Lab: A sandbox, a courier check and dry runs

Module 10 · Lab · About 80 minutes · Needs: the Module 3 lab's tools.py, and lessons 1, 2 and 3 (run_limited, summarise and code mode; the guarded browser and check_courier_tracking; dry_run, review and apply)

What you are building

This lab moves the three lessons' code into the returns-desk/ repository as three files that import Module 3's tools.py and never replace it. returns_desk/sandbox.py runs model-written Python inside walls and lets scripts call the read tools. returns_desk/courier.py gives R-1005 its courier check. returns_desk/dry_run.py runs the agent against a copy, turns its writes into a plan, and applies only a plan Priya approved. Tests pin every wall. Everything at BrightCart is simulated.

One warning first: a subprocess with limits is not a sandbox. It runs as you, on your kernel, and can read any file you can. The Python shows how each wall behaves; Docker gives real walls, and step 3 gives the exact command.