Step 4: run it

# from the directory that contains the minharness/ package
python -m minharness.main /path/to/your/project

With the scripted call_model, you will see the agent take a request, ask to run ls (gated by the approval policy and the hook), feed the result back, and finish with a text message. Replace the body of call_model with a real API call and the same harness becomes a real coding agent: now the model decides what to run, and every safety, memory, context, and budget mechanism you built is already wrapped around it.