed
• npm run format:check — passed
• npm run build — passed
• npm run typecheck — passed
• ./node_modules/.bin/knip --workspace apps/desktop — passed
• ./node_modules/.bin/knip --workspace packages/ui — passed
• Desktop smoke test — selected deepseek-v4-flash-vision-exp, attached an image, and received an image-grounded response

Before: Maka classifies the model as text-only and filters the attached image.

After: Maka recognizes the model as multimodal, sends the same image, and receives an image-grounded response.

Full npm test was not completed locally: an earlier run stalled in packages/eval/dist/__tests__/provider-admission-integration.test.js and was interrupted. The affected Core suite completed without failures, and required CI will exercise the remaining workspaces.

Scope

• The generated models.dev snapshot is unchanged; this is an explicit static entry until the upstream catalog includes the model.
• No lifecycle is assigned because DeepSeek calls the model experimental, while Maka has no exact matching lifecycle value.
• Protocol selection is unchanged; the model continues to use Chat Completions, whose image input path DeepSeek documents and the Desktop smoke test verified.
• The model is not added to the offline fallback list.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope:

OpenAI Codex assisted with official-documentation review, diagnosis, implementation, regression tests, local verification, and drafting this description. I reproduced the issue, reviewed the final diff and test results, and performed the manual smoke test.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

Thanks for filling in this metadata, especially for including the official docs link and the context window.

I am closing this because #3605 is doing the same thing — registering deepseek-v4-flash-vision-exp in STATIC_MODEL_METADATA.deepseek. #3605 is out of draft and under review, while this one has been in draft without updates since 21 August, so I am closing this side to avoid two parallel efforts.

Two things here differ from #3605, and I have raised both there — please take a look:

1. Your thinkingOptions.efforts is ['low', 'high', 'max']; fix(core): add DeepSeek V4 Flash Vision to model metadata #3605 has ['high', 'max'], without the low tier.
2. Your capabilities has no webSearch; fix(core): add DeepSeek V4 Flash Vision to model metadata #3605 adds webSearch: true.

You also included model-catalog.test.ts and model-metadata.test.ts, and #3605 has no tests at all. If you confirm those two facts and move the tests over to #3605, nothing from this PR is lost.

简体中文

谢谢你补这条元数据，尤其是把官方文档链接和上下文窗口这些都写全了。

这条我准备关掉，因为 #3605 在做同一件事——同样在 STATIC_MODEL_METADATA.deepseek 里注册 deepseek-v4-flash-vision-exp。#3605 已经脱离 draft 并在 review 中，这条从 8 月 21 日起一直是 draft 没有再动，为了避免两条并行我先关这边。

不过你这条有两处内容是 #3605 没有的，我已经在那边提出来了，麻烦你也过去看一眼：

1. thinkingOptions.efforts 你写的是 ['low', 'high', 'max']，fix(core): add DeepSeek V4 Flash Vision to model metadata #3605 写的是 ['high', 'max']，少一档 low。
2. capabilities 你没有 webSearch，fix(core): add DeepSeek V4 Flash Vision to model metadata #3605 加了 webSearch: true。

另外你还带了 model-catalog.test.ts 和 model-metadata.test.ts 两个测试，#3605 没有测试。如果你愿意把这两处事实确认一下、再把测试补到 #3605 上，这条的价值就都保住了。