Windows CI is currently failing in what appears to be an unrelated UI test:

src/ui/components/InputPrompt.test.tsx > InputPrompt > prompt suggestions > accepts and submits the prompt suggestion on Enter when the buffer is empty

It failed twice on Windows with:

AssertionError: expected "spy" to be called with arguments: [ 'commit this' ]

This PR only changes packages/core/src/models/modelsConfig.ts and packages/core/src/models/modelsConfig.test.ts. The same CI matrix passed on macOS and Ubuntu, and the failing UI test passes locally on Windows in both targeted and full-file runs:

• npm run test --workspace=packages/cli -- src/ui/components/InputPrompt.test.tsx -t "accepts and submits the prompt suggestion on Enter when the buffer is empty"
• npm run test --workspace=packages/cli -- src/ui/components/InputPrompt.test.tsx

I do not have permission to rerun the upstream failed job directly, so I force-pushed the same content once to retrigger CI; the same unrelated Windows UI test failed again.

Updated the PR description to the latest template, including Reviewer Test Plan, Evidence, Tested on, Risk & Scope, and the Chinese summary.

I rechecked the failing Windows job. The only failure is still the unrelated CLI UI test:

src/ui/components/InputPrompt.test.tsx > InputPrompt > prompt suggestions > accepts and submits the prompt suggestion on Enter when the buffer is empty
AssertionError: expected "spy" to be called with arguments: [ 'commit this' ]
Number of calls: 0

This PR only changes packages/core/src/models/modelsConfig.ts and packages/core/src/models/modelsConfig.test.ts; macOS and Ubuntu CI passed, and I previously verified the failing UI test locally on Windows in both targeted and full-file runs.

Rebased this branch onto current origin/main and force-pushed with lease to retrigger CI after the unrelated Windows UI-test failure.

Local validation on the rebased head c5159e18b:

• npm run test --workspace=packages/core -- src/models/modelsConfig.test.ts -t "raw model|explicitly configured modalities"
• npm run test --workspace=packages/core -- src/models/modelsConfig.test.ts
• npm run test --workspace=packages/core -- src/core/openaiContentGenerator/converter.test.ts -t "image modality|mixed content|defaults to text-only"
• npx prettier --check packages/core/src/models/modelsConfig.ts packages/core/src/models/modelsConfig.test.ts
• npx eslint packages/core/src/models/modelsConfig.ts packages/core/src/models/modelsConfig.test.ts
• npm run lint --workspace=packages/core
• npm run build --workspace=packages/core
• npm run typecheck --workspace=packages/core

Note: I hit the known local coverage .tmp race when running core vitest commands concurrently, then reran the affected converter test serially and it passed.

🔍 PR #4517 验证报告

概要

问题描述

从 multimodal provider model（支持 image/video）切换到 raw text-only model（如 qwen3.7-max）时，旧 provider 的 multimodal 配置会残留在 model config 中，导致 OpenAI converter 继续向 text-only model 发送 image parts。

修复方案

1. 重新计算 derived defaults: 切换到 raw model ID 时重新计算 model-derived defaults
2. 触发 owner refresh: raw model 更新会触发 owner refresh callback，让 content-generator 使用刷新后的 model config
3. 保留显式配置: 保留 explicitly configured modalities
4. 失败回滚: owner refresh 失败时回滚 raw model state

静态验证

单元测试 (69 tests)

关键测试覆盖：

• Raw model 切换后 modalities 被正确重置为 text-only
• Explicitly configured modalities 被保留
• Owner refresh callback 在 raw model 切换后触发
• Image modality 相关的 converter 行为正确

TUI 真实用户测试 (tmux)

使用 tmux 模拟真实用户在 --approval-mode yolo 下的完整交互流程：

关键验证点：

• ✅ /model 对话框正确显示 Modality: text-only，证明 PR fix(models): refresh raw model-derived defaults #4517 的修复生效
• ✅ qwen3.7-max 作为 text-only raw model 正常工作
• ✅ 没有继承 multimodal provider 的 image/video capabilities

核心修复验证

Before PR #4517：

从 multimodal provider (image: true, video: true) 切换到 qwen3.7-max
→ modalities 残留: { image: true, video: true }
→ OpenAI converter 向 text-only model 发送 image parts ❌

After PR #4517：

从 multimodal provider 切换到 qwen3.7-max
→ modalities 重新计算: { text: true }
→ Modality: text-only ✅
→ Converter 正确处理为 text-only content

代码变更分析

packages/core/src/models/modelsConfig.ts (+55 / -6)：

1. setModel() 方法增加 raw model 检测逻辑
2. Raw model 切换时调用 computeModelDerivedDefaults() 重新计算 defaults
3. 触发 ownerRefreshCallback，让 content-generator 使用刷新后的 config
4. 保留 explicitly configured modalities（用户显式设置的优先级最高）
5. Owner refresh 失败时回滚 raw model state

packages/core/src/models/modelsConfig.test.ts (+118 / -1)：

• 新增 4 个测试用例覆盖 raw model 切换场景
• 测试 modalities 清理、owner refresh 调用、失败回滚
• 测试 explicitly configured modalities 的保留

风险评估

结论

✅ PASS — 推荐 merge

PR #4517 成功修复了 raw model 切换时 multimodal 配置残留的问题：

• 单元测试: 69 个相关测试全部通过
• 静态检查: typecheck / lint / prettier 全部通过
• TUI 测试: qwen3.7-max 正确显示 Modality: text-only
• 代码质量: 实现清晰，测试覆盖充分，包含失败回滚逻辑
• 向后兼容: 保留 explicitly configured modalities

此修复解决了 #4513 中报告的问题，确保 text-only raw models 不会继承 multimodal provider 的 capabilities。

日志产物: tmp/pr4517-model-tmux-20260527-083953/tmux-readable-full.log (208 行)