fix(core): recognize new DeepSeek/GLM vision models in modality defaults

[jarvislee90s-dot]

What this PR does

Extends the modality auto-detection table (MODALITY_PATTERNS in packages/core/src/core/modalityDefaults.ts) so the new vision-capable models are recognized without a manual modalities override: DeepSeek vision variants (deepseek-v4-flash-vision-exp, and vl-suffixed naming), the GLM v-suffix family (glm-4.5v/glm-4.6v/glm-5v-turbo), and glm-5.3-flash (image + video + pdf per BigModel docs, 2026-08). Text-only defaults for the rest of both families are unchanged, and explicit modalities overrides keep taking precedence.

Why it's needed

Fixes #10270. The documented behavior in docs/users/configuration/settings.md says Qwen Code "automatically detects supported modalities … based on model name pattern matching"; the detection table predates these models, so users who don't know about the override silently lose image input (the image is swapped for a text placeholder with only a debug-level log, and these models are excluded from vision-bridge candidates). The triage on #10270 confirmed the gap and this fix direction.

Reviewer Test Plan

How to verify

1. npx vitest run src/core/modalityDefaults.test.ts (from packages/core) — new cases cover deepseek-v4-flash-vision-exp, deepseek-vl2/deepseek-vl-chat, glm-4.6v, glm-5v-turbo, glm-5.3-flash, plus negative controls (deepseek-v3 v-prefixed text model, glm-5.3 non-flash) that pin the pattern boundaries.
2. Optionally, build from source and run with a custom OpenAI-compatible provider: select deepseek-v4-flash-vision-exp without any modalities override and reference an image — the model should now describe the image instead of receiving the [Unsupported image file: …] placeholder.

Evidence (Before & After)

Before (v0.22.2, no override — image silently replaced by the placeholder, model reports it cannot see it; full transcript evidence on #10270):

After (this PR, local source build, same model entry still without override, non-interactive qwen -m deepseek-v4-flash-vision-exp -p "@<image> 这张图片里是什么内容？请具体描述", DeepSeek official endpoint):

识别出 Qwen Code CLI v0.22.2 深色终端界面、QWEN 蓝粉渐变 logo、状态栏，并正确读出截图内嵌的历史对话内容（"宝蓝色西装外套、白色衬衫未系领带、佩戴眼镜、讲台前对着麦克风发言"）

The session transcript confirms the image entered the request as an inlineData part (both the user-message path and the read_file tool-result path) with no [Unsupported image file placeholder. Same result with glm-5.3-flash via the GLM coding-plan endpoint (model describes the terminal screenshot contents in detail).

Tested on

OS	Status
🍏 macOS	⚠️ not tested
🪟 Windows	✅ tested
🐧 Linux	⚠️ not tested

Environment (optional)

Local source build (npm run build + non-interactive CLI), Node.js ≥22 on Windows; unit tests (60/60 in modalityDefaults.test.ts), ESLint, typecheck, and Prettier pass on the changed files; end-to-end verified against the DeepSeek official API and the Zhipu GLM Coding Plan endpoint.

Risk & Scope

• Main risk or tradeoff: a text-only endpoint that (incorrectly) names a model matching the new patterns would now receive multipart input and may reject it — the same tradeoff fix(core): preserve images for multimodal DeepSeek #9854 accepted (declared/detected capability beats per-family guessing).
• Not validated / out of scope: future vision models without vision/vl/v naming still need new entries (structural maintenance cost of the table, noted on Modality auto-detection doesn't recognize new DeepSeek/GLM vision models — silent image placeholder unless manually overriding modalities #10270); making the placeholder replacement a user-visible warning is tracked as an optional follow-up on Modality auto-detection doesn't recognize new DeepSeek/GLM vision models — silent image placeholder unless manually overriding modalities #10270.
• Breaking changes / migration notes: none — behavior only changes for model ids that currently get silently-downgraded image input.

