feat(chat): server-side interview turn plan (mode, anchor slot, snippet)

- Add plan_interview_turn: emotion_first / memoir_push / follow_user_only
- Inject hard directive block at top of guided system prompt
- Pass stage_switched_this_turn from ChatOrchestrator after stage detection
- Log interview_turn_plan for observability; add unit tests

api/app/agents/chat/interview_agent.py

@@ -15,6 +15,7 @@ from app.agents.chat.interview_state_hints import (
     extract_recent_questions,
     update_recent_questions,
 )
+from app.agents.chat.interview_turn_plan import plan_interview_turn
 from app.agents.chat.personas import normalize_interview_persona
 from app.agents.chat.prompt_context import ChatPromptContext
 from app.agents.chat.prompts_conversation import (
@@ -99,6 +100,7 @@ class InterviewAgent:
         occupation: str = "",
         profile_birth_year: int | None = None,
         profile_era_place: str = "",
+        stage_switched_this_turn: bool = False,
     ) -> AgentChatTurn:
         """生成状态感知的访谈回复，不持久化（由 Orchestrator 负责）"""
         if not self.llm:
@@ -133,6 +135,20 @@ class InterviewAgent:
             max_tokens = int(settings.chat_interview_max_tokens)
             max_chars = int(settings.chat_interview_max_chars_per_segment)
 
+            turn_plan = plan_interview_turn(
+                current_stage=memoir_state.current_stage,
+                empty_slots=empty_slots,
+                normalized_user_message=text_for_model,
+                memory_evidence_text=memory_evidence_text,
+                stage_switched_this_turn=stage_switched_this_turn,
+            )
+            logger.info(
+                "event=interview_turn_plan mode={} anchor_slot={} snippet_len={}",
+                turn_plan.mode,
+                turn_plan.anchor_slot_key or "-",
+                len(turn_plan.anchor_snippet or ""),
+            )
+
             ctx = ChatPromptContext(
                 current_stage=memoir_state.current_stage,
                 empty_slots=empty_slots,
@@ -149,6 +165,7 @@ class InterviewAgent:
                 known_facts=memoir_state.known_facts,
                 persona_threads=memoir_state.persona_threads,
                 recent_questions=recent_questions or memoir_state.recent_questions,
+                turn_plan=turn_plan,
             )
             system_prompt = ctx.guided_system_prompt()
             messages: List[Any] = [SystemMessage(content=system_prompt)]

