Architecture
The runtime exposes OpenAI-compatible endpoints and stores each uploaded image as a persistent visual variable. The original bytes are retained in object storage, while derived views are computed lazily by replaceable lenses and cached by image hash, lens name, lens version, and canonical parameters.
GLM-5.2 is used only for reasoning over compact evidence, plan generation, ambiguous references, and natural-language responses. Pixel analysis, OCR, chart measurement, arithmetic, persistence, and PPTX serialization remain outside the model.
The current MVP uses local JSON/object storage so it can run immediately on a CPU Hugging Face Docker Space. The module boundaries mirror the target PostgreSQL plus S3/MinIO deployment.