Rajal-ui/InboxOps ? reverse-engineered prompt

Build me a small Python app called InboxOps that works like a deterministic inbox triage benchmark. It should simulate an internal operations inbox where an agent gets a task, sees the business context and risk details, then chooses the best next action from a small fixed set like routing, escalating, replying with a template, or resolving.

I want it to be easy to run locally, easy to test, and easy to deploy in Docker and on Hugging Face Spaces. Please include a simple FastAPI server with endpoints to reset the scenario, take a step, inspect the current state, check metadata, and analyze which action would score best. The scoring should be deterministic, give partial credit for reasonable but not perfect choices, and make bad choices get a tiny score instead of crashing.

Also include a basic inference script that can run through the whole episode as a baseline, plus a few scripts for smoke testing and local validation. Keep it lightweight, clean, and production shaped, with clear models and readable code.