dsh-vision-subagent
为 DeepSeek Harness 提供基于能力的视觉支持：多模态模型原生支持图像输入，纯文本模型则使用隔离的一次性视觉路由回退方案。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-vision-subagent@0.5.0Vision Subagent 以 dsh-vision-subagent 发布，当前版本为 0.5.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Capability-aware vision for DeepSeek Harness: multimodal current models receive pasted images through DSH's native path; text-only models delegate image reading to a one-shot subagent on a separately configured vision route (MiniMax / Kimi / any OpenAI-compatible provider). On the delegated path, image bytes and the vision model's intermediate context stay out of the main session — only the final text answer comes back.
sh dsh plugin --profile web add /path/to/dsh-vision-subagent
Then edit ~/.dsh/profiles/web/cordis.patch.yml:
`yaml
Restart dsh web, open a new session, and ask: "Look at ~/Desktop/error.png — what is the error?" The model calls vision_agent(images=[...], question=...) on its own.
| Paste & ask | Analyzing | Clean bubble | Lightbox |
|---|---|---|---|
| composer with pasted image thumbnail | analysis progress capsule | bubble: thumbnails plus your words only | lightbox with full analysis |
The Web composer already supports native image blocks. The plugin defaults to capability-aware pasteMode: auto:
sendSession(text, imageIds, mode, signal). The current model sees the original pixels with the full conversation context; no analysis capsule or pre-description is produced.The screenshots above show the delegated fallback path. Its chat bubble contains only your own words plus thumbnails; the analysis lives in the thumbnail lightbox. Need the original bytes later (image editing, pixel-level inspection)? vision_image_fetch materializes the full-fidelity file into .dsh-vision/.
Set pasteMode: delegate to force the isolated route even for multimodal models, or pasteMode: native to bypass the plugin and leave all admission to DSH. On delegated-analysis failure the message is not sent and the composer draft is preserved.
| Provider | baseURL | Vision models | Key env |
|---|---|---|---|
| Kimi (Moonshot) | https://api.moonshot.cn/v1 | k3 / kimi-k3 / moonshot-v1-8k-vision-preview | MOONSHOT_API_KEY |
| MiniMax | https://api.minimaxi.com/v1 | MiniMax-VL-01 | MINIMAX_API_KEY |
| MiniMax CN | (built-in llm-pi-ai minimax-cn route) | MiniMax-M3 | MINIMAX_CN_API_KEY |
If a route already exists in Settings/Models (e.g. kimi-coding, minimax-cn), the plugin config only names provider + model — the key stays in the route's credential reference. The plugin itself never touches secrets.
| Field | Default | Meaning |
|---|---|---|
| enabled | true | Master switch |
| provider / model | '' (dormant) | Vision route; must be set together |
| subagentProvider | spawn | ctx.subagents provider |
| maxDepth | 1 | Absolute delegation-depth cap for the spawned child; 1 lets the vision child run but forbids further delegation (0 would reject the child itself — a top-level agent's child is depth 1) |
| maxImages | 4 | Images per call |
| maxImageBytes | 10 MiB | Per-image byte cap |
| maxPromptChars | 8000 | Question length cap |
| maxOutputChars | 32000 | Returned text truncation |
| maxTokens | 4096 | Output token cap for delegated vision calls; 0 leaves it to the route |
| pasteMode | auto | auto: native for image-capable current models, delegated fallback for text-only; delegate/native force one path |
| allowRemoteUrls | false | Reserved (v0.1 supports local paths only) |
| allowOutsideWorkspace | false | Workspace containment bypass |
| extraAllowedRoots | [] | Extra allowed image roots |
| guidance | '' | Extra instructions appended to the child prompt |
vision_image_fetch writes only under the session workspace's .dsh-vision/ with self-generated content-hashed filenames; no caller-controlled path segment reaches the diskMain model (text-only) └─ vision_agent(images, question) ──┐ │ 1. admission: ext / containment / symlink / byte cap │ 2. ctx.attachments.saveImage → durable content-addressed refs │ 3. ctx.subagents.start('spawn', { agentOptions: {provider, model} }) ▼ One-shot subagent (MiniMax/Kimi vision route, own context) └─ final text ──► main session (only this message enters main context)
The plugin consumes harness services structurally (duck-typed) and is rc-version tolerant. Runtime dependencies: @deepseek-ai/dsh-tools (defineTool) and @deepseek-ai/schemastery (config schema) only.
vision_image_fetch: materialize pasted originals into .dsh-vision/ for editing (v0.3)sh npm install && npm run typecheck && npm test && npm run build
Version 0.4.2 is developed and release-tested against DSH 0.1.1-rc.2. The delegated Web path uses rc.2's atomic batch image admission, forwards submit cancellation, returns the native SubmitOutcome, and releases draft images only after a successful Host admission. DSH runtime packages are optional exact peers rather than ordinary dependencies, preventing a profile install from adding a second Harness runtime copy.
MIT