diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_utils.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_utils.py
index f5877ecb..9d2085d7 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_utils.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/_utils.py
@@ -15,10 +15,15 @@ from langchain_core.runnables import RunnableConfig
 logger = logging.getLogger(__name__)
 
 
-def get_model(config: RunnableConfig):
+def get_model(config: RunnableConfig, model_name: str | None = None):
+    """创建对话模型。
+
+    model_name 显式传入时优先使用（如素材收集专家的检索词/排序专用快模型），
+    否则回退到 configurable.model_name（用户在写作表单里选的主模型）。
+    """
     from deerflow.models import create_chat_model
-    model_name = config.get("configurable", {}).get("model_name", "") or None
-    return create_chat_model(name=model_name)
+    name = model_name or config.get("configurable", {}).get("model_name", "") or None
+    return create_chat_model(name=name)
 
 
 # 弱模型常把结构性标点写成中文全角符号，仅在常规解析失败后作为修复手段
@@ -220,11 +225,16 @@ class ThinkTagParser:
         return 0
 
 
-async def llm_json(system: str, user: str, fallback: dict, config: RunnableConfig) -> dict:
-    """调用 LLM，解析 JSON 输出，失败用 fallback。"""
-    model = get_model(config)
+async def llm_json(system: str, user: str, fallback: dict, config: RunnableConfig, model_name: str | None = None) -> dict:
+    """调用 LLM，解析 JSON 输出，失败用 fallback。
+
+    model_name 可覆盖默认模型（如检索词/排序使用配置的专用快模型）。
+    """
     messages = [SystemMessage(content=system), HumanMessage(content=user)]
     try:
+        # get_model 放进 try：模型名在 config.yaml 找不到时 create_chat_model 会抛
+        # ValueError，若留在 try 外会逐层上抛使整个图硬崩溃。
+        model = get_model(config, model_name=model_name)
         resp = await model.ainvoke(messages)
         return parse_json(_content_to_text(resp.content), fallback)
     except Exception as e:
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/editor.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/editor.py
index 438dac98..6aa7019b 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/editor.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/editor.py
@@ -89,7 +89,19 @@ async def editor_node(state: AIWritingState, config: RunnableConfig) -> dict:
         ("language", LANGUAGE_REVIEW_USER.format(article=article)),
     ]
 
-    model = get_model(config)
+    # 模型创建守卫：模型名配错时 create_chat_model 抛 ValueError，
+    # 不拦截会硬崩溃整个图；此处转成可恢复的 status:error。
+    try:
+        model = get_model(config)
+    except Exception as e:
+        logger.error("editor model creation failed: %s", e, exc_info=True)
+        return {
+            "status": "error",
+            "progress_events": [{
+                "type": "error", "agent_name": "编辑",
+                "title": "模型错误", "detail": str(e), "message": "审核失败",
+            }],
+        }
     results: dict[str, dict] = {}
     review_errors: list[str] = []  # 分项审查的模型异常，供完成后非致命提示
 
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/intent_parser.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/intent_parser.py
index a575d902..92825d68 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/intent_parser.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/intent_parser.py
@@ -1,6 +1,8 @@
 """意图解析节点：将用户自然语言意图结构化为写作配置。"""
 from __future__ import annotations
 
+import re
+
 from langchain_core.runnables import RunnableConfig
 
 from ..state import AIWritingState
@@ -43,6 +45,33 @@ def _build_type_options(types: list[dict]) -> str:
     return " / ".join(f'{t["key"]}（{t["label"]}）' for t in types)
 