Linked Issues

Fixes #10270

中文说明

本 PR 的改动

扩展模态自动检测表（packages/core/src/core/modalityDefaults.ts 的 MODALITY_PATTERNS），使新发布的视觉模型无需手动 modalities 覆盖即可被识别：DeepSeek 视觉变体（deepseek-v4-flash-vision-exp 及 vl 后缀命名）、GLM v 后缀家族（glm-4.5v/glm-4.6v/glm-5v-turbo）、以及 glm-5.3-flash（按 BigModel 文档 2026-08 支持 image + video + pdf）。两个家族其余模型的纯文本默认不变，显式 modalities 覆盖依然优先。

为什么需要此改动

修复 #10270。docs/users/configuration/settings.md 文档承诺 Qwen Code 会"基于模型名模式自动检测支持的模态"；检测表早于这些模型，不知道覆盖项的用户会静默丢失图片输入（图片被替换为文本占位符、仅有一条 debug 级日志，且这些模型被排除在 vision bridge 候选之外）。#10270 的 triage 已确认该缺口与修复方向。

审查者测试计划

验证方式

1. 在 packages/core 下运行 npx vitest run src/core/modalityDefaults.test.ts —— 新用例覆盖 deepseek-v4-flash-vision-exp、deepseek-vl2/deepseek-vl-chat、glm-4.6v、glm-5v-turbo、glm-5.3-flash，以及钉住模式边界的负例（deepseek-v3 v 前缀文本模型、glm-5.3 非 flash）。
2. 可选：源码构建，配置自定义 OpenAI 兼容 provider，选择 deepseek-v4-flash-vision-exp 且不写任何 modalities 覆盖，引用一张图片 —— 模型应能直接描述图片，而不是收到 [Unsupported image file: …] 占位符。

证据（修复前后）

修复前（v0.22.2，无覆盖 —— 图片被静默替换为占位符，模型表示看不到；完整 transcript 证据见 #10270）:

修复后（本 PR，本地源码构建，同样不写覆盖，非交互 qwen -m deepseek-v4-flash-vision-exp -p "@<图片> 这张图片里是什么内容？请具体描述"，DeepSeek 官方端点）:

识别出 Qwen Code CLI v0.22.2 深色终端界面、QWEN 蓝粉渐变 logo、状态栏，并正确读出截图内嵌的历史对话内容（"宝蓝色西装外套、白色衬衫未系领带、佩戴眼镜、讲台前对着麦克风发言"）

会话 transcript 确认图片以 inlineData part 进入请求（用户消息路径与 read_file 工具结果路径均验证），[Unsupported image file 占位符完全消失。GLM coding plan 端点 + glm-5.3-flash 同样验证通过（模型详细描述了终端截图内容）。

测试平台

操作系统	状态
macOS	⚠️ 未测试
Windows	✅ 已测试
Linux	⚠️ 未测试

环境（可选）

本地源码构建（npm run build + 非交互 CLI），Windows / Node.js ≥22；改动文件的单测（modalityDefaults.test.ts 60/60）、ESLint、typecheck、Prettier 全部通过；端到端对 DeepSeek 官方 API 与智谱 GLM Coding Plan 端点实测通过。

风险与范围

• 主要风险或取舍：命名命中新模式的纯文本端点现在会收到 multipart 输入并可能拒绝请求 —— 与 fix(core): preserve images for multimodal DeepSeek #9854 的取舍一致（检测到的能力优先于按家族猜测）。
• 未验证 / 不在范围内：未来不带 vision/vl/v 命名的视觉模型仍需补条目（检测表的结构性维护成本，已在 Modality auto-detection doesn't recognize new DeepSeek/GLM vision models — silent image placeholder unless manually overriding modalities #10270 说明）；把占位符替换改为用户可见警告是 Modality auto-detection doesn't recognize new DeepSeek/GLM vision models — silent image placeholder unless manually overriding modalities #10270 记录的可选后续。
• 破坏性变更 / 迁移说明：无 —— 仅影响当前图片输入被静默降级的模型 id。

关联 Issue

Fixes #10270

[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。

[yiliang114]

Thanks for the detailed PR and the before/after evidence — the investigation on #10270 was genuinely helpful.

