Skip to content

v1.2.1

Choose a tag to compare

@github-actions github-actions released this 15 Aug 16:23
· 2 commits to main since this release
1b7470b

v1.2.1

修复 / Fixed

  • 像素工具直接接受附件 IDvision_ground / vision_detect / vision_crop / vision_present / vision_pixel_diff / vision_colors / vision_ocr / vision_long_screenshot_ocr / vision_trace / vision_extract_foreground 的 image 参数现在可以直接传上传图片的附件 ID(如 sha256:…),不再报 cannot read …/sha256:…: not foundvision_describepaths 同样支持。
  • Pixel tools accept attachment ids: the image argument of all ten pixel tools — and vision_describe's paths — now resolves uploaded-image attachment ids like sha256:… directly instead of failing with cannot read …/sha256:…: not found.
  • 产物文件名不再互相覆盖:artifact 文件名在截断的 basename 之外追加完整输入的短指纹,64 字符 sha256 附件名与其裁剪产物不再塌缩成同一个 …-ground.png 互相覆盖。
  • Artifact names no longer collide: stems append a short fingerprint of the full input, so a 64-char sha256 attachment name and its crops no longer collapse onto one …-ground.png and overwrite each other.
  • vision_ground 对退化框(如 1 像素宽)自动重试一次,重试仍退化则明确报错。
  • vision_ground retries once when the vision model returns a degenerate box (e.g. 1px wide) instead of accepting it silently.

改进 / Changed

  • 模型引导可完整重放:「重新查看模型引导」现在先退出设置页(设置面板监听标准 Escape 键关闭)、重新弹出三步总览,并依次走完第 1 步(会话/文字模型)、第 2 步(视觉模型)、第 3 步(高亮视觉链);引导状态改为 v2 步骤标记存储。
  • The model guide replays fully: “re-view model guide” closes the settings modal (its panel closes on the standard Escape keydown), reopens the 3-step overview, and walks step 1 (session/text model), step 2 (vision model) and step 3 (highlighted chain) in order; guide state is stored as a v2 step marker.
  • 设置卡片滚动流畅:长区块滚出视口即跳过绘制(content-visibility),各 provider 的模型选项列表缓存复用,卡片以稳定 props + memo 跳过无关重渲染;openrouter 单 provider 276 个模型下滚动不再卡顿。
  • Smooth settings scrolling: long sections paint only when scrolled into view (content-visibility), per-provider model option lists are cached, and the memoized card skips unrelated re-renders — scrolling stays smooth even with 276 openrouter models.

Commits

  • release: prepare v1.2.1 (#71)
  • fix: attachment-id pixel tools, collision-free artifacts, replayable guide, smooth settings scroll (#70)
  • docs: 挂载 dsh-recommend 认证与分数徽章 (#69)
  • fix: load sharp lazily so a missing/broken sharp cannot crash dsh web (#68)
  • assets: 替换损坏的 Pixel loop in practice 配图(en/zh 实测横幅) (#67)
  • docs: fix duplicated demo and pixel-loop rendering (#66)
  • docs: add QQ user community group (#65)