Multi-agent dialogue: off-topic replies, role confusion, premature discussion end

[cosarah]

Problem

Multi-agent classroom dialogue has three recurring quality issues:

1. Off-topic responses (答非所问) — agents reply without addressing the actual question content.
2. Role confusion between user and student agents — teacher/director treats the human user as if they were a student agent, or attributes user statements to student agents (and vice versa).
3. Director ends discussion prematurely — when the human user asks a substantive question, the director closes the discussion (已结束) instead of routing it to the appropriate agent for a real answer.

Example

Topic: axis of symmetry. User pushes back on whether the Tiananmen gate (a 3D building) can be called an axisymmetric figure.

奇奇:    哇,蝴蝶真的好漂亮!我想到了,咱们的天安门城楼是不是也是轴对称的呀?
林老师:  奇奇观察得真仔细!天安门城楼确实是轴对称的。... 谁来大胆猜猜看?
思思:    肯定可以
你(用户): 天安门城楼是建筑物,为什么也是轴对称图形呢?蝴蝶和剪纸我觉得是对折能够
         完全重合是对称轴对称图形,那那个城楼对折能对折吗?城楼它是三D的怎么办?
思思:    我觉得肯定能重合!只要找到正中间那条对称轴,两边的点就像照镜子一样一对一地对上呢。
林老师:  思思说得真妙,"照镜子"抓住了点对点的本质!虽然建筑是立体的,但在数学里,
         我们观察的是它的正面轮廓。... 严丝合缝地重合啦!
你(用户): 我们当然可以说天安门城楼的照片是轴对称图形,但我们不能说天安门城楼这个物体
         是一个轴对称的吧。
已结束

Issues visible in this trace:

• Off-topic: User's first question targets the 3D-vs-2D distinction. 思思 ignores it and just reasserts symmetry. 林老师's reply pivots to "照镜子" without addressing the dimensionality concern raised by the user.
• Role confusion: 林老师 attributes the "照镜子" insight to 思思, but 思思 only said "肯定能重合" — the framing came from the user's challenge. Teacher conflates whose contribution is whose.
• Premature end: User's second message is a direct correction/objection ("不能说天安门城楼这个物体是一个轴对称的"). Director responds with 已结束 instead of routing to teacher or peer agents for engagement.

Expected behavior

• Agents must address the actual content of the prior turn, not generic topic continuation.
• Director/teacher must distinguish user turns from student-agent turns, and never re-attribute user statements to student agents.
• User questions — especially substantive challenges or objections — should trigger continued discussion, not termination. Termination should be gated on resolution, not on user input arrival.

Suggested investigation

• Review director prompt for end-of-discussion trigger conditions; check whether user turn is being misclassified as a closing signal.
• Review teacher/agent prompts for speaker-attribution rules; ensure user is always tagged distinctly from student agents in the dialogue history fed back to the LLM.
• Check whether agent turns are conditioned on the latest user/agent message specifically, not just the rolling topic.