refactor(agent): inject deps via context_schema, LLM-driven photo/confirm nodes

- graph.py: LangGraph-native DI - StateGraph(context_schema=GraphDeps),
  nodes take (state, runtime: Runtime[GraphDeps]); build_graph() decoupled
  from LLM; deleted scaffold _event_reply/_fake_llm_propose/stub nodes
- photo.py: model proposes stage+body via per-mode prompts (direct/addon);
  validator enforces clientMode exits; deterministic event forcing removed
- confirm.py: single runtime for single/dual confirm, extraction call
  isolated per person_index
- llm.py: workflow_fake_llm() (offline workflow-faithful responder,
  history-aware) for smoke/tests
- backend: LangGraphBackend(graph, llm) passes context per call
- tests: invoke helpers inject GraphDeps; illegal-proposal test asserts
  error + state untouched via second well-behaved fake

Verified: 143 passed; smoke turns 1/2/1 with original welcome text

examples/langgraph_smoke.py

@@ -1,28 +1,31 @@
-"""LangGraph 脚手架冒烟脚本（作业 1）。
+"""LangGraph 脚手架冒烟脚本。
 
 同一 thread 输入两次、另一个 thread 输入一次，打印轮数和文本。
-预期输出轮数依次为 1、2、1。
+预期轮数依次为 1、2、1；首轮为 welcome（1001 固定问候，不调模型），
+第二轮起由注入的 workflow fake LLM 按提示词语义回复。
 
 本脚本不导入 FastGPT、不读取其密钥；离线即可运行：
 
-    python examples/langgraph_smoke.py
+    python -m examples.langgraph_smoke
 """
 
 import asyncio
 
-from src.agent.graph import build_graph
+from src.agent.graph import GraphDeps, build_graph, workflow_fake_llm
 
 
 async def main() -> None:
     graph = build_graph()
+    deps = GraphDeps(llm=workflow_fake_llm())
 
     async def ask(thread_id: str, text: str) -> None:
-        # 只传本轮变化的字段；会话字段（turn_count 等）由 checkpoint 保管。
+        # 只传本轮变化的字段；会话字段（turn_count/stage_code 等）
-        # 注意：对已有 checkpoint 的 thread 传入完整初始状态，
+        # 由 checkpoint 保管。注意：对已有 checkpoint 的 thread 传
-        # 其中的 turn_count=0 会覆盖已累计的值。
+        # 完整初始状态会覆盖已累计的值。
         result = await graph.ainvoke(
             {"session_id": thread_id, "input_text": text},
             config={"configurable": {"thread_id": thread_id}},
+            context=deps,
         )
         print(
             f"thread={thread_id} turn_count={result['turn_count']} "

src/agent/confirm.py

@@ -0,0 +1,154 @@
+"""信息核实节点（作业 5，LLM 决策版）。
+
+设计：提示词即程序——核实链的顺序、话术、阈值、全局中断全部
+在 ``prompts/single_confirm.md`` / ``dual_confirm.md`` 中（与工作流
+chatNode 提示词一致），代码不维护状态链。
+
+节点职责（薄运行时）：
+1. 把上下文注入提示词：已核实字段（person_fields，模型据此跳过
+   已确认项）、当前人员序号（person_index，双车两方隔离）、
+   对话历史、无回复计数；
+2. 一次 LLM 调用产出 ``<state>`` 前缀回复；前缀缺失/非法提议
+   受控修复一次（复用 collect 的语义），仍失败抛 CollectReplyError；
+3. 字段提交走**独立提取调用**（计划作业 4-4 允许，需记录次数）：
+   从本轮输入提取语义明确的新字段，validate_candidate 校验后
+   按 person_index 隔离写入 person_fields / turn_patch；
+4. 阶段码一律经白名单校验（3001/3002 → 0000/0001/0002/0003/0004）。
+
+模型决定“问什么、去哪一步”；代码决定“能不能去、字段能不能写”。
+"""
+
+import json
+
+from .collect import (
+    _bookkeeping,
+    context_prompt,
+    conversation_history,
+    validated_reply,
+)
+from .fields import FieldCandidate, FieldStatus, validate_candidate
+from .llm import LLMService
+from .prompts import DUAL_CONFIRM, SINGLE_CONFIRM, render_prompt
+from .state import AgentState
+
+# 提取调用的系统提示标记（测试与分流依据；不以 <state> 开头，
+# 故不会被误当对话回复解析）
+_EXTRACTION_MARKER = "【字段提取】"
+
+_EXTRACTION_PROMPT = (
+    _EXTRACTION_MARKER
+    + "从【用户输入】中提取本次新提供且语义明确的当事人信息，"
+    "只输出 JSON 对象（键限于给定字段表，值用原文；无新字段输出 {{}}）。"
+    "模糊、破碎、待确认的输入不提取。reserved 键 person_index（1 或 2）"
+    "仅在对话明确切换到另一方当事人时给出。\n"
+    "字段表（当前人员）：{keys}\n"
+    "【用户输入】{user_input}\n"
+    "【模型回复】{reply}"
+)
+
+# 各方人员字段（来源：endpoints.get_info 分组；与 fields.KNOWN_FIELD_KEYS 对齐）
+_PERSON1_KEYS = ("xm1", "sfzmhm1", "sfzmwh1", "sjhm1", "sjwh1", "csbw1")
+_PERSON2_KEYS = ("xm2", "sfzmhm2", "sfzmwh2", "sjhm2", "sjwh2", "csbw2")
+
+
+def _person_keys(person_index: int) -> tuple[str, ...]:
+    return _PERSON1_KEYS if person_index != 2 else _PERSON2_KEYS
+
+
+async def _extract_turn_fields(
+    state: AgentState, llm: LLMService, reply: str
+) -> tuple[list[FieldCandidate], int | None]:
+    """独立提取调用：返回 (候选列表, 模型声明的 person_index 或 None)。
+
+    调用次数由调用方计入（计划作业 4-4：记录原因/次数/延迟）。
+    """
+    person_index = state.get("person_index") or 1
+    raw = await llm.complete(
+        _EXTRACTION_PROMPT.format(
+            keys="、".join(_person_keys(person_index)),
+            user_input=state.get("input_text", ""),
+            reply=reply,
+        ),
+        "",
+    )
+    try:
+        data = json.loads(raw.strip())
+        if not isinstance(data, dict):
+            return [], None
+    except json.JSONDecodeError:
+        return [], None
+    next_person = data.pop("person_index", None)
+    if next_person not in (1, 2):
+        next_person = None
+    candidates = [
+        FieldCandidate(key=key, value=value, source_turn=state.get("turn_count") or 0)
+        for key, value in data.items()
+        if key in _person_keys(person_index)
+    ]
+    return candidates, next_person
+
+
+async def _run_confirm(state: AgentState, llm: LLMService, prompt_name: str) -> dict:
+    """核实节点共享运行时（单车/双车仅提示词不同）。"""
+    system_prompt = render_prompt(prompt_name, state) + context_prompt(state)
+    user_message = state.get("input_text", "")
+    current = state.get("stage_code") or "3001"
+    raw, stage = await validated_reply(
+        llm, system_prompt, user_message, current, history=conversation_history(state)
+    )
+
+
+    # 字段提取与按人员隔离提交（提取调用本身可能失败，不阻塞对话）
+    person_fields = dict(state.get("person_fields") or {})
+    confirm_status = dict(state.get("confirm_status") or {})
+    patch: dict = {}
+    new_person_index = state.get("person_index") or 1
+    try:
+        candidates, declared = await _extract_turn_fields(state, llm, raw)
+    except Exception:
+        candidates, declared = [], None
+    for cand in candidates:
+        validated = validate_candidate(cand)
+        confirm_status[validated.key] = validated.status
+        if validated.status is FieldStatus.CONFIRMED:
+            person_fields[validated.key] = validated.value
+            patch[validated.key] = validated.value
+    if declared is not None:
+        new_person_index = declared
+
+    no_reply_count, _clarify = _bookkeeping(state)
+    # clarify_count 的“计入”是模型按提示词（连续 4 次）判断的语义，
+    # 代码侧只做簿记透传，不用 collect 的近似增量
+    clarify_count = state.get("clarify_count") or 0
+    dialogue = list(state.get("dialogue") or [])
+    dialogue.append({"role": "user", "content": state.get("input_text", "")})
+    dialogue.append({"role": "assistant", "content": raw})
+
+    return {
+        "turn_count": (state.get("turn_count") or 0) + 1,
+        "no_reply_count": no_reply_count,
+        "clarify_count": clarify_count,
+        "response_text": raw,
+        "stage_code": stage,
+        "raw_status_code": stage,
+        "person_fields": person_fields,
+        "confirm_status": confirm_status,
+        "turn_patch": patch,
+        "person_index": new_person_index,
+        "dialogue": dialogue,
+    }
+
+
+async def run_single_confirm(state: AgentState, llm: LLMService) -> dict:
+    """单车信息确认节点（提示词：prompts/single_confirm.md）。"""
+    return await _run_confirm(state, llm, SINGLE_CONFIRM)
+
+
+async def run_dual_confirm(state: AgentState, llm: LLMService) -> dict:
+    """双车信息确认节点（提示词：prompts/dual_confirm.md）。
+
+    与单车共用运行时；两方的区分靠注入的 person_fields 与
+    person_index，提取调用的字段表按当前人员隔离，模型声明切换
+    到另一方（person_index）时由代码应用。
+    """
+    return await _run_confirm(state, llm, DUAL_CONFIRM)

