Replanning After Failure

The problem, from first principles

Plans meet reality. A step fails because a tool is down, a page does not load or an API refuses the request, and a robust agent does not start over or give up. It replans: keep everything already done, replace the failed step's tool with a fallback, and continue with the rest. The remaining steps that depended on completed work no longer need those dependencies, because the results already exist. Implementing this as a pure function on plan data makes the recovery logic testable without running any tools.

From theory to code

Implement replan.

Constraints

• plan is a list of step dicts with keys id, tool and depends_on. completed is a set of finished step ids, failed_id is the id of the step that failed and fallbacks maps a tool name to its fallback tool name.
• Return a new plan containing, in the original order, every step whose id is not in completed. The failed step's tool is replaced by fallbacks[tool]; if it has no fallback, raise ValueError.
• In the returned steps, remove from depends_on every id in completed. Other fields are kept unchanged.
• failed_id must not be in completed. Do not modify the input plan.

Hints