Step 1: the model seam

Everything depends on one function. This is the only place that talks to a real model, so it is the only place you change to go from toy to real.

# minharness/model.py
"""The single seam between the harness and a real model.
Swap the body of call_model for an Anthropic / OpenAI / local call.
It must return one of:
  {"type": "text", "content": "..."}                      -> turn ends
  {"type": "tool_calls", "calls": [{"call_id","name","args"}, ...]}
"""
def call_model(request: dict) -> dict:
    # request has keys: instructions (str), tools (list), input (list)  [Chapter 3]
    #
    # REAL VERSION (Anthropic example, pseudocode):
    #   resp = anthropic.messages.create(
    #       system=request["instructions"], tools=request["tools"],
    #       messages=to_messages(request["input"]), model="claude-sonnet-4-6")
    #   return parse(resp)
    #
    # For a runnable demo without an API key, we script a tiny behavior:
    history = request["input"]
    ran_tool = any(it.get("type") == "tool_result" for it in history)
    if not ran_tool:
        return {"type": "tool_calls", "calls": [
            {"call_id": "c1", "name": "shell", "args": {"command": "ls"}}]}
    return {"type": "text", "content": "Done. I listed the directory."}