v0.2.2

Fixes

• llm_vision_check testCall probe: the 1×1 probe image was rejected by qwen3-vl-plus's minimum-size rule, so the diagnostic failed with HTTP 400 even when the pipeline was healthy. The probe is now a 64×64 white PNG and the check passes end to end (verified against DashScope qwen3-vl-plus).
• Tool/param descriptions and the testCall success detail now say 64×64 instead of 1×1.

Verification

• 217 offline tests green
• dsh plugin --profile web add dsh-llm-vision@0.2.2 installs from npm and the bundle layer activates

v0.2.0 — free provider presets, batch reads, health check, HEIC

What's new in 0.2.0

Free provider presets — one provider switch fills the whole endpoint config:

• zhipu: free GLM-4V-Flash (open.bigmodel.cn) — best default in mainland China
• gemini: free key from Google AI Studio (generativelanguage.googleapis.com/v1beta/openai/)
• dashscope: the default models as a preset
  Explicit fields always override the preset; existing configs (provider unset = custom) are untouched.

Batch image reads — describe_image accepts up to 8 images via images[] in one call (compare screenshots, spot a shared visual family). Single-image cache keys keep the v1 layout, so previously cached answers still hit. One failed image fails the whole call with a clear error.

llm_vision_check diagnostic tool — verifies config, API-key resolution, and a keyed GET /models probe (401/404/network classified), plus an optional end-to-end testCall on a 1×1 image. The API key never appears in the report.

HEIC/HEIF support — the tools read local HEIC/HEIF paths and URLs directly; macOS preprocessing re-encodes them to JPEG (endpoints support HEIC unevenly), and platforms without sips fail loudly instead of shipping a format endpoints reject. The paste/upload channel stays on the official four media types with a hint to use paths instead.

Also: loadImage file-read errors now carry the llm-vision: prefix (error contract), and the attach body cap scales with maxBytes (large-image pastes no longer hit the fixed 16 MiB ceiling).

216 offline tests (was 183), typecheck + CI green.

dsh-llm-vision v0.1.0

dsh-llm-vision v0.1.0 — 首个 npm 发布

给纯文本模型可靠视觉 + OCR 的 DeepSeek Harness 插件。

工具

• describe_image：normal / critical 双视角图像理解
• extract_text：专用 OCR 模型文字提取

DSH 原生体验

• 粘贴/拖拽图片自动改写为附件引用，会话内联缩略图
• 设置卡（官方白名单限制下引导至 patch 层配置）
• 三种输入：本地路径 / http(s) URL（拒重定向）/ 附件引用

可靠性工程

• macOS sips 自动预处理（超 1568px 缩放、超 1.5MB 重压，失败静默降级）
• 瞬时错误重试（指数退避、预算递减）
• 内容寻址持久缓存（SHA-256，TTL 30 天）

验证

• 真实 DashScope 端点端到端验证（qwen3-vl-plus / qwen3.5-ocr）
• 183 个离线测试，CI 全绿

安装

dsh plugin --profile web add dsh-llm-vision

Apache-2.0，NOTICE 署名 deepseek-harness / dsh-web-ui / llm_vision。