v1.2.0

v1.2.0

新增 / Added

• 真正零配置的匿名视觉兜底：内置 OVHcloud 五模型视觉链固定作为最终兜底，免注册、免 Key；用户自己配置的视觉模型仍优先。OVH 匿名额度按每 IP、每模型 2 次/分钟独立计算，五个模型理论合计约 10 RPM，实际以 OVH 限流为准。
• Zero-config anonymous vision fallback: a five-model OVHcloud image chain is now the fixed final fallback with no signup or API key; user-configured vision models still run first. Anonymous OVH buckets are 2 RPM per IP per model, roughly 10 RPM in theory across five independent models, subject to OVH rate limiting.
• 安全的图片展示：新增 vision_present，生成 / 编辑 / 截图得到的图片可作为 DSH 持久附件直接内联展示；刷新后仍可查看，并避免把展示图片塞回纯文本 DeepSeek 请求。
• Safe image presentation: vision_present stores generated/edited/screenshot images as durable DSH attachments for inline display and refresh persistence without feeding display images back into text-only DeepSeek requests.
• 诊断与修复工具：新增 npx dsh-vision-router doctor / repair，可在 DSH 因 profile UTF-8 BOM 无法启动时独立检测并安全移除 BOM。
• Doctor / repair CLI: npx dsh-vision-router doctor and repair can diagnose and safely remove a profile UTF-8 BOM even when DSH cannot boot.
• 版本检查与安全更新：设置页会明确提示发现的新版本；继承的 npm/pnpm registry 失败时自动回退 npm 官方源；自动更新不可用或失败时显示 pnpm / npx 手动命令、项目主页与 Releases。
• Update checks and safe updates: the settings card explicitly reports newer versions, falls back from a broken inherited npm/pnpm registry to npmjs, and shows pnpm/npx manual commands plus Project/Releases links when automatic update is unavailable or fails.

改进 / Changed

• 聊天模型和看图模型彻底分开：聊天页右下角只负责会话模型；设置页「视觉后端链」一行一个用户视觉模型；内部 Vision HTTP 不再暴露给用户，OVH 匿名链固定在最后自动兜底。
• Chat and vision model UX is fully separated: the lower-right chat picker only chooses the conversation model; settings rows choose user vision models; internal Vision HTTP is hidden and the anonymous OVH chain remains the automatic final fallback.
• 首次引导升级为三步模型说明，标题统一为 “Vision Router 已准备好 🎉”；设置卡提示改为更自然的「聊天与看图分别设置」。
• First-run guidance now explains the two model roles in three steps under “Vision Router is ready 🎉”; the permanent settings hint is now “Chat and vision are configured separately.”
• OpenAI-compatible 视觉 HTTP 后端现在保留用户问题，GLM 等模型的输出 token 差异通过集中兼容规则与有限纠错处理。
• OpenAI-compatible vision HTTP calls now preserve the user's question, with centralized compatibility handling for GLM-style output-token limits and bounded corrective retries.

修复 / Fixed

• 修复纯文本 DeepSeek 会话被图片工具结果永久污染：递归清理嵌套 tool-result 中的图片块，保留原始会话附件供 UI / vision_describe 使用，同时确保纯文本适配器永远收不到不支持的 image content。
• Fixed text-only DeepSeek history poisoning from image tool results: nested tool-result image blocks are sanitized before model calls while durable attachments remain available to the UI and vision_describe.
• 匿名 OVH 429 不再在单模型内部长时间等待，立即切换到下一个独立限额模型；Windows/macOS/Linux 的 Chrome/Edge 截图路径发现也更完整。
• Anonymous OVH 429 responses now move immediately to the next independently rate-limited model instead of sleeping inside one bucket; Chrome/Edge discovery for screenshots is also broader across Windows/macOS/Linux.

Commits

• release: prepare v1.2.0 (#64)
• fix: restore ready onboarding title (#63)
• fix: hide internal Vision HTTP and clarify OVH fallback UX (#62)
• feat: guide users to vision model settings (#61)
• docs: prefer doctor repair for BOM startup failures (#59)
• feat: add standalone doctor and BOM repair CLI
• feat: add safe update checks and one-click updates
• docs: add troubleshooting for UTF-8 BOM startup failure
• fix: harden OpenAI-compatible vision providers (#54)
• fix: harden generated-image delivery and screenshot portability
• feat: native-first vision fallback + multi-model OVH rotation
• fix: filter vision backends by image capability (#51)
• feat: add auto-vision onboarding and clearer model guidance (#50)
• fix: unify auto-vision model display names (#47)
• fix: skip Star History workflow in forks
• docs: correct v1.1.1 verified test count (#46)