dsh-tool-imagegen

DeepSeek Harness has no built-in image generation, and dsh-tool-imagegen closes that gap. It exposes a single generate_image tool backed by OpenRouter's unified Image API, letting the model request image creation directly inside a conversation. Results are written to the workspace with an exact USD cost returned, and the generated images render inline in the chat so the model can read them on the next turn and iterate.

The design centers on capability gating: every parameter—resolution, aspect ratio, count, seed, output format, and more—is validated against the model's live capability record fetched from the models listing endpoint rather than hard-coded. An unsupported parameter named in an explicit call produces a clear error that names both the parameter and the model; config defaults the model cannot honor are silently dropped and reported in droppedDefaults. Image-to-image editing works through input_references, accepting workspace paths or http(s) URLs with per-model limits on reference count (up to 14 for Seedream, 16 for the GPT-Image family). Files are identified by magic bytes, never by extension. Aliases serve double duty as the model's vocabulary and as an access allowlist: a model without an alias simply cannot be reached, eliminating silent failures from mistyped names. The Plugins settings card lets you edit the alias table and tunables at runtime with immediate effect and no restart.

It suits Harness users who need frequent image generation or reference-based editing in their workflows, especially those routing through OpenRouter for multi-model image access and who want the model itself to read generated results for multi-turn iteration.