car-active-planner

Active planner for Common Agent Runtime.

What it does

Composes inference (candidate generation) with static scoring (car-planner) to produce, evaluate, and select the best ActionProposal for a given goal.

Goal / FailureContext
  → generate_candidates()      [N parallel inference calls, strategy-diverse]
  → parse()                    [LLM text → ActionProposal, with JSON repair]
  → planner.rank()             [static verification + cost scoring]
  → best proposal

Also provides ActiveReplanAdapter — a ReplanCallback implementation that generates diverse alternatives when the primary proposal fails at execution.

Where it fits

Sits above car-planner (which is a pure scorer) and below the agent layer. Use car-planner directly when you already have N candidates; use car-active-planner when you also want CAR to generate the candidates via inference.