src/agent/graph.py

@@ -1,19 +1,19 @@
 """LangGraph 事故流程图。
 
-架构（对照 ``workflow/20261016/事故信息采集2025118.json``，作业 3 修订）：
+架构（对照 ``workflow/20261020/事故信息采集2025118.json``）：
 
-- 每个主要阶段是独立节点，与工作流 chatNode 一一对应：
+- 每个主要阶段一个节点，与工作流 chatNode 一一对应：
-  ``collect``（事故信息收集）、``single_photo``（单车拍照）、
+  ``collect``（事故信息收集）、``single_photo`` / ``dual_photo``
-  ``dual_photo``（双车拍照）、``single_confirm`` / ``dual_confirm``
+  （拍照）、``single_confirm`` / ``dual_confirm``（信息确认）；
-  （信息确认，作业 5 实现）；
+- 节点间大切换由 ``route_by_stage`` 条件边按 ``stage_code`` 分发
-- 节点间的大切换由 ``route_by_stage`` 条件边按 ``stage_code`` 路由
+  （对照 Condition#2/#7/#9 等）；一轮 HTTP 请求 = 一次图执行，
-  （对照工作流的 Condition#2/#7/#9 等）；LLM 接入后由节点内的
+  跨轮记忆由 checkpoint 承载；
-  “解析状态”代码写出 ``stage_code``，条件边只做分发，不含业务规则；
+- 节点内阶段码一律由 LLM 提议（提示词即程序），代码只校验
-- 节点内部的阶段推进（拍照步骤、失败/无回复处理、分流）由节点内
+  （前缀格式、白名单、clientMode 出口）与落库；
-  LLM 提议目标 ``stage_code``（脚手架期由 ``_fake_llm_propose``
+- 依赖注入走 LangGraph 原生机制：``context_schema=GraphDeps``，
-  模拟），代码按 ``transitions`` 白名单校验、非法提议不写状态；
+  节点签名 ``(state, runtime: Runtime[GraphDeps])``，调用时
-  代码只负责事件解析与计数等上下文簿记，不代替模型决定；
+  ``context=GraphDeps(llm=...)``。LLM client 属于运行时依赖，
-- LLM 接入前（作业 4 之前），节点共用固定话术模板。
+  不进 state、不进 checkpoint（计划作业 3-2）。
 
 图结构::
 
@@ -21,40 +21,43 @@
                                      ├→ collect ───────────→ END
                                      ├→ single_photo ──────→ END
                                      ├→ dual_photo ────────→ END
-                                     ├→ single_confirm ────→ END  (作业 5)
+                                     ├→ single_confirm ────→ END
-                                     ├→ dual_confirm ──────→ END  (作业 5)
+                                     ├→ dual_confirm ──────→ END
                                      └→ END                        (终态 0000—0005)
 
-首轮语义（对照工作流 work_start 的 is_first_message → Assigned Reply#3）：
+离线运行示例（默认注入 workflow fake LLM）：
-新会话第一条消息由 ``welcome`` 节点以固定话术返回 ``1001``
-（未准备好通话），不调用模型；之后用户说【继续办理】或继续对话，
-由 collect 节点的 LLM 提议 ``1001 → 1002``（白名单已含该转换）。
-
-离线运行示例：
 
     python -m examples.langgraph_smoke
 
 对应测试：test/agent/test_scaffold.py、test/agent/test_events.py、
-test/agent/test_transitions.py
+test/agent/test_transitions.py、test/agent/test_photo.py、
+test/agent/test_confirm.py
 """
 
+from dataclasses import dataclass
+
 from langgraph.checkpoint.memory import InMemorySaver
 from langgraph.graph import END, START, StateGraph
+from langgraph.runtime import Runtime
 