api/app/agents/chat/interview_turn_plan.py

@@ -0,0 +1,198 @@
+"""
+访谈轮次编排（方案 A）：由服务端显式给出 turn_mode / 主槽 / 挂钩摘录，
+减少仅靠长 prompt 软约束时模型「随便问、不往回忆录引」的漂移。
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Literal
+
+from app.agents.chat.prompts_conversation import SLOT_NAME_MAP
+from app.agents.stage_constants import STAGE_SLOT_KEYS
+
+InterviewTurnMode = Literal["emotion_first", "memoir_push", "follow_user_only"]
+
+
+@dataclass(frozen=True)
+class InterviewTurnPlan:
+    """单轮访谈的硬目标（供注入 system prompt 顶部，优先级高于一般性建议）。"""
+
+    mode: InterviewTurnMode
+    anchor_slot_key: str | None
+    anchor_slot_readable: str
+    anchor_snippet: str
+
+
+def primary_empty_slot(stage: str, empty_slots: list[str]) -> str | None:
+    """按 STAGE_SLOT_KEYS 顺序取第一个仍空的槽。"""
+    if not empty_slots:
+        return None
+    order = STAGE_SLOT_KEYS.get(stage, ())
+    for key in order:
+        if key in empty_slots:
+            return key
+    return empty_slots[0]
+
+
+def _strip_scene_hint(memory_evidence_text: str) -> str:
+    raw = (memory_evidence_text or "").strip()
+    if "[场景氛围提示" in raw:
+        raw = raw.split("[场景氛围提示", 1)[0].strip()
+    return raw
+
+
+def extract_anchor_snippet(
+    *,
+    memory_evidence_text: str,
+    user_message: str,
+    max_chars: int = 180,
+) -> str:
+    """优先记忆摘录，其次用户原话（用于追问挂钩，非事实断言）。"""
+    mem = _strip_scene_hint(memory_evidence_text)
+    if mem and len(mem) >= 4:
+        return mem[:max_chars].strip()
+    um = (user_message or "").strip()
+    if len(um) >= 10:
+        return um[:max_chars].strip()
+    return ""
+
+
+_EMOTION_MARKERS: tuple[str, ...] = (
+    "哭",
+    "难受",
+    "委屈",
+    "害怕",
+    "后悔",
+    "恨",
+    "舍不得",
+    "崩溃",
+    "绝望",
+    "心疼",
+    "哽咽",
+    "咽不下",
+    "睡不着",
+    "想哭",
+    "好难",
+    "太难了",
+    "挺不住",
+    "扛不住",
+    "放不下",
+    "意难平",
+)
+
+
+def _is_emotion_heavy(text: str) -> bool:
+    t = (text or "").strip()
+    if not t:
+        return False
+    if any(m in t for m in _EMOTION_MARKERS):
+        return True
+    if len(t) >= 40 and ("！" in t or "!" in t) and (".." in t or "…" in t or "唉" in t):
+        return True
+    return False
+
+
+def plan_interview_turn(
+    *,
+    current_stage: str,
+    empty_slots: list[str],
+    normalized_user_message: str,
+    memory_evidence_text: str,
+    stage_switched_this_turn: bool,
+) -> InterviewTurnPlan:
+    """
+    粗规则（可迭代）：
+    - 情绪浓：先共情，不强推叙述槽搜集问。
+    - 刚切换人生阶段：跟着用户节奏，不做「新阶段问卷首开」。
+    - 当前阶段无空槽：深度跟进，不重启盘点。
+    - 默认：memoir_push，锁一个主槽 + 挂钩摘录。
+    """
+    snippet = extract_anchor_snippet(
+        memory_evidence_text=memory_evidence_text,
+        user_message=normalized_user_message,
+    )
+
+    if _is_emotion_heavy(normalized_user_message):
+        slot = primary_empty_slot(current_stage, empty_slots)
+        readable = (
+            SLOT_NAME_MAP.get(slot, slot or "")
+            if slot
+            else "（情绪优先时可暂不强绑某一槽位）"
+        )
+        return InterviewTurnPlan(
+            mode="emotion_first",
+            anchor_slot_key=slot,
+            anchor_slot_readable=readable,
+            anchor_snippet=snippet,
+        )
+
+    if stage_switched_this_turn:
+        return InterviewTurnPlan(
+            mode="follow_user_only",
+            anchor_slot_key=None,
+            anchor_slot_readable="（刚自然谈到本阶段，先顺着对方语势，勿问卷式首开）",
+            anchor_snippet=snippet,
+        )
+
+    if not empty_slots:
+        return InterviewTurnPlan(
+            mode="follow_user_only",
+            anchor_slot_key=None,
+            anchor_slot_readable="（本阶段主要叙述槽已有素材）请 depth-first：接续画面或情绪线，别重启童年在哪长大式盘点",
+            anchor_snippet=snippet,
+        )
+
+    slot = primary_empty_slot(current_stage, empty_slots)
+    assert slot is not None
+    return InterviewTurnPlan(
+        mode="memoir_push",
+        anchor_slot_key=slot,
+        anchor_slot_readable=SLOT_NAME_MAP.get(slot, slot),
+        anchor_snippet=snippet,
+    )
+
+
+def format_interview_turn_directive_block(plan: InterviewTurnPlan) -> str:
+    """注入 guided prompt 顶部的硬指令块。"""
+    snippet_line = (
+        plan.anchor_snippet
+        if plan.anchor_snippet
+        else "（无可用摘录时，必须从用户本轮原话里抽词作挂钩，禁止编造）"
+    )
+
+    if plan.mode == "emotion_first":
+        mode_rules = (
+            "- **情绪优先**：本轮以承接、并肩与安全感为主，**不要**为推进「叙述槽大纲」而追加信息搜集型追问。\n"
+            "- 若末尾带问，只能是**贴着用户当前情绪或原词**的极轻一句；禁止切到盘点式下一题。\n"
+            "- 参考主槽「"
+            + plan.anchor_slot_readable
+            + "」仅供你心里知道后续方向，**不要**在本轮用问卷口吻硬推该槽。"
+        )
+    elif plan.mode == "follow_user_only":
+        mode_rules = (
+            "- **跟话头**：本轮禁止问卷式首开、禁止重启式盘点；顺着用户刚展开的画面、人物或情绪自然往下。\n"
+            "- 若带问句，最多**一个**，且必须**从用户原词或下面摘录**长出来，禁止空泛「还有吗」。"
+        )
+    else:
+        mode_rules = (
+            "- **回忆推进（memoir_push）**：对用户可见回复中须有**恰好一个**开放式回忆追问，\n"
+            "  且意图明显在补足下面「主追问方向」；问句必须挂住**挂钩摘录**或**用户本轮原词**（二者至少其一）。\n"
+            "- 禁止用日常寒暄替代该追问；仍遵守全篇「最多一个问句」「禁止晚会硬切」。"
+        )
+
+    return f"""## 本轮编排指令（硬规则，优先于后文一般性建议）
+{mode_rules}
+- **主追问方向（叙述槽）**：{plan.anchor_slot_readable}
+- **挂钩摘录**（仅作衔接线索，**不是**用户本轮新说的内容；禁止写成就等于用户刚讲的原话）：{snippet_line}
+"""
+
+
+__all__ = [
+    "InterviewTurnMode",
+    "InterviewTurnPlan",
+    "extract_anchor_snippet",
+    "format_interview_turn_directive_block",
+    "plan_interview_turn",
+    "primary_empty_slot",
+]

