Around the user's given goal, clarify premises and test trade-offs until the key decisions are sufficient to support the next delivery. Use a decision tree to record choices and their dependencies.
Questioning boundaries
- Facts are verified by the Agent; routine implementation choices are judged on their own based on requirements and the project's current state.
- Only hand over to the user the key questions that depend on user preferences, business trade-offs, risk acceptance, or authorization. Reuse answers already confirmed in the current session.
- When called by another process for a specific question, resolve only that question and the prerequisite decisions that affect the answer, then return to the calling process.
Advance in rounds
In each round, ask the questions that can currently be answered and that will change the direction of delivery, explaining the recommended choice and the main trade-offs. Control the number of questions according to actual decision needs, and retain the following question format:
❓ **Q1** - **<问题标题>**: <问题正文，可多段，可含选项>
➡️ <推荐答案>
❓ Q2 - <question title>: <question body, may have multiple paragraphs, may include options>
➡️ <recommended answer>
等待这些决定的回答，同时继续不依赖答案的事实查找。答案依赖本轮其他选择的问题留到下一轮。只有事实调研会阻塞其他独立问题时，才使用子 Agent 并行调研。
用户作出选择后更新决策树，只继续追问仍会影响目标、业务规则或交付边界的分支。
## 完成
关键取舍已明确、能够进入约定的下一步时结束。用户明确要求完整遍历时，遍历约定范围内的决策树；普通调用不以穷尽所有可能性为完成条件。
输出已确认决策及其依赖、影响交付的依据，以及仍需外部输入的未决项。未决项明确标注，不写成用户已确认的选择。