blendsmith

BlendSmith is a model-independent production control loop for Blender agents. It does not replace the model, Blender, MCP, bpy, or GUI automation, but provides production structure by deciding whether real-world knowledge is needed, turning researched facts into practice rules, decomposing the job, exposing method-relevant operations, choosing dedicated Blender methods before scratch work, tracking dependencies, inspecting rendered and live-GUI evidence, classifying what went wrong, and returning to the right upstream decision before continuing. The core idea is that a visible defect does not tell you where the mistake was made, so the right response may be a local edit, a different Blender method, a Production Graph change, an upstream contract revision, or a return to the domain knowledge that produced the requirement. BlendSmith forces that distinction before the agent is allowed to keep patching. It is not a crutch for weak models; capable models already make strong 3D assets, and BlendSmith is the control layer for keeping a capable model on a coherent production path across a long run. The production loop includes a Domain Research Gate, Method Selection, Change Impact Gate, and Exploratory Live GUI Review, and it ends deliberately at AI_ACCEPTED != HUMAN_ACCEPTED, where the agent proposes actions, BlendSmith Core decides which transitions are valid, and the owner decides whether the exact reviewed artifact is accepted.