refactor(agents): 抽取阶段常量与对话上下文；快档 LLM；图片 prompt 可禁止回退

访谈与阶段
- 新增 app/agents/stage_constants.py：集中 CHAT_STAGES、章节分类/顺序、阶段到默认 memoir 类别等，与 MemoirState 默认槽位顺序对齐；减少散落在 prompts 内的重复常量。
- 新增 app/agents/chat/prompt_context.py：以 ChatPromptContext 汇总 guided 系统提示所需字段（阶段、槽位、轮次、人设、记忆证据、回复长度模式、背景声线、职业等），统一走 get_guided_conversation_prompt。
- 大幅收敛 app/agents/chat/prompts_conversation.py；调整 prompts.py、stage_prompts.py、stage_detection.py；同步 interview_agent、profile_agent、helpers 与 state_schema，使对话侧构造提示的方式一致、可测。

回忆录流水线
- memoir/prompts.py 删除已迁至 stage_constants / 独立模板的大段常量与图片占位相关逻辑；classification / extraction / fidelity / narrative agents 与 orchest（全量历史仍可用于计数，注入模型时按轮次与字符上限截断）、image_prompt_fallback_disabled。
- dependencies 增加 get_llm_provider_fast（LRU 缓存，可与默认共用密钥与 base_url）。

任务与编排
- memoir_tasks：prepare_batches 注入 llm_fast；开启独立快档模型时打结构化日志。
- chapter_cover_tasks、story_image_tasks：与图片 prompt / JSON 工具路径或策略变更对齐（import 与行为一致）。
- story_pipeline_sync 等小处同步。

其它核心
- langchain_llm、text_normalize 随上述调用链微调。

开发者体验
- .cursor/settings.json：启用 redis-development、postman 插件。

测试
- 新增 test_image_prompt_policy：覆盖「禁止回退」等图片 prompt 策略。
- 更新 test_interview_prompts、test_interview_reply_length、test_experience_regressions、test_json_and_memory_utils，匹配新常量位置、json_utils 与对话/长度行为。

api/app/agents/chat/interview_agent.py

@@ -3,19 +3,19 @@ InterviewAgent：正式访谈 Specialist
 负责状态感知回复、开场白，不负责 Redis 持久化（由 Orchestrator 统一处理）
 """
 
+import time
 from typing import Any, List, Optional
 
-from app.agents.chat.agent_turn import AgentChatTurn
-from app.agents.chat.stage_detection import keyword_fallback_primary_stage
-from app.core.dependencies import get_llm_provider
-from app.core.logging import get_logger
+from langchain_core.messages import HumanMessage, SystemMessage
 
-from app.agents.chat.helpers import format_history_string, get_history_messages
+from app.agents.chat.agent_turn import AgentChatTurn
+from app.agents.chat.helpers import format_history_string, get_history_with_window
 from app.agents.chat.personas import normalize_interview_persona
+from app.agents.chat.prompt_context import ChatPromptContext
+from app.agents.chat.stage_detection import keyword_fallback_primary_stage
 from app.agents.chat.interview_reply_length import compute_reply_plan
 from app.agents.chat.prompts_conversation import (
     SLOT_NAME_MAP,
-    get_guided_conversation_prompt,
     get_opening_prompt,
 )
 from app.agents.state_schema import MemoirStateSchema
@@ -30,6 +30,8 @@ from app.core.agent_logging import (
     log_agent_summary,
 )
 from app.core.config import settings
+from app.core.dependencies import get_llm_provider
+from app.core.logging import get_logger
 from app.features.conversation.input_normalize import normalize_chat_input_for_agent
 
 logger = get_logger(__name__)
@@ -46,6 +48,15 @@ def _get_langchain_llm():
         return None
 
 
+def _message_contents_char_count(messages: List[Any]) -> int:
+    n = 0
+    for m in messages:
+        c = getattr(m, "content", None)
+        if isinstance(c, str):
+            n += len(c)
+    return n
+
+
 class InterviewAgent:
     """正式访谈 Specialist Agent"""
 
