zjcdkj/dsh-plugins--packages-qwen-image ↗★ 1
dsh-plugin-qwen-image
DeepSeek Harness out-of-tree plugin: give a text-only coding model eyes by routing images to a Qwen-VL (DashScope) route through ctx.llm and returning text.
安装
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:zjcdkj/dsh-plugins#bc7362ffdde2079c3f27bbc2b9c65fb7b9450c8c&path:packages/qwen-image说明文档
阅读完整 README ↗Configuration
Installing puts this package in the profile's dsh.profile.bundles, and its own bundle patch already inserted the qwen-image row. To change settings, override that row by id in the profile's cordis.patch.yml:
- id: qwen-image
name: dsh-plugin-qwen-image # optional assertion: a name mismatch skips with a warning
config:
provider: dashscope
model: qwen3-vl-flash
Do not write another insert: — inserting the same id twice fails the boot with duplicate loader entry id. Also config replaces wholesale rather than deep-merging, so spell out every field you want to depart from the defaults.
| Field | Default | Meaning |
|---|---|---|
provider | dashscope | preferred provider id; falls back to any discovered vision route |
model | qwen3-vl-plus | preferred vision model id; falls back the same way |
systemPrompt | see source | system prompt sent to the vision model |
maxOutputTokens | 1024 | output cap for one vision answer |
timeoutMs | 120000 | cooperative timeout budget per call |
provider and model are a preference, not a requirement: naming a route that does not exist costs a log line, not a failure. Name one you actually have when you want a specific model — with several vision routes configured, the scan takes the first registered, which is the deployment's own order and not necessarily the one you meant.