dsh-tool-imagegen

Working with DeepSeek Harness Desktop often means asking the model to generate an image and then switching to a side panel to view the result, re-uploading it, and re-prompting for every revision. dsh-tool-imagegen collapses that entire loop into the chat itself: the model hears a drawing request, invokes the generate_image tool on its own, and the picture appears as an inline card in the message stream. Hover over the card for full-screen view, download, open-in-system-viewer, and a one-click edit that turns the image into a ready-to-send reference for the next iteration—no external panel, no manual switching.

The standout design is dual-model compatibility. The plugin uses custom generated-image and uploaded-image content blocks instead of the platform-wide image block, so text-only models never hit UNSUPPORTED_CONTENT during generation, upload, or editing, while vision-capable models work just as well—both share the exact same code path with zero extra configuration. Image-to-image requests go through the standard OpenAI-compatible /images/edits endpoint (multipart form-data) and text-to-image through /images/generations (JSON), so any gateway that supports both works out of the box. Iterative editing is genuinely smooth: click edit on any generated image to make it the pending reference, type your new brief, hit send, and get a fresh image; repeat as many times as you like. The model can also directly reference the auto-saved file path in the workspace to call generate_image(image=path), skipping the manual upload entirely.

It also handles the engineering details that usually cause friction: automatic retry with exponential backoff for network blips and 5xx errors (up to three attempts) while correctly skipping 4xx parameter errors; a settings card with endpoint, API key, model, size, count, and six optional parameters (quality, format, background, style, moderation tier, watermark) each exposed as a checkbox plus a free-form dropdown—unchecked parameters are neither sent upstream nor visible to the model; the API key lives only in the local settings.yaml and never enters the browser, with every bridge (settings, attachment, local-open, storage) fenced to loopback. If you use DeepSeek Harness Desktop and want the draw-view-edit cycle to live inside a single conversation—even on a text-only model—this plugin is built for exactly that.