-from .events import ClientEvent, parse_client_event
+from .collect import run_collect
+from .confirm import run_dual_confirm, run_single_confirm
+from .events import parse_client_event
+from .llm import LLMService, workflow_fake_llm
+from .photo import run_photo
 from .state import AgentState
-from .transitions import TERMINAL_STAGES, is_legal_transition
+from .transitions import TERMINAL_STAGES
 
-# 缺省阶段（首轮 welcome 之后的兜底）；后续由规则/模型驱动。
-DEFAULT_STAGE_CODE = "1002"
 
-# 首轮固定问候（对照工作流 Assigned Reply#3 + Variable Update#6）：
+@dataclass
-# 新会话首条消息返回 1001，不调用模型。
+class GraphDeps:
-WELCOME_STAGE = "1001"
+    """图节点的运行时依赖。
-WELCOME_TEXT = (
+
-    "<state>1001</state>您好，请问您是来办理交通事故信息采集的吗？"
+    通过 ``context_schema`` 注入（``ainvoke(..., context=GraphDeps(...))``），
-    "请说“继续办理”开始。"
+    只存在于调用期，不序列化、不进入 checkpoint。
-)
+    """
+
+    llm: LLMService
 
 
 def normalize_input(state: AgentState) -> dict:
@@ -99,157 +102,79 @@ def route_by_stage(state: AgentState) -> str:
     return "collect"
 
 
-def welcome(state: AgentState) -> dict:
+# 首轮固定问候（对照工作流 Assigned Reply#3）：新会话首条消息返回
-    """首轮问候节点：新会话第一条消息的固定话术（对照工作流
+# 1001，不调用模型；话术与 Variable Update#6 后的记录一致。
-    work_start is_first_message → Assigned Reply#3）。
+WELCOME_STAGE = "1001"
+WELCOME_TEXT = (
+    "<state>1001</state>您好，这里是无锡交警，我将为您远程处理交通事故。"
+    "请将人员撤离至路侧安全区域，开启危险报警双闪灯、放置三角警告牌、"
+    "做好安全防护，谨防二次事故伤害。若您已经准备好了，请点击继续办理，"
+    "如需人工服务，请说转人工。"
+)
 
-    确定性节点，不调用模型；返回 1001（未准备好通话），
+
-    用户确认后由 collect 节点的 LLM 经白名单守卫进入 1002。
+def welcome(state: AgentState) -> dict:
+    """首轮问候节点（确定性，不调用模型）。
+
+    返回 1001（未准备好通话），并把问候写入 dialogue，
+    供后续节点作为聊天历史注入（对齐 FastGPT 的 chatId 历史）。
     """
     return {
         "turn_count": (state.get("turn_count") or 0) + 1,
         "response_text": WELCOME_TEXT,
         "stage_code": WELCOME_STAGE,
         "raw_status_code": WELCOME_STAGE,
+        "dialogue": [
+            {"role": "user", "content": state.get("input_text", "")},
+            {"role": "assistant", "content": WELCOME_TEXT},
+        ],
     }
 
 
-def _fake_llm_propose(state: AgentState, event: ClientEvent | None) -> str | None:
+async def collect_node(state: AgentState, runtime: Runtime[GraphDeps]) -> dict:
-    """假 LLM：提出下一步阶段码（作业 4 起由真实 LLM 取代）。
+    """事故信息收集节点（工作流「事故信息收集」chatNode）。"""
-
+    return await run_collect(state, runtime.context.llm)
-    模拟工作流 chatNode 提示词的判断（来源见各分支注释）：
-    - 会话标记且当前 1001 → 1002（work_start is_first_message）；
-    - 【拍摄完成】拍照链前进一步（单车/双车拍照提示词）；
-    - 拍照失败标记 → 0005（客户端已统计三次，提示词要求直接转人工）；
-    - 连续第二次【用户无回复】 → 0004（no_reply_count 由代码簿记，
-      本函数读取的是本轮计数前的历史值）。
-
-    提议是否生效由 ``is_legal_transition`` 校验。
-    """
-    current = state.get("stage_code") or ""
-    if event in (ClientEvent.SESSION_START, ClientEvent.CONTINUE):
-        if current == "1001":
-            return "1002"
-    if event is ClientEvent.PHOTO_DONE:
-        chain = {
-            "2000": "2001", "2001": "2002", "2002": "2003",
-            "2003": "2005",
-            "2010": "2011", "2011": "2012", "2012": "2013",
-            "2013": "2014", "2014": "2015",
-        }
-        return chain.get(current)
-    if event is ClientEvent.PHOTO_FAILED_3X:
-        return "0005"
-    if event is ClientEvent.NO_REPLY and (state.get("no_reply_count") or 0) >= 1:
-        return "0004"
-    return None
 
 
-def _event_reply(state: AgentState) -> dict:
+async def single_photo_node(state: AgentState, runtime: Runtime[GraphDeps]) -> dict:
-    """事件簿记 + 阶段提议/校验 + 固定话术（作业 3 脚手架版）。
+    """单车拍照节点（工作流「单车拍照」/「单车拍照#6」chatNode）。"""
-
+    return await run_photo(state, runtime.context.llm, kind="single")
-    - 代码只维护上下文：``no_reply_count`` 累计（NO_REPLY）与
-      重置（其他输入）；
-    - 阶段码（含失败/无回复处理）一律由 ``_fake_llm_propose`` 提议
-      （作业 4 换真实 LLM），经白名单校验，非法提议保持当前阶段；
-    - 轮数 +1，响应带 ``<state>`` 前缀（API 层解析）。
-
-    TODO(作业 4): 真实 LLM 接入后，话术与阶段提议移入各节点；
-    TODO(作业 4-10): 对非法提议做一次受控修复重试（当前直接拒绝）。
-    """
-    raw_event = state.get("client_event") or ""
-    event = ClientEvent(raw_event) if raw_event else None
-
-    no_reply_count = state.get("no_reply_count") or 0
-    if event is ClientEvent.NO_REPLY:
-        no_reply_count += 1
-    elif (state.get("input_text") or "").strip():
-        no_reply_count = 0
-
-    current = state.get("stage_code") or DEFAULT_STAGE_CODE
-    proposed = _fake_llm_propose(state, event)
-    if proposed and is_legal_transition(current, proposed):
-        stage_code = proposed
-    else:
-        stage_code = current
-
-    response_text = (
-        f"<state>{stage_code}</state>"
-        f"已收到：{state['input_text']}"
-    )
-    return {
-        "turn_count": (state.get("turn_count") or 0) + 1,
-        "no_reply_count": no_reply_count,
-        "response_text": response_text,
-        "stage_code": stage_code,
-    }
 
