Robust tool-use agents must navigate mixed-intent interactions, deciding whether to ask for missing information, answer directly, or invoke tools. We tackle this challenge by introducing a novel task formulation that structures the planner’s action space into explicit interaction instructions over an explicit Ask/Answer/Act space, thereby making these behavioral boundaries an explicit target of optimization. In this framework, a planner predicts the interaction mode and semantic payload, while a frozen executor realizes the instruction into an observable behavior. Coupled with this structured decision process, we propose Executor-Grounded Failure-Routed Planner Learning (EFPL). EFPL optimizes the planner via executor-observed outcomes and strategically leverages privileged feedback for targeted correction when outcome rewards provide insufficient guidance, further refining these behavioral boundaries. Experiments on BFCL and When2Call across three backbone sizes show that EFPL consistently improves clarification, abstention, when-to-call decisions, and robustness to irrelevant tools, while fully preserving strong multi-turn tool-calling performance.