api/app/agents/chat/orchestrator.py

@@ -254,12 +254,14 @@ class ChatOrchestrator:
             is_from_voice=is_from_voice,
         )
         state = await get_or_create_state(user_id, db)
+        stage_before = state.current_stage
         detected = await detect_primary_life_stage(
             normalized_user_message,
             state.current_stage,
             self.interview_agent.llm,
         )
-        if detected != state.current_stage:
+        stage_switched_this_turn = detected != stage_before
+        if stage_switched_this_turn:
             state = await switch_stage(user_id, detected, db)
 
         if conversation and conversation.conversation_stage != state.current_stage:
@@ -317,6 +319,7 @@ class ChatOrchestrator:
             occupation=occupation,
             profile_birth_year=profile_birth_year,
             profile_era_place=profile_era_place,
+            stage_switched_this_turn=stage_switched_this_turn,
         )
         recent_questions = prompt_state.recent_questions
         if turn.interview_state_meta and isinstance(turn.interview_state_meta, dict):
@@ -413,6 +416,7 @@ class ChatOrchestrator:
         occupation: str = "",
         profile_birth_year: int | None = None,
         profile_era_place: str = "",
+        stage_switched_this_turn: bool = False,
     ) -> AgentChatTurn:
         """委托 InterviewAgent 生成访谈回复（持久化由调用方负责）。"""
         return await self.interview_agent.generate_response_with_state(
@@ -427,6 +431,7 @@ class ChatOrchestrator:
             occupation=occupation,
             profile_birth_year=profile_birth_year,
             profile_era_place=profile_era_place,
+            stage_switched_this_turn=stage_switched_this_turn,
         )
 
     def detect_user_stage(self, user_message: str) -> str:

api/app/agents/chat/prompt_context.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 from dataclasses import dataclass
 from typing import Dict, List, Optional
 
+from app.agents.chat.interview_turn_plan import InterviewTurnPlan
 from app.agents.state_schema import KnownFact, PersonaThread
 
 
@@ -27,11 +28,20 @@ class ChatPromptContext:
     known_facts: List[KnownFact] | None = None
     persona_threads: List[PersonaThread] | None = None
     recent_questions: List[str] | None = None
+    turn_plan: InterviewTurnPlan | None = None
 
     def guided_system_prompt(self) -> str:
         """用户原话仅以对话历史 + HumanMessage 注入模型。"""
+        from app.agents.chat.interview_turn_plan import (
+            format_interview_turn_directive_block,
+        )
         from app.agents.chat.prompts_conversation import get_guided_conversation_prompt
 