@@ -120,11 +131,13 @@ class InterviewAgent:
                 du = detected_user_stage
             else:
                 du = self._detect_user_stage(text_for_model)
-            history_messages = await get_history_messages(conversation_id)
-            conversation_turn = len(history_messages) // 2
-            same_topic_turns = self._estimate_same_topic_turns(
-                history_messages, filled_slots
+            hw = await get_history_with_window(
+                conversation_id,
+                max_pairs=settings.chat_history_max_pairs,
+                max_chars=settings.chat_history_max_chars,
             )
+            conversation_turn_total = hw.turn_total
+            same_topic_turns = self._estimate_same_topic_turns(hw.window, filled_slots)
             all_stages_coverage = memoir_state.all_stages_coverage()
             persona = normalize_interview_persona(settings.chat_interview_persona)
             reply_plan = compute_reply_plan(
@@ -132,12 +145,12 @@ class InterviewAgent:
                 background_voice=background_voice,
                 settings=settings,
             )
-            system_prompt = get_guided_conversation_prompt(
+            ctx = ChatPromptContext(
                 current_stage=memoir_state.current_stage,
                 empty_slots=empty_slots,
                 filled_slots=filled_slots,
                 user_message=text_for_model,
-                conversation_turn=conversation_turn,
+                conversation_turn_total=conversation_turn_total,
                 same_topic_turns=same_topic_turns,
                 all_stages_coverage=all_stages_coverage,
                 detected_user_stage=du,
@@ -148,19 +161,46 @@ class InterviewAgent:
                 background_voice=background_voice,
                 occupation=occupation,
             )
-            history_string = format_history_string(history_messages)
-            full_prompt = f"{system_prompt}\n\n{history_string}\n\nHuman: {text_for_model}\n\nAssistant:"
+            system_prompt = ctx.guided_system_prompt()
+            messages: List[Any] = [SystemMessage(content=system_prompt)]
+            messages.extend(hw.window)
+            messages.append(HumanMessage(content=text_for_model))
+            history_pairs_windowed = len(hw.window) // 2
+            window_chars = sum(len(getattr(m, "content", "") or "") for m in hw.window)
+            logger.info(
+                "event=history_window_applied total={} windowed={} chars={}",
+                conversation_turn_total,
+                history_pairs_windowed,
+                window_chars,
+            )
             log_agent_payload(
-                logger, "InterviewAgent.generate_response.prompt", full_prompt
+                logger,
+                "InterviewAgent.generate_response.prompt",
+                format_history_string(messages),
             )
             chat_llm = self.llm.bind(max_tokens=reply_plan.max_tokens)
+            prompt_chars = _message_contents_char_count(messages)
+            llm_t0 = time.perf_counter()
             with agent_span(
                 logger,
                 "InterviewAgent.generate_response.llm",
                 conversation_id=conversation_id,
                 stage=memoir_state.current_stage,
             ):
-                response = await chat_llm.ainvoke(full_prompt)
+                logger.info(
+                    "event=chat_prompt_built agent=InterviewAgent.generate_response_with_state "
+                    "prompt_chars={} history_pairs_total={} history_pairs_windowed={}",
+                    prompt_chars,
+                    conversation_turn_total,
+                    history_pairs_windowed,
+                )
+                response = await chat_llm.ainvoke(messages)
+            response_ms = (time.perf_counter() - llm_t0) * 1000
+            logger.info(
+                "event=chat_llm_done agent=InterviewAgent.generate_response_with_state "
+                "response_latency_ms={:.2f}",
+                response_ms,
+            )
             response_text = (
                 response.content if hasattr(response, "content") else str(response)
             )
@@ -218,15 +258,47 @@ class InterviewAgent:
                 background_voice=background_voice,
                 occupation=occupation,
             )
-            full_prompt = f"{prompt}\n\nAssistant:"
-            log_agent_payload(logger, "InterviewAgent.opening.prompt", full_prompt)
+            hw = await get_history_with_window(
+                conversation_id,
+                max_pairs=settings.chat_history_max_pairs,
+                max_chars=settings.chat_history_max_chars,
+            )
+            messages: List[Any] = [SystemMessage(content=prompt)]
+            messages.extend(hw.window)
+            if not hw.window:
+                messages.append(
+                    HumanMessage(content="（对话刚开始，请自然地说出你的开场白。）")
+                )
+            else:
+                messages.append(
+                    HumanMessage(content="（请根据上文，自然接续并说出你的开场白。）")
+                )
+            log_agent_payload(
+                logger,
+                "InterviewAgent.opening.prompt",
+                format_history_string(messages),
+            )
             opening_llm = self.llm.bind(max_tokens=settings.chat_opening_max_tokens)
+            prompt_chars = _message_contents_char_count(messages)
+            llm_t0 = time.perf_counter()
             with agent_span(
                 logger,
                 "InterviewAgent.opening.llm",
                 conversation_id=conversation_id,
             ):
-                response = await opening_llm.ainvoke(full_prompt)
+                logger.info(
+                    "event=chat_prompt_built agent=InterviewAgent.generate_opening_message "
+                    "prompt_chars={} history_pairs_total={} history_pairs_windowed={}",
+                    prompt_chars,
+                    hw.turn_total,
+                    len(hw.window) // 2,
+                )
+                response = await opening_llm.ainvoke(messages)
+            logger.info(
+                "event=chat_llm_done agent=InterviewAgent.generate_opening_message "
+                "response_latency_ms={:.2f}",
+                (time.perf_counter() - llm_t0) * 1000,
+            )
             response_text = (
                 response.content if hasattr(response, "content") else str(response)
             )