 
+def _coerce_word_count(raw) -> int | None:
+    """稳健地把模型返回的字数字段转成正整数；无法解析时返回 None。
+
+    弱模型（尤其老模型被陌生词带偏时）常返回「约800字」「两千」「800.0」
+    这类非纯数字形式，直接 int() 会抛 ValueError/TypeError 使整个图崩溃。
+    此处用「提取首个数字串」策略容错，解析不出就返回 None 让调用方保留默认值。
+    """
+    if raw is None:
+        return None
+    if isinstance(raw, bool):  # bool 是 int 子类，需先排除
+        return None
+    if isinstance(raw, (int, float)):
+        value = int(raw)
+        return value if value > 0 else None
+    if not isinstance(raw, str):
+        # dict / list 等结构化类型：str() 化后可能误含数字，直接判为无法解析
+        return None
+    m = re.search(r"\d[\d,]*", raw)
+    if not m:
+        return None
+    try:
+        value = int(m.group().replace(",", ""))
+    except ValueError:
+        return None
+    return value if value > 0 else None
+
+
 async def intent_parser_node(state: AIWritingState, config: RunnableConfig) -> dict:
     # 从 configurable 读取动态类型列表，未配置时退回默认值
     configurable = (config or {}).get("configurable", {})
@@ -73,10 +102,19 @@ async def intent_parser_node(state: AIWritingState, config: RunnableConfig) -> d
         config=config,
     )
 
-    updates: dict = {
-        "status": "researching",
-        "progress_events": [{"type": "step_done", "agent_name": "系统", "message": "意图解析完成"}],
-    }
+    # 意图解析模型调用失败时以非致命 notice 提示用户（llm_json 已用 fallback 兜底，
+    # 流程仍可继续，但用户需知道结构化配置可能不准）。
+    progress_events: list[dict] = []
+    if result.get("_llm_error"):
+        progress_events.append({
+            "type": "notice",
+            "agent_name": "系统",
+            "title": "模型错误",
+            "detail": f"意图解析失败，已按默认配置兜底。原因：{result['_llm_error']}",
+        })
+    progress_events.append({"type": "step_done", "agent_name": "系统", "message": "意图解析完成"})
+
+    updates: dict = {"status": "researching", "progress_events": progress_events}
 
     # 校验 LLM 返回的 article_type 是否在有效范围内，不合法时回退到默认值
     raw_type = result.get("article_type", "")
@@ -84,8 +122,9 @@ async def intent_parser_node(state: AIWritingState, config: RunnableConfig) -> d
 
     if result.get("target_audience"):
         updates["target_audience"] = result["target_audience"]
-    if result.get("word_count_target"):
-        updates["word_count_target"] = int(result["word_count_target"])
+    word_count = _coerce_word_count(result.get("word_count_target"))
+    if word_count is not None:
+        updates["word_count_target"] = word_count
     if result.get("refined_intent"):
         updates["user_intent"] = result["refined_intent"]
 
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py
index 3b8f19b4..688b9a02 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/researcher.py
@@ -40,6 +40,12 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
     sse_queue = cfg.get("sse_queue")
     sse_session_id: str = cfg.get("sse_session_id", "")
 
+    # 素材收集专家专用模型：检索词生成 / 素材排序 可在 config.yaml 的
+    # ai_writing 段单独配置（keyword_model / rank_model），留空则回退写作主模型。
+    ai_writing_config = cfg.get("ai_writing_config", {}) or {}
+    keyword_model: str | None = ai_writing_config.get("keyword_model") or None
+    rank_model: str | None = ai_writing_config.get("rank_model") or None
+
     async def _emit(message: str) -> None:
         if sse_queue is None:
             return
@@ -59,13 +65,13 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
         ),
         fallback={"keywords": [intent]},
         config=config,
+        model_name=keyword_model,
     )
     keywords: list[str] = kw_result.get("keywords", [intent])
     if extra_keywords:
         keywords = list(dict.fromkeys(keywords + extra_keywords))
 