+        directive = (
+            format_interview_turn_directive_block(self.turn_plan)
+            if self.turn_plan is not None
+            else ""
+        )
         return get_guided_conversation_prompt(
             current_stage=self.current_stage,
             empty_slots=self.empty_slots,
@@ -48,4 +58,5 @@ class ChatPromptContext:
             known_facts=self.known_facts or [],
             persona_threads=self.persona_threads or [],
             recent_questions=self.recent_questions or [],
+            turn_directive_block=directive,
         )

api/app/agents/chat/prompts_conversation.py

@@ -229,6 +229,7 @@ def get_guided_conversation_prompt(
     known_facts: list[KnownFact] | None = None,
     persona_threads: list[PersonaThread] | None = None,
     recent_questions: list[str] | None = None,
+    turn_directive_block: str = "",
 ) -> str:
     """生成状态感知的对话提示词；用户原话仅以 HumanMessage 传入，不写入本 system 文本。"""
     persona_key = normalize_interview_persona(persona)
@@ -365,7 +366,9 @@ def get_guided_conversation_prompt(
         current_stage, empty_slots
     )
 
-    return f"""你是「岁月知己」——**主持式知己**：语气像最懂我的老朋友，**职责是帮用户把人生故事口述清楚**。{tone_line}
+    _prefix = f"{turn_directive_block.rstrip()}\n\n" if (turn_directive_block or "").strip() else ""
+
+    return f"""{_prefix}你是「岁月知己」——**主持式知己**：语气像最懂我的老朋友，**职责是帮用户把人生故事口述清楚**。{tone_line}
 
 {topic_desc}
 

api/tests/test_interview_turn_plan.py

@@ -0,0 +1,65 @@
+"""interview_turn_plan：轮次模式与主槽选择（服务端硬编排）。"""
+
+from app.agents.chat.interview_turn_plan import (
+    extract_anchor_snippet,
+    plan_interview_turn,
+    primary_empty_slot,
+)
+
+
+def test_primary_empty_slot_order():
+    assert primary_empty_slot("childhood", ["emotion", "place"]) == "place"
+    assert primary_empty_slot("childhood", ["emotion"]) == "emotion"
+
+
+def test_extract_anchor_snippet_prefers_memory():
+    mem = "摘录的一段记忆\n\n[场景氛围提示"
+    assert "摘录的一段记忆" in extract_anchor_snippet(
+        memory_evidence_text=mem, user_message="用户说很长一句" * 3
+    )
+
+
+def test_plan_memoir_push():
+    p = plan_interview_turn(
+        current_stage="childhood",
+        empty_slots=["place", "people"],
+        normalized_user_message="我小时候住在河边，夏天常去玩水。",
+        memory_evidence_text="",
+        stage_switched_this_turn=False,
+    )
+    assert p.mode == "memoir_push"
+    assert p.anchor_slot_key == "place"
+    assert p.anchor_snippet
+
+
+def test_plan_emotion_first():
+    p = plan_interview_turn(
+        current_stage="childhood",
+        empty_slots=["place"],
+        normalized_user_message="想起来还是很难受，忍不住想哭。",
+        memory_evidence_text="",
+        stage_switched_this_turn=False,
+    )
+    assert p.mode == "emotion_first"
+
+
+def test_plan_follow_on_stage_switch():
+    p = plan_interview_turn(
+        current_stage="education",
+        empty_slots=["school", "city"],
+        normalized_user_message="后来我去省城读中学了。",
+        memory_evidence_text="",
+        stage_switched_this_turn=True,
+    )
+    assert p.mode == "follow_user_only"
+
+
+def test_plan_follow_when_no_empty_slots():
+    p = plan_interview_turn(
+        current_stage="childhood",
+        empty_slots=[],
+        normalized_user_message="嗯。",
+        memory_evidence_text="",
+        stage_switched_this_turn=False,
+    )
+    assert p.mode == "follow_user_only"