 
-def collect(state: AgentState) -> dict:
+async def dual_photo_node(state: AgentState, runtime: Runtime[GraphDeps]) -> dict:
-    """事故信息收集节点（工作流“事故信息收集” chatNode）。
+    """双车拍照节点（工作流「双车拍照」/「双车拍照#7」chatNode）。"""
-
+    return await run_photo(state, runtime.context.llm, kind="dual")
-    TODO(作业 4): 接入 LLM——多轮采集、跳过已知信息、一次补齐分流；
-    模型回复中的状态前缀由代码校验（解析状态，对照工作流
-    “解析状态” code 节点），不由模型自由写阶段。
-    """
-    return _event_reply(state)
 
 
-def single_photo(state: AgentState) -> dict:
+async def single_confirm_node(state: AgentState, runtime: Runtime[GraphDeps]) -> dict:
-    """单车拍照节点（工作流“单车拍照” chatNode）。
+    """单车信息确认节点（工作流「单车信息确认」chatNode）。"""
-
+    return await run_single_confirm(state, runtime.context.llm)
-    TODO(作业 5): 按 2000→2001→2002→2003→2005 生成对应拍照话术；
-    步骤推进的确定性部分已在 ``transitions`` 表覆盖。
-    """
-    return _event_reply(state)
 
 
-def dual_photo(state: AgentState) -> dict:
+async def dual_confirm_node(state: AgentState, runtime: Runtime[GraphDeps]) -> dict:
-    """双车拍照节点（工作流“双车拍照” chatNode）。
+    """双车信息确认节点（工作流「双车信息确认」chatNode）。"""
-
+    return await run_dual_confirm(state, runtime.context.llm)
-    TODO(作业 5): 按 2010→…→2015 生成对应话术；direct 模式末次
-    拍摄完成进入 3002（同轮询问第一方身份）。
-    """
-    return _event_reply(state)
 
 
-def single_confirm(state: AgentState) -> dict:
+def build_graph(checkpointer=None):
-    """单车信息确认节点（工作流“单车信息确认” chatNode，作业 5）。"""
-    raise NotImplementedError("TODO(作业 5): 实现单车信息确认节点")
-
-
-def dual_confirm(state: AgentState) -> dict:
-    """双车信息确认节点（工作流“双车信息确认” chatNode，作业 5）。"""
-    raise NotImplementedError("TODO(作业 5): 实现双车信息确认节点")
-
-
-def build_graph(checkpointer=None, llm=None):
     """装配并编译图。
 
     参数:
         checkpointer: 自定义 checkpointer；缺省 ``InMemorySaver``，
             作业 7 替换为 ``AsyncPostgresSaver``。
-        llm: LLM 服务（``llm.LLMService``）；提供后 collect 节点
-            走 ``collect.run_collect``（作业 4），否则保持离线
-            脚手架行为（``_event_reply``，作业 1—3 的测试依赖）。
 
-    调用方通过 ``config={"configurable": {"thread_id": <会话>}}`` 区分会话。
+    LLM 依赖不在此处绑定——调用时通过
+    ``config={"configurable": {"thread_id": <会话>}}`` 区分会话，
+    ``context=GraphDeps(llm=<LLMService>)`` 注入运行时依赖。
     """
-    builder = StateGraph(AgentState)
+    builder = StateGraph(AgentState, context_schema=GraphDeps)
     builder.add_node("normalize_input", normalize_input)
     builder.add_node("welcome", welcome)
-    if llm is None:
-        collect_node = collect
-    else:
-        from .collect import run_collect
-
-        async def collect_node(state):  # noqa: F811
-            return await run_collect(state, llm)
     builder.add_node("collect", collect_node)
-    builder.add_node("single_photo", single_photo)
+    builder.add_node("single_photo", single_photo_node)
-    builder.add_node("dual_photo", dual_photo)
+    builder.add_node("dual_photo", dual_photo_node)
-    builder.add_node("single_confirm", single_confirm)
+    builder.add_node("single_confirm", single_confirm_node)
-    builder.add_node("dual_confirm", dual_confirm)
+    builder.add_node("dual_confirm", dual_confirm_node)
 
     builder.add_edge(START, "normalize_input")
     builder.add_conditional_edges(
@@ -265,10 +190,17 @@ def build_graph(checkpointer=None, llm=None):
             END: END,
         },
     )
-    builder.add_edge("welcome", END)
+    for name in (
-    builder.add_edge("collect", END)
+        "welcome", "collect", "single_photo", "dual_photo",
-    builder.add_edge("single_photo", END)
+        "single_confirm", "dual_confirm",
-    builder.add_edge("dual_photo", END)
+    ):
-    builder.add_edge("single_confirm", END)
+        builder.add_edge(name, END)
-    builder.add_edge("dual_confirm", END)
     return builder.compile(checkpointer=checkpointer or InMemorySaver())
+
+
+__all__ = [
+    "AgentState",
+    "GraphDeps",
+    "build_graph",
+    "workflow_fake_llm",
+]

src/agent/photo.py

@@ -0,0 +1,83 @@
+"""拍照节点：按工作流分支选提示词，由 LLM 同时决定下一阶段和话术。
+
+来源：workflow/20261020/事故信息采集2025118.json 的四个拍照 chatNode。
+【拍摄完成】、拍照失败、无回复均作为模型输入；代码校验前缀、转换
+白名单和 clientMode 出口，不覆盖模型的判断或正文。
+"""
+
+from .collect import (
+    _bookkeeping,
+    _build_user_message,
+    context_prompt,
+    conversation_history,
+    validated_reply,
+)
+from .llm import LLMService
+from .prompts import (
+    DUAL_PHOTO,
+    DUAL_PHOTO_ADDON,
+    SINGLE_PHOTO,
+    SINGLE_PHOTO_ADDON,
+    render_prompt,
+)
+from .state import AgentState
+from .transitions import is_legal_transition
+
+PHOTO_PROMPTS = {
+    ("single", "direct"): SINGLE_PHOTO,
+    ("single", "browser_addon"): SINGLE_PHOTO_ADDON,
+    ("dual", "direct"): DUAL_PHOTO,
+    ("dual", "browser_addon"): DUAL_PHOTO_ADDON,
+}
+
+
+def is_legal_photo_transition(source: str, target: str, *, mode: str) -> bool:
+    """按工作流 Condition#11/#12 校验模式出口；不替模型改写目标。
+
+    direct 拍照结束进入 3001/3002；addon 拍照结束进入 0000。
+    两种模式均允许提示词定义的转人工码和保持当前阶段。
+    """
+    if not is_legal_transition(source, target):
+        return False
+    if mode == "browser_addon" and target in {"3001", "3002"}:
+        return False
+    if mode == "direct" and target == "0000":
+        return False
+    return True
+
+
+async def run_photo(state: AgentState, llm: LLMService, *, kind: str) -> dict:
+    """调用对应拍照模型一次；输出不合规时最多修复一次。
+
+    历史包含上一轮已经提出的拍照/身份问题，随 checkpoint 传到核实
+    节点，避免切换节点后丢失上下文。车损位置等自然语言保留原样。
+    """
+    mode = state.get("client_mode") or "direct"
+    prompt_name = PHOTO_PROMPTS[(kind, mode)]
+    current = state["stage_code"]
+    raw, stage = await validated_reply(
+        llm,
+        render_prompt(prompt_name, state) + context_prompt(state),
+        _build_user_message(state),
+        current,
+        history=conversation_history(state),
+        validator=lambda source, target: is_legal_photo_transition(
+            source, target, mode=mode
+        ),
+    )
+    no_reply_count, clarify_count = _bookkeeping(state)
+    dialogue = list(state.get("dialogue") or [])
+    dialogue.extend([
+        {"role": "user", "content": state.get("input_text", "")},
+        {"role": "assistant", "content": raw},
+    ])
+    return {
+        "turn_count": (state.get("turn_count") or 0) + 1,
+        "response_text": raw,
+        "stage_code": stage,
+        "raw_status_code": stage,
+        "no_reply_count": no_reply_count,
+        "clarify_count": clarify_count,
+        "dialogue": dialogue,
+        "turn_patch": {},
+    }