This one is a duplicate of #10278, which already landed on main and fixed the same gap (the new DeepSeek/GLM vision model patterns in the modality auto-detection table). #10270 was closed as fixed by that merge, about an hour before this PR was opened, so there is nothing left for this PR to fix. Closing in favor of #10278 — if you spot any model pattern that #10278 still misses, a follow-up PR or a comment on #10270 would be welcome.

中文说明

感谢这个 PR 和修复前后的验证证据 —— #10270 上的排查很有价值。

本 PR 与 #10278 重复：#10278 已合入 main，修复了同一个问题（模态自动检测表对新 DeepSeek/GLM 视觉模型的识别）。#10270 已在那次合并后约一小时被关闭（已修复），因此本 PR 没有可修复的剩余问题。现关闭，以 #10278 为准 —— 如果你发现 #10278 仍遗漏了某些模型模式，欢迎提后续 PR 或在 #10270 留言。

[qwen-code-ci-bot]

@jarvislee90s-dot The linked issue #10270 was already fixed before this PR was opened: #10278 landed on main about an hour earlier and extended the same MODALITY_PATTERNS table. Comparing this PR's production diff (only packages/core/src/core/modalityDefaults.ts) against main:

Already on main (via #10278):

• DeepSeek vision naming — main matches /^deepseek-.*vision/, which covers deepseek-v4-flash-vision-exp
• GLM v-suffix family — main matches /^glm-[0-9.]+v/, which covers glm-4.5v / glm-4.6v / glm-5v-turbo and is broader than this PR's variant
• glm-5.3-flash — main already returns image

Remaining delta against main:

1. The vl alternative: your /^deepseek-.*(vision|vl)/ also matches deepseek-vl2 / deepseek-vl-chat, which on main still fall through to text-only.
2. glm-5.3-flash gains video + pdf here; main only grants image.

So this is not a clean duplicate, but most of it is already shipped. Please rebase onto main and reduce this PR to those two deltas — keeping only the tests that pin them (the deepseek-v4-flash-vision-exp, glm-4.6v, and glm-5v-turbo cases pass against main as-is; the deepseek-vl* cases and the glm-5.3-flash video/pdf assertions are the ones that capture the delta) — or close it if you agree nothing is left worth keeping.

中文说明

@jarvislee90s-dot 关联 issue #10270 在本 PR 提交前已经修复：#10278 约一小时前已合入 main，扩展的正是同一张 MODALITY_PATTERNS 表。将本 PR 的生产代码 diff（仅 packages/core/src/core/modalityDefaults.ts）与 main 对比：

已在 main 上（来自 #10278）：

• DeepSeek vision 命名 —— main 的 /^deepseek-.*vision/ 已覆盖 deepseek-v4-flash-vision-exp
• GLM v 后缀家族 —— main 的 /^glm-[0-9.]+v/ 已覆盖 glm-4.5v / glm-4.6v / glm-5v-turbo，且比本 PR 的写法更宽
• glm-5.3-flash —— main 已返回 image

相对 main 的剩余差异：

1. vl 分支：你的 /^deepseek-.*(vision|vl)/ 还能匹配 deepseek-vl2 / deepseek-vl-chat，而它们在 main 上仍会落入纯文本默认。
2. glm-5.3-flash 在本 PR 中额外获得 video + pdf；main 只给了 image。

因此这不是纯粹的重复，但大部分改动已经合入。请 rebase 到 main，把本 PR 缩减到上述两处差异——只保留钉住这两处差异的测试（deepseek-v4-flash-vision-exp、glm-4.6v、glm-5v-turbo 的用例在 main 上本来就能通过；deepseek-vl* 用例和 glm-5.3-flash 的 video/pdf 断言才是真正捕获差异的部分）——或者如果你认为没有剩余价值，也可以直接关闭。

— Qwen Code · qwen3.8-max