Planning and decomposition

Steps you can check beat steps you can only watch

A model handed "migrate the database and email everyone" will attempt it in one leap and you will have no idea which part failed. Decomposition — asking for a list of steps first, then executing them one at a time — buys three things:

1. Inspectability. You can read the plan before anything happens.
2. Recoverability. Step 4 failing does not mean re-running steps 1 to 3.
3. A stopping condition. A plan has a length; a vibe does not.

The plan is data, so treat it as data: parse it, validate it, and refuse a step that names a tool you never provided — before executing anything, not partway through.

Your turn: write validate_plan(steps, allowed_tools) returning (ok, problem). A plan is bad if it is empty, longer than eight steps, or names an unknown tool. Check the whole plan up front — discovering step 6 is invalid after running five is the failure this prevents.

MAX_STEPS = 8


def validate_plan(steps, allowed_tools):
    """Return (ok, problem). problem is None when the plan is fine."""
    return (True, None)