src/backends/langgraph.py

@@ -14,18 +14,21 @@
 
 from collections.abc import AsyncIterator
 
-from ..agent.graph import build_graph
+from ..agent.graph import GraphDeps, build_graph, workflow_fake_llm
 from .chat import ChatInput, ChatResult, ChatStreamEvent, ChatBackend, TextDelta
 
 
 class LangGraphBackend(ChatBackend):
     """Adapt a compiled LangGraph to the neutral chat contract."""
 
-    def __init__(self, graph=None):
+    def __init__(self, graph=None, llm=None):
         """参数:
-        graph: 编译后的图；缺省调用 ``build_graph()``（内存 checkpoint）。
+        graph: 编译后的图；缺省调用 ``build_graph()``。
+        llm: LLM 服务（运行时依赖）；缺省为离线 workflow fake。
+            生产环境由 lifespan 注入 ``build_llm_service()`` 的实例。
         """
         self._graph = graph if graph is not None else build_graph()
+        self._llm = llm if llm is not None else workflow_fake_llm()
 
     async def complete(self, chat_input: ChatInput) -> ChatResult:
         """非流式：执行一轮图并翻译结果。
@@ -40,8 +43,13 @@ class LangGraphBackend(ChatBackend):
           （作业 6 起实现）。
         """
         result = await self._graph.ainvoke(
-            {"session_id": chat_input.session_id, "input_text": chat_input.text},
+            {
+                "session_id": chat_input.session_id,
+                "input_text": chat_input.text,
+                "client_mode": chat_input.client_mode,
+            },
             config={"configurable": {"thread_id": chat_input.session_id}},
+            context=GraphDeps(llm=self._llm),
         )
         return ChatResult(
           content = result["response_text"],

src/core/fastgpt_client.py

@@ -26,8 +26,8 @@ async def lifespan(app: FastAPI):
         from ..agent.llm import build_llm_service
 
         global _langgraph_graph
-        _langgraph_graph = build_graph(llm=build_llm_service())
+        _langgraph_graph = build_graph()
-        _chat_backend = LangGraphBackend(_langgraph_graph)
+        _chat_backend = LangGraphBackend(_langgraph_graph, llm=build_llm_service())
     else:
         _fastgpt_client = AsyncChatClient(
             api_key=Config.FASTGPT_API_KEY,

test/agent/test_collect_llm.py

@@ -26,7 +26,8 @@ from src.agent.prompts import COLLECTION, load_prompt
 def collect_state(**overrides) -> dict:
     state = {
         "session_id": "s1",
-        "input_text": "你好",
+        # 使用完整句而非短促词，避免触发分流守卫（“你好”<=2 字）
+        "input_text": "发生了交通事故，两辆车追尾",
         "stage_code": "1002",
         "turn_count": 1,
         "no_reply_count": 0,
@@ -138,13 +139,16 @@ async def test_collect_illegal_stage_proposal_not_committed():
 
 @pytest.mark.asyncio
 async def test_graph_collect_node_with_llm():
-    # 图级：build_graph(llm=...) 时首轮走 welcome（不调模型），
+    # 图级：首轮走 welcome（确定性，不调模型）；第二轮起 collect
-    # 第二轮起 collect 节点走 run_collect
+    # 节点经 Runtime 注入的 llm 走 run_collect
+    from src.agent.graph import GraphDeps
+
     llm = FakeLLMService(lambda s, u: "<state>1002</state>收到。")
-    graph = build_graph(llm=llm)
+    graph = build_graph()
     first = await graph.ainvoke(
         {"session_id": "s1", "input_text": "追尾了"},
         config={"configurable": {"thread_id": "s1"}},
+        context=GraphDeps(llm=llm),
     )
     assert len(llm.calls) == 0  # welcome 是确定性节点
     assert first["stage_code"] == "1001"
@@ -152,6 +156,7 @@ async def test_graph_collect_node_with_llm():
     result = await graph.ainvoke(
         {"session_id": "s1", "input_text": "继续办理"},
         config={"configurable": {"thread_id": "s1"}},
+        context=GraphDeps(llm=llm),
     )
     assert len(llm.calls) == 1
     assert result["response_text"].startswith("<state>1002</state>")

