Plan Validation

The problem, from first principles

Language models write plans as structured data: a list of steps, each naming a tool, its arguments and the earlier steps it depends on. Before executing anything, a safe agent checks the plan against what is actually available, because models routinely invent tools that do not exist, omit required arguments, add arguments the tool does not accept and reference steps that come later or never. Catching these statically is cheap, and the error messages are valuable: returned to the model they let it repair its own plan.

From theory to code

Implement validate_plan.

Constraints

• plan is a list of dicts with keys id, tool, args (a dict) and depends_on (a list of step ids). tools maps tool name to {'required': [...], 'optional': [...]}.
• Check steps in plan order and return a list of error strings (empty if the plan is valid). For each step, in this order: if the tool is unknown, add step {id}: unknown tool '{tool}' and skip the argument checks; otherwise add step {id}: missing argument '{a}' for each required argument absent from args (alphabetical), then step {id}: unexpected argument '{a}' for each argument in neither list (alphabetical).
• Then, for each id in depends_on in the listed order: if it is not the id of an earlier step, add step {id}: invalid dependency '{d}'.
• Step ids are unique strings.

Hints