wbb316/dsh-comfyui-image

dsh-comfyui-image is a native Cordis plugin built specifically for the DeepSeek Harness (DSH) ecosystem, exposing a single Agent tool named generate_image that bridges a local ComfyUI server into DSH and gives the agent text-to-image, image-to-image, inpainting, and background-removal capabilities. The package declares dsh.engines.dsh >= 0.2.0-rc.1 and depends on DSH's tools registry; installation is performed with one command, dsh plugin --profile <profile> add github:wbb316/dsh-comfyui-image, which the README reports completes in roughly 21.6 seconds and ships 13 ready-to-run files under lib/ with no TypeScript build step required. Alternative mounts include pointing the DSH Settings → Plugins panel at a local link: path or hand-editing the active profile's package.json to add the dependency plus an entry in dsh.profile.bundles, followed by a restart.

The typical workflow is conversational: the user asks the agent to draw, for example, a corgi astronaut, and the plugin auto-infers the text2img mode, runs an SD1.5 checkpoint at 512×512 with seed 42 and dpmpp_2m/karras, then saves the PNG into the current project root and returns an absolute path that DSH can open directly. A matching .json sidecar stores every parameter and the exact ComfyUI workflow for later reproduction, and a custom extra_options.workflow_path token-substitution hook lets users swap in upscalers, outpainting, batch jobs, or any other custom graph without editing plugin source. It targets DSH users who want local image generation without third-party commercial APIs, plus creators who need reproducible prompt-to-image pipelines.

Requirements are Node.js ≥ 20 (only built-in fetch, FormData, and Blob are used, so the runtime carries zero third-party dependencies), a reachable ComfyUI instance (default http://127.0.0.1:8188), and at least one checkpoint under ComfyUI/models/checkpoints/; a CUDA-capable GPU accelerates but is not mandatory since CPU inference still completes. Background removal is the only mode needing extra ComfyUI custom nodes — any one of ComfyUI_essentials, ComfyUI-BRIA_AI-RMBG, ComfyUI-LayerStyle, or WAS Node Suite works, with pip install "rembg[cpu]" explicitly required to avoid the missing onnxruntime trap. Config edits in config.json apply immediately without restarting DSH, existing files are never overwritten (-1, -2 suffixes are appended automatically), the test suite contains 88 cases with a mock ComfyUI that performs real PNG encoding and polling, and the project is released under the MIT license. The plugin is not yet published to npm, so the GitHub shorthand is the only registry-based install path.