test/agent/test_confirm.py

@@ -0,0 +1,193 @@
+"""作业 5 验收测试：信息核实节点（LLM 决策版）。
+
+测试用脚本化 FakeLLM 模拟“行为良好的模型”：按用户最新输入
+返回对应的带前缀回复；字段提取调用按提示词返回 JSON。
+断言仍是行为级的：stage 推进、person_fields 落库、person_index
+切换、模糊不计入失败。
+"""
+
+import json
+
+import pytest
+
+from src.agent.graph import GraphDeps, build_graph, workflow_fake_llm
+from src.agent.llm import FakeLLMService
+
+
+def latest_input(message: str) -> str:
+    return message.split("【用户最新输入】")[-1].strip()
+
+
+def make_llm(dialogue_script, extraction_script=None):
+    """dialogue_script(latest_input) -> 带前缀回复；
+    extraction_script(extraction_system_prompt) -> JSON 字符串。"""
+
+    def responder(system, user):
+        if system.startswith("【字段提取】"):
+            if extraction_script:
+                return extraction_script(system)
+            return "{}"
+        return dialogue_script(latest_input(user))
+
+    return FakeLLMService(responder)
+
+
+async def invoke(graph, thread_id, text, llm=None, **initial):
+    payload = {"session_id": thread_id, "input_text": text}
+    payload.update(initial)
+    return await graph.ainvoke(
+        payload,
+        config={"configurable": {"thread_id": thread_id}},
+        context=GraphDeps(llm=llm or workflow_fake_llm()),
+    )
+
+
+# ---------- 单车核实 ----------
+
+
+@pytest.mark.asyncio
+async def test_entry_consumes_pending_identity_question():
+    # 用户确认车主身份 → 模型直接问姓名（身份问题的提问发生在
+    # 拍照节点进入 3001 的同一回复里，这里不得重复）
+    llm = make_llm(
+        lambda q: "<state>3001</state>好的，请问您的姓名是什么？"
+    )
+    graph = build_graph()
+    result = await invoke(
+        graph, "t-entry", "是的，我是车主", llm=llm,
+        stage_code="3001", client_mode="direct",
+    )
+
+    assert result["stage_code"] == "3001"
+    assert "姓名" in result["response_text"]
+    assert "车主" not in result["response_text"]
+
+
+@pytest.mark.asyncio
+async def test_single_confirm_full_chain_to_0000():
+    def dialogue(q):
+        if q == "是的":
+            return "<state>3001</state>好的，请问您的姓名是什么？"
+        if q == "张三":
+            return "<state>3001</state>张三您好，请问您的身份证号码后四位是否正确？"
+        if q == "是对的":
+            return "<state>3001</state>请问您的手机号码后四位是否正确？"
+        return "<state>0000</state>好的，信息已全部确认无误。感谢您的配合。"
+
+    def extraction(system):
+        if "【用户输入】张三" in system:
+            return json.dumps({"xm1": "张三"}, ensure_ascii=False)
+        return "{}"
+
+    llm = make_llm(dialogue, extraction)
+    graph = build_graph()
+
+    r1 = await invoke(graph, "t-chain", "是的", llm=llm, stage_code="3001")
+    assert "姓名" in r1["response_text"]
+    r2 = await invoke(graph, "t-chain", "张三", llm=llm)
+    assert "身份证" in r2["response_text"]
+    assert r2["person_fields"]["xm1"] == "张三"
+    r3 = await invoke(graph, "t-chain", "是对的", llm=llm)
+    assert "手机" in r3["response_text"]
+    r4 = await invoke(graph, "t-chain", "没错", llm=llm)
+    assert r4["stage_code"] == "0000"
+    assert r4["response_text"].startswith("<state>0000</state>")
+
+
+@pytest.mark.asyncio
+async def test_deny_collects_correct_value():
+    llm = make_llm(
+        lambda q: "<state>3001</state>抱歉，请告诉我您正确的身份证号码后四位。",
+        lambda system: "{}",
+    )
+    graph = build_graph()
+    result = await invoke(
+        graph, "t-deny", "不对", llm=llm,
+        stage_code="3001", person_fields={"xm1": "张三"}, client_mode="direct",
+    )
+
+    assert result["stage_code"] == "3001"
+    assert "身份证" in result["response_text"]
+    assert result["person_fields"].get("sfzmwh1") in (None, "")
+
+
+@pytest.mark.asyncio
+async def test_vague_answer_stays_without_counting_failure():
+    llm = make_llm(
+        lambda q: "<state>3001</state>请问您是否是这辆车的车主？请回答“是”或“不是”。",
+        lambda system: "{}",
+    )
+    graph = build_graph()
+    result = await invoke(
+        graph, "t-vague", "嗯", llm=llm,
+        stage_code="3001", clarify_count=0, client_mode="direct",
+    )
+
+    assert result["stage_code"] == "3001"
+    assert "车主" in result["response_text"]
+    assert (result.get("clarify_count") or 0) == 0
+
+
+# ---------- 双车核实 ----------
+
+
+@pytest.mark.asyncio
+async def test_dual_person_index_switches_after_handoff():
+    llm = make_llm(
+        lambda q: "<state>3002</state>您好。请问这辆车是由您驾驶的吗？",
+        lambda system: json.dumps({"person_index": 2}),
+    )
+    graph = build_graph()
+    result = await invoke(
+        graph, "t-dual", "移交完毕", llm=llm,
+        stage_code="3002", person_index=1,
+        person_fields={"xm1": "张三"}, client_mode="direct",
+    )
+
+    assert result["stage_code"] == "3002"
+    assert result["person_index"] == 2
+    assert "驾驶" in result["response_text"]
+
+
+@pytest.mark.asyncio
+async def test_dual_second_person_fields_isolated():
+    def dialogue(q):
+        if q == "是的":
+            return "<state>3002</state>好的，驾驶员您好。请问您的姓名是什么？"
+        return "<state>3002</state>请问您的身份证号码后四位是否正确？"
+
+    def extraction(system):
+        if "【用户输入】李四" in system:
+            return json.dumps({"xm2": "李四"}, ensure_ascii=False)
+        return "{}"
+
+    llm = make_llm(dialogue, extraction)
+    graph = build_graph()
+
+    await invoke(
+        graph, "t-dual2", "是的", llm=llm,
+        stage_code="3002", person_index=2,
+        person_fields={"xm1": "张三", "sjhm1": "13800001111"},
+        client_mode="direct",
+    )
+    result = await invoke(graph, "t-dual2", "李四", llm=llm)
+
+    assert result["person_fields"]["xm2"] == "李四"
+    assert result["person_fields"]["xm1"] == "张三"
+
+
+@pytest.mark.asyncio
+async def test_dual_completion_goes_to_0000():
+    llm = make_llm(
+        lambda q: "<state>0000</state>好的，两位驾驶员的信息均已核实完毕。感谢配合。",
+        lambda system: "{}",
+    )
+    graph = build_graph()
+    result = await invoke(
+        graph, "t-dual-done", "没错", llm=llm,
+        stage_code="3002", person_index=2,
+        person_fields={"xm1": "张三", "xm2": "李四"},
+        client_mode="direct",
+    )
+
+    assert result["stage_code"] == "0000"