     # Step 2: 并发搜索
-    ai_writing_config = config.get("configurable", {}).get("ai_writing_config", {})
     provider = get_search_provider(ai_writing_config)
 
     search_keywords = keywords[:6]
@@ -136,6 +142,7 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
             RANK_MATERIALS_USER.format(user_intent=intent, materials_list=materials_list),
             fallback={"ranked_ids": [m["id"] for m in raw_materials[:_MAX_MATERIALS]], "summary": f"共搜索到{len(raw_materials)}条素材"},
             config=config,
+            model_name=rank_model,
         )
         rank_error = rank_result.get("_llm_error", "")
         ranked_ids: list[str] = rank_result.get("ranked_ids", [m["id"] for m in raw_materials])
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_draft.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_draft.py
index bf3da60f..14013dd1 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_draft.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_draft.py
@@ -73,7 +73,19 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
     def _section_title(idx: int, sec: dict) -> str:
         return sec.get("section_title") or sec.get("sectionTitle") or f"章节{idx + 1}"
 
-    model = get_model(config)
+    # 模型创建守卫：模型名配错时 create_chat_model 抛 ValueError，
+    # 不拦截会硬崩溃整个图；此处转成可恢复的 status:error。
+    try:
+        model = get_model(config)
+    except Exception as e:
+        logger.error("draft model creation failed: %s", e, exc_info=True)
+        return {
+            "status": "error",
+            "progress_events": [{
+                "type": "error", "agent_name": "作家",
+                "title": "模型错误", "detail": str(e), "message": "草稿生成失败",
+            }],
+        }
     section_titles: list[str] = []
     section_bufs: list[str] = []   # 各章节正文（不含 ## 标题），供扩写兜底逐节替换
     section_errors: list[str] = []  # 分节写作的模型异常，供完成后非致命提示
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_outline.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_outline.py
index 404f9d5d..bd3ad1e6 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_outline.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_outline.py
@@ -62,11 +62,13 @@ async def writer_outline_node(state: AIWritingState, config: RunnableConfig) ->
     sse_queue: asyncio.Queue | None = cfg.get("sse_queue")
     session_id: str = cfg.get("sse_session_id", "")
 
-    model = get_model(config)
     full_text = ""   # 只累积正文（非思考）用于 JSON 解析
     parser = ThinkTagParser()
 
     try:
+        # get_model 放进 try：模型名配错时 create_chat_model 抛 ValueError，
+        # 留在 try 外会硬崩溃；放进来则统一走下方的 status:error 兜底。
+        model = get_model(config)
         async for chunk in model.astream([
             SystemMessage(content=OUTLINE_SYSTEM),
             HumanMessage(content=user_prompt),
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/prompts/researcher_prompts.py b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/prompts/researcher_prompts.py
index 70b41eba..b8f7e576 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/prompts/researcher_prompts.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/prompts/researcher_prompts.py
@@ -6,6 +6,11 @@ EXTRACT_KEYWORDS_SYSTEM = """你是一名专业情报研究员（素材收集专
 1. 从写作意图中提取 3-6 个搜索关键词
 你不能写文章，不能评价内容好坏，只做关键词提取。
 
+【严禁思考】
+- 这是一个简单的结构化抽取任务，禁止任何思考、分析、推理或规划过程。
+- 禁止输出 <think></think> 标签及其中的任何内容，禁止输出"让我想想""首先"之类的过渡语。
+- 收到输入后必须立即、直接输出最终 JSON，不得有任何前置内容。
+
 【输出要求】
 - 只输出一个 JSON 对象，不要任何解释文字，不要 markdown 代码块。
 - 必须使用英文双引号 "，不要用中文引号 “”。
@@ -31,6 +36,11 @@ RANK_MATERIALS_SYSTEM = """你是一名专业情报研究员（素材收集专
 1. 从提供的素材列表中，选出与写作意图最相关的素材ID，按相关度从高到低排序
 你不能写文章，只做素材筛选排序。
 
+【严禁思考】
+- 这是一个简单的筛选排序任务，禁止任何思考、分析、推理或规划过程。
+- 禁止输出 <think></think> 标签及其中的任何内容，禁止输出"让我想想""首先"之类的过渡语。
+- 收到输入后必须立即、直接输出最终 JSON，不得有任何前置内容。
+
 【输出要求】
 - 只输出一个 JSON 对象，不要任何解释文字，不要 markdown 代码块。
 - 必须使用英文双引号 "。