Keep interrogating the user until consensus is reached. Map it as adecision tree: each decision branches to the subordinate decisions that depend on it.
Work inrounds.The frontieris all decisions whose prerequisites have been settled—the questions you can ask right now without guessing at answers you have not yet heard. Each round, ask the entire frontier: number each question and give your recommended answer. Then wait for the user's answers before proceeding to the next round.
Each round of user answers reshapes the decision tree—settled decisions push the frontier outward, unlocking the pending questions that depend on them. Recompute the frontier, then proceed to the next round. If a question's answer depends on other questions in this round that have not yet been answered, it belongs toa later round, not the current round.
Discovering facts is your job, never the user's. When a frontier question needs facts obtained from the environment (filesystem, tools, etc.), dispatch a subagent to find them—do not ask the user for anything you can look up yourself. Do not block: an ongoing investigation is an unresolved prerequisite, and only downstream questions wait for the subagent's report—for now, ask the rest of the frontier.Decisionsare the user's—submit each decision to them, then wait.
When the frontier is empty, the session ends: every branch of the decision tree has been traversed, with no unspoken assumptions. Do not take any action until the user confirms that consensus has been reached.