test/agent/test_photo.py

@@ -0,0 +1,116 @@
+"""拍照节点回归：LLM 提议阶段与正文，运行时校验并保存上下文。
+
+fake 输出仅验证接线与校验，不代替真实模型对提示词服从性的评测。
+"""
+
+import pytest
+
+from src.agent.collect import CollectReplyError
+from src.agent.graph import GraphDeps, build_graph, workflow_fake_llm
+from src.agent.llm import FakeLLMService
+
+
+async def invoke(graph, thread_id, text, llm=None, **initial):
+    return await graph.ainvoke(
+        {"session_id": thread_id, "input_text": text, **initial},
+        config={"configurable": {"thread_id": thread_id}},
+        context=GraphDeps(llm=llm or workflow_fake_llm()),
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("source", "target"),
+    [("2000", "2001"), ("2001", "2002"), ("2002", "2003"), ("2003", "2005"),
+     ("2010", "2011"), ("2011", "2012"), ("2012", "2013"),
+     ("2013", "2014"), ("2014", "2015")],
+)
+async def test_photo_applies_model_proposal(source, target):
+    raw = f"<state>{target}</state>模型生成的本轮拍照引导。"
+    llm = FakeLLMService(lambda s, u: raw)
+    result = await invoke(build_graph(), "photo", "【拍摄完成】", llm=llm,
+                          stage_code=source, client_mode="direct")
+    assert result["stage_code"] == target
+    assert result["response_text"] == raw
+    assert len(llm.calls) == 1
+    assert "【拍摄完成】" in llm.calls[0]["user_message"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("text", ["【拍摄完成】", "【客户端连续3次拍摄识别失败：图片过于模糊】", "【用户无回复】"])
+async def test_event_alone_does_not_force_stage(text):
+    # 模型保持阶段时，运行时不会仅凭事件强制跳转。
+    llm = FakeLLMService(lambda s, u: "<state>2012</state>请确认当前操作。")
+    result = await invoke(build_graph(), "hold", text, llm=llm,
+                          stage_code="2012", client_mode="direct")
+    assert result["stage_code"] == "2012"
+    assert len(llm.calls) == 1
+
+
+@pytest.mark.asyncio
+async def test_failure_stage_also_comes_from_model():
+    llm = FakeLLMService(lambda s, u: "<state>0005</state>正在转接人工。")
+    result = await invoke(build_graph(), "fail", "【客户端连续3次拍摄识别失败：图片过于模糊】", llm=llm,
+                          stage_code="2012", client_mode="direct")
+    assert result["stage_code"] == "0005"
+    assert len(llm.calls) == 1
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("source", "mode", "target", "node"),
+    [("2005", "direct", "3001", "单车拍照"),
+     ("2005", "browser_addon", "0000", "单车拍照#6"),
+     ("2015", "direct", "3002", "双车拍照"),
+     ("2015", "browser_addon", "0000", "双车拍照#7")],
+)
+async def test_mode_selects_workflow_prompt(source, mode, target, node):
+    from src.agent.prompts import load_prompt
+    from src.agent.photo import PHOTO_PROMPTS
+
+    kind = "single" if source == "2005" else "dual"
+    expected_prompt = load_prompt(PHOTO_PROMPTS[(kind, mode)])
+    expected_body = expected_prompt.split("--- 原文开始 ---", 1)[1].split("--- 原文结束 ---", 1)[0].strip()
+    raw = f"<state>{target}</state>对应模式的结束话术。"
+    llm = FakeLLMService(lambda s, u: raw)
+    text = "车子左后方" if source == "2005" else "【拍摄完成】"
+    result = await invoke(build_graph(), "end", text, llm=llm,
+                          stage_code=source, client_mode=mode)
+    assert result["stage_code"] == target
+    assert len(llm.calls) == 1
+    assert f"节点「{node}」" in expected_prompt
+    # 模型收到原文，预填变量已经渲染，来源头部不进入模型上下文。
+    assert "{{$" not in llm.calls[0]["system_prompt"]
+    assert expected_body.splitlines()[0] in llm.calls[0]["system_prompt"]
+
+
+@pytest.mark.asyncio
+async def test_photo_preserves_guidance_and_handoff_history():
+    replies = iter([
+        "<state>2015</state>您可以点击手机界面中间的拍照按钮。",
+        "<state>3002</state>请问第一辆车是由您驾驶的吗？",
+    ])
+    llm = FakeLLMService(lambda s, u: next(replies))
+    graph = build_graph()
+    first = await invoke(graph, "history", "拍照按钮在哪里？", llm=llm,
+                         stage_code="2015", client_mode="direct")
+    second = await invoke(graph, "history", "【拍摄完成】", llm=llm)
+    assert "手机界面中间" in first["response_text"]
+    assert llm.calls[1]["history"][0] == {"role": "user", "content": "拍照按钮在哪里？"}
+    assert llm.calls[1]["user_message"] == "【拍摄完成】"
+    assert second["dialogue"][-1]["content"] == second["response_text"]
+    assert "驾驶" in second["dialogue"][-1]["content"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(("source", "mode", "target"), [
+    ("2000", "direct", "2003"),
+    ("2015", "browser_addon", "3002"),
+    ("2005", "direct", "0000"),
+])
+async def test_illegal_proposal_fails_after_one_repair(source, mode, target):
+    llm = FakeLLMService(lambda s, u: f"<state>{target}</state>非法提议")
+    with pytest.raises(CollectReplyError):
+        await invoke(build_graph(), "illegal", "继续", llm=llm,
+                     stage_code=source, client_mode=mode)
+    assert len(llm.calls) == 2

test/agent/test_scaffold.py

@@ -11,15 +11,16 @@
 
 import pytest
 
-from src.agent.graph import build_graph
+from src.agent.graph import GraphDeps, build_graph, workflow_fake_llm
 
 
-async def ask(graph, thread_id: str, text: str) -> dict:
+async def ask(graph, thread_id: str, text: str, llm=None) -> dict:
     # 只传本轮变化的字段；对已有 checkpoint 的 thread 传完整初始状态，
     # 其中的 turn_count=0 会覆盖已累计的值（LangGraph 输入按通道覆盖）。
     return await graph.ainvoke(
         {"session_id": thread_id, "input_text": text},
         config={"configurable": {"thread_id": thread_id}},
+        context=GraphDeps(llm=llm or workflow_fake_llm()),
     )
 
 

test/agent/test_transitions.py

@@ -12,8 +12,9 @@
 
 import pytest
 
-import src.agent.graph as graph_mod
+from src.agent.llm import FakeLLMService
-from src.agent.graph import build_graph
+
+from src.agent.graph import GraphDeps, build_graph, workflow_fake_llm
 from src.agent.transitions import (
     TERMINAL_STAGES,
     TransitionSpec,
@@ -21,12 +22,13 @@ from src.agent.transitions import (
 )
 
