Build a harness

You will have built your own coding agent, the one that drives every repair problem, one small function at a time: four that describe and run tools, three that keep the rules the model cannot be trusted with, three that shape the conversation, and two that make the loop. Each step is a few lines with its own tests, and the earlier functions are provided to the later ones. The twelve compose into one file, `harness.py`, and every step you pass is written into your harness on the harness page as you go. When the last one passes, your harness is complete: check it there, and repair problems run on it. Nothing fancy. Five tools, a refusal to touch the tests, a system prompt that states the rules, a loop that always says why it stopped. It is the harness the simple starter should have been, and the base to grow anything from.