dsh-draw-mode

Links a local ComfyUI server into DeepSeek Harness and gives the agent a drawing mode: five draw_* tools, a curated workflow library, per-model prompt dialects, and a sidebar panel that shows the bridge state.

What it does

• Five tools
  • draw_status — is ComfyUI reachable, what GPU/VRAM does it have, how deep is the queue; optionally start it headlessly.
  • draw_workflows — list the installed workflows with their default parameters.
  • draw_models — model families, checkpoints, default resolution/steps/CFG/sampler, and the LoRA files available for each family.
  • draw_generate — run a workflow: text-to-image, image-to-image, or edit, with LoRA injection and a seed reported back.
  • draw_reverse — reverse-engineer an image into a prompt written in the target family's dialect (WD14 tags for anime models, JoyCaption prose for the others).
• A workflow library for three families — Anima (anime / Danbooru tags), Krea 2 (photographic prose), Qwen-Image 2.1 (general and instruction editing) — including a hand-fix pass and two reverse-engineering graphs.
• An agent preset (画图) that wires a persona, the drawing tools and the shell tools together, so a session can be started directly in drawing mode.
• A sidebar panel listing the bridge state, the GPU, the workflow library and the latest thumbnails. Its data comes from the host half's read-only HTTP route, same origin as the Web GUI.
• Optional auto-start / idle shutdown: start ComfyUI when a tool needs it and stop it again after the queue has been idle, releasing VRAM.

Requirements

• A local (or LAN) ComfyUI instance with the models you intend to use.
• Node >= 22. No @deepseek-ai/* package is imported; the bundle only references host-provided services.

Install

dsh plugin --profile <your-profile> add dsh-draw-mode

Configuration

Set these in your profile's cordis.patch.yml. comfyRoot ships empty on purpose — point it at your own ComfyUI installation if you want auto-start:

- insert:
    - id: dsh-draw
      name: 'dsh-draw-mode'
      config:
        comfyBaseUrl: 'http://127.0.0.1:8188'
        comfyRoot: ''            # e.g. 'D:/ComfyUI' — needed only for autoStart
        autoStart: false         # start ComfyUI on demand
        startTimeoutMs: 240000
        timeoutMs: 900000        # per-generation timeout
        saveToWorkspace: true
        workspaceFolder: 'draw-mode'
        idleShutdownMs: 900000   # 0 disables idle shutdown
        idlePollMs: 30000

Notes

• The bundle also inserts an agent preset; delete that insert block if you only want the tools.
• Prompt dialects differ per family and the preset documents the rules — Anima takes Danbooru tags with a fixed quality prefix, Krea 2 and Qwen take natural-language prose.
• The panel does not require ComfyUI to be running: it reports the disconnected state instead of failing.

中文

把本机 ComfyUI 接进 DeepSeek Harness，并给 agent 一个「画图模式」：五个 draw_* 工具、一套策展好的工作流库、按模型族分化的提示词方言，以及一个显示桥接状态的侧栏面板。

• 五个工具：draw_status（连通性 / GPU 显存 / 队列深度，可无头拉起）、draw_workflows（工作流清单与默认参数）、draw_models（模型族、检查点、默认参数、可用 LoRA）、draw_generate（文生图 / 图生图 / 改图，支持 LoRA 注入并回传 seed）、draw_reverse（把图片反推成目标族的提示词：动漫族出 Danbooru 标签，其余出自然语言长句）。
• 工作流库覆盖三族：Anima（动漫 / Danbooru 标签）、Krea 2（摄影感自然语言）、Qwen-Image 2.1（通用与指令改图），另带手脸修复与两条反推图。
• Agent 预设「画图」把人格、绘图工具与 shell 工具串起来，可以直接用画图模式开会话。
• 侧栏面板显示桥接状态、显卡、工作流库与最近出图缩略图，数据来自 host 半只读 HTTP 路由，与 Web GUI 同源。
• 可选自动启停：需要时拉起 ComfyUI，队列空闲一段时间后自动关掉释放显存。

配置项同上，comfyRoot 默认留空，只有想用自动启动时才需要填你自己的 ComfyUI 路径。

License

MIT