 
-def invoke(graph, thread_id, text, **initial):
+def invoke(graph, thread_id, text, llm=None, **initial):
     payload = {"session_id": thread_id, "input_text": text}
     payload.update(initial)
     return graph.ainvoke(
         payload,
         config={"configurable": {"thread_id": thread_id}},
+        context=GraphDeps(llm=llm or workflow_fake_llm()),
     )
 
 
@@ -82,12 +84,21 @@ async def test_valid_reply_resets_no_reply_count():
 
 
 @pytest.mark.asyncio
-async def test_illegal_proposal_keeps_stage(monkeypatch):
+async def test_illegal_proposal_raises_and_state_untouched():
-    # LLM 提议非法跳步（2001 → 2003）：校验拒绝，状态不被污染
+    # LLM 坚持非法跳步（2001 → 2003）：白名单拒绝 + 一次修复后
+    # 仍非法 → CollectReplyError，非法阶段不写状态
+    from src.agent.collect import CollectReplyError
+
+    llm = FakeLLMService(lambda s, u: "<state>2003</state>跳步")
     graph = build_graph()
-    monkeypatch.setattr(graph_mod, "_fake_llm_propose", lambda s, e: "2003")
+    with pytest.raises(CollectReplyError):
-    result = await invoke(graph, "t-illegal", "【拍摄完成】", stage_code="2001")
+        await invoke(graph, "t-illegal", "【拍摄完成】", llm=llm, stage_code="2001")
-    assert result["stage_code"] == "2001"
+    assert len(llm.calls) == 2
+
+    # 会话未被污染：下一轮用守规矩的 fake 继续，阶段仍是 2001
+    good = FakeLLMService(lambda s, u: "<state>2002</state>已收到。")
+    result = await invoke(graph, "t-illegal", "【拍摄完成】", llm=good, stage_code="2001")
+    assert result["stage_code"] == "2002"
 
 
 @pytest.mark.asyncio

test/agent/test_workflow_prompts.py

@@ -0,0 +1,101 @@
+"""20261020 工作流的提示词移植与模型决策回归。"""
+
+from datetime import datetime, timezone
+import json
+from pathlib import Path
+
+import pytest
+
+from src.agent.collect import CollectReplyError, run_collect
+from src.agent.llm import FakeLLMService
+from src.agent.prompts import COLLECTION, load_prompt, render_prompt
+
+WORKFLOW = Path(__file__).resolve().parents[2] / "workflow/20261020/事故信息采集2025118.json"
+NODES = {
+    "collection": "事故信息收集",
+    "single_photo": "单车拍照",
+    "dual_photo": "双车拍照",
+    "single_photo_addon": "单车拍照#6",
+    "dual_photo_addon": "双车拍照#7",
+    "single_confirm": "单车信息确认",
+    "dual_confirm": "双车信息确认",
+}
+
+
+@pytest.mark.parametrize(("prompt", "node_name"), NODES.items())
+def test_archived_prompt_matches_workflow(prompt, node_name):
+    workflow = json.loads(WORKFLOW.read_text(encoding="utf-8-sig"))
+    node = next(n for n in workflow["nodes"] if n["name"] == node_name)
+    original = next(i["value"] for i in node["inputs"] if i["key"] == "systemPrompt")
+    archived = load_prompt(prompt).split("--- 原文开始 ---", 1)[1].split("--- 原文结束 ---", 1)[0]
+    assert archived.strip() == original.strip()
+
+
+def test_render_replaces_time_and_person_variables():
+    now = datetime(2026, 10, 20, 12, 0, tzinfo=timezone.utc)
+    rendered = render_prompt(COLLECTION, {}, now=now)
+    assert "2026-10-20 12:00:00 +0000" in rendered
+    rendered = render_prompt("dual_confirm", {
+        "person_fields": {"hphm1": "测试A", "hphm2": "测试B", "sfzmwh1": "1234"}
+    })
+    assert "测试A" in rendered and "测试B" in rendered
+    assert "1234" in rendered
+    assert "{{$" not in rendered
+    assert "未提供" in rendered
+
+
+@pytest.mark.asyncio
+async def test_short_answer_does_not_override_model_routing():
+    # “在”可能是明确回答；运行时不得按字数一律拒绝分流。
+    llm = FakeLLMService(lambda s, u: "<state>2000</state>信息已齐，请拍摄车辆前方。")
+    result = await run_collect({
+        "stage_code": "1002", "input_text": "在", "clarify_count": 0,
+        "dialogue": [{"role": "assistant", "content": "您还在事故现场吗？"}],
+    }, llm)
+    assert result["stage_code"] == "2000"
+    assert result["clarify_count"] == 0
+    assert len(llm.calls) == 1
+
+
+@pytest.mark.asyncio
+async def test_format_and_stage_errors_share_one_repair_budget():
+    replies = iter(["缺少前缀", "<state>9999</state>非法阶段"])
+    llm = FakeLLMService(lambda s, u: next(replies))
+    with pytest.raises(CollectReplyError):
+        await run_collect({"stage_code": "1002", "input_text": "追尾了"}, llm)
+    assert len(llm.calls) == 2
+
+
+@pytest.mark.asyncio
+async def test_collect_sends_history_as_messages():
+    history = [
+        {"role": "user", "content": "没有人受伤"},
+        {"role": "assistant", "content": "<state>1002</state>请问有几辆车？"},
+    ]
+    llm = FakeLLMService(lambda s, u: "<state>1002</state>请问您还在现场吗？")
+    await run_collect({
+        "stage_code": "1002", "input_text": "两辆", "dialogue": history,
+    }, llm)
+    assert llm.calls[0]["history"] == history
+    assert llm.calls[0]["user_message"] == "两辆"
+    assert "没有人受伤" not in llm.calls[0]["user_message"]
+
+
+@pytest.mark.asyncio
+async def test_openai_adapter_preserves_message_roles():
+    from types import SimpleNamespace
+    from unittest.mock import AsyncMock
+    from src.agent.llm import OpenAILLMService
+
+    create = AsyncMock(return_value=SimpleNamespace(choices=[
+        SimpleNamespace(message=SimpleNamespace(content="<state>1002</state>收到。"))
+    ]))
+    llm = OpenAILLMService(model="test", api_key="test", max_retries=0)
+    llm._client = SimpleNamespace(chat=SimpleNamespace(completions=SimpleNamespace(create=create)))
+    history = [{"role": "assistant", "content": "之前的问题"}]
+    await llm.complete("系统提示词", "本轮输入", history=history)
+    assert create.call_args.kwargs["messages"] == [
+        {"role": "system", "content": "系统提示词"},
+        {"role": "assistant", "content": "之前的问题"},
+        {"role": "user", "content": "本轮输入"},
+    ]