diff --git a/core/main_terminal_parts/tools_definition/search_web_tools.py b/core/main_terminal_parts/tools_definition/search_web_tools.py
index a2ebe6dc..d98b089c 100644
--- a/core/main_terminal_parts/tools_definition/search_web_tools.py
+++ b/core/main_terminal_parts/tools_definition/search_web_tools.py
@@ -92,7 +92,7 @@ class ToolsDefinitionSearchWebToolsMixin:
                         "type": "function",
                         "function": {
                             "name": "web_search",
-                            "description": "当现有资料不足时搜索外部信息。调用前说明目的，精准撰写 query，并合理设置时间/主题参数；避免重复或无意义的搜索。",
+                            "description": "当现有资料不足时搜索外部信息。调用前说明目的，精准撰写 query，并合理设置时间/主题参数；避免重复或无意义的搜索。结果中每条来源带 [src_xxx] 引用 ID，回答中标注事实来源时使用。",
                             "parameters": {
                                 "type": "object",
                                 "properties": self._inject_intent({
@@ -145,7 +145,7 @@ class ToolsDefinitionSearchWebToolsMixin:
                         "type": "function",
                         "function": {
                             "name": "extract_webpage",
-                            "description": "在 web_search 结果不够详细时提取网页正文。调用前说明用途，注意提取内容会消耗大量 token，超过80000字符将被拒绝。",
+                            "description": "在 web_search 结果不够详细时提取网页正文。调用前说明用途，注意提取内容会消耗大量 token，超过80000字符将被拒绝。结果会带来源引用 ID。",
                             "parameters": {
                                 "type": "object",
                                 "properties": self._inject_intent({
diff --git a/core/main_terminal_parts/tools_execution.py b/core/main_terminal_parts/tools_execution.py
index 83544bd1..b253e2ab 100644
--- a/core/main_terminal_parts/tools_execution.py
+++ b/core/main_terminal_parts/tools_execution.py
@@ -1697,13 +1697,35 @@ class MainTerminalToolsExecutionMixin:
                         )
 
                         if search_response["success"]:
+                            # 注册网页来源 citation，并让模型可见的 summary 带上 [src_xxx] 前缀
+                            from modules.citations import get_registry
+                            registry = get_registry(self)
+                            citations = []
+                            results_list = search_response.get("results", [])
+                            for item in results_list:
+                                ann = registry.register_url(
+                                    title=item.get("title") or item.get("url") or "",
+                                    url=item.get("url") or "",
+                                    snippet=item.get("content", ""),
+                                    published_date=item.get("published_date") or None,
+                                    source_tool="web_search",
+                                )
+                                if ann:
+                                    item["citation_id"] = ann["id"]
+                                    citations.append(ann)
                             result = {
                                 "success": True,
-                                "summary": search_response["summary"],
+                                "summary": self.search_engine.build_summary_text(
+                                    search_response.get("query") or arguments["query"],
+                                    results_list,
+                                    search_response.get("filters", {}),
+                                    search_response.get("timestamp", ""),
+                                ),
                                 "filters": search_response.get("filters", {}),
                                 "query": search_response.get("query"),
-                                "results": search_response.get("results", []),
-                                "total_results": search_response.get("total_results", 0)
+                                "results": results_list,
+                                "total_results": search_response.get("total_results", 0),
+                                "citations": citations,
                             }
                         else:
                             result = {
@@ -1738,10 +1760,19 @@ class MainTerminalToolsExecutionMixin:
                                     "url": url
                                 }
                             else:
+                                from modules.citations import get_registry
+                                ann = get_registry(self).register_url(
+                                    title=url,
+                                    url=url,
+                                    snippet=full_content[:500],
+                                    source_tool="extract_webpage",
+                                )
                                 result = {
                                     "success": True,
                                     "url": url,
-                                    "content": full_content
+                                    "content": full_content,
+                                    "citation_id": ann.get("id") if ann else None,
+                                    "citations": [ann] if ann else [],
                                 }
                         except Exception as e:
                             result = {
diff --git a/modules/citations.py b/modules/citations.py
new file mode 100644
index 00000000..0fcdd544
--- /dev/null
+++ b/modules/citations.py
@@ -0,0 +1,283 @@
+"""Inline Citations：引用注册表、marker 解析与校验。
+
+数据流：
+  工具执行（web_search / extract_webpage）→ CitationRegistry.register_url()
+  → 工具结果文本带 [src_xxx]（模型可见）
+  → 模型输出 【cite:src_xxx】/【file:相对路径】 marker
+  → assistant 消息落库前 finalize_message_citations() 校验 + 挂载 message.citations
+  → 前端渲染 citation chip
+
+设计要点：
+- 网页来源用 cite: 前缀 + ID（src_<sha1(规范化 url) 前 10 位>），确定性、同 URL 天然去重；
+- 文件来源用 file: 前缀 + 工作区相对路径（可带 #L120-148 / #p7 locator），不经过 registry；
+- registry 是会话运行期的临时查找表，不落盘；历史消息靠 message.citations 自包含恢复。
+"""
+
+from __future__ import annotations
+
+import hashlib
+import os
+import re
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+from urllib.parse import urlparse
+
+# marker 语法：【cite:src_xxx】/【file:AGENTS.md】/同前缀多来源逗号并列（中英文逗号均可）
+CITATION_MARK_RE = re.compile(r"【(cite|file):([^】]+)】")
+
+# locator 后缀：#L120 / #L120-148 / #L120-L148（GitHub 风格）/ #p7
+_LOCATOR_RE = re.compile(r"^(?P<path>.*?)#(?:L(?P<line_start>\d+)(?:-L?(?P<line_end>\d+))?|p(?P<page>\d+))$", re.IGNORECASE)
+
+_SRC_ID_RE = re.compile(r"^src_[0-9a-f]{10}x*$")
+
+SNIPPET_MAX_CHARS = 300
+
+# 文件 snippet 读取上限（前 64KB 足够覆盖 locator 行段与开头摘要）
+_FILE_SNIPPET_READ_BYTES = 65536
+
+# 图片扩展名：引用弹层渲染 <img> 而不是文本摘要
+_IMAGE_EXTS = {".png", ".jpg", ".jpeg", ".gif", ".webp", ".svg", ".bmp", ".ico", ".avif"}
+
+
+def _read_file_snippet(target: Path, ref: Dict[str, Any]) -> Optional[str]:
+    """读取文本文件内容片段作为引用预览：有 #L 定位时取对应行段，否则取开头。
+
+    二进制文件（含空字节）返回 None。空白压缩为单行，限长 SNIPPET_MAX_CHARS。
+    """
+    try:
+        with open(target, "rb") as f:
+            raw = f.read(_FILE_SNIPPET_READ_BYTES)
+    except Exception:
+        return None
+    if b"\x00" in raw:
+        return None
+    text = raw.decode("utf-8", errors="replace")
+    line_start = ref.get("line_start")
+    if line_start:
+        lines = text.splitlines()
+        start = max(1, int(line_start))
+        end = min(len(lines), int(ref.get("line_end") or start))
+        snippet = "\n".join(lines[start - 1 : end]) if start <= len(lines) else ""
+    else:
+        snippet = text
+    # 保留换行（弹层 pre-line 渲染），仅压缩行内空白、去空行
+    snippet = "\n".join(ln.strip() for ln in snippet.splitlines() if ln.strip())
+    return snippet[:SNIPPET_MAX_CHARS] or None
+
+
+def normalize_url(url: str) -> str:
+    """URL 规范化：小写 scheme/host、去末尾斜杠。用于同来源去重。"""
+    url = (url or "").strip()
+    try:
+        p = urlparse(url)
+        scheme = (p.scheme or "https").lower()
+        netloc = p.netloc.lower()
+        path = p.path.rstrip("/")
+        query = f"?{p.query}" if p.query else ""
+        return f"{scheme}://{netloc}{path}{query}"
+    except Exception:
+        return url
+
+
+def domain_of(url: str) -> str:
+    try:
+        return urlparse(url).netloc.lower().lstrip("www.")
+    except Exception:
+        return ""
+
+
+def make_url_citation_id(url: str) -> str:
+    digest = hashlib.sha1(normalize_url(url).encode("utf-8")).hexdigest()[:10]
+    return f"src_{digest}"
+
+
+class CitationRegistry:
+    """会话级网页来源注册表（运行期内存态，不落盘）。"""
+
+    def __init__(self) -> None:
+        self._by_id: Dict[str, Dict[str, Any]] = {}
+        self._id_by_url: Dict[str, str] = {}
+
+    def register_url(
+        self,
+        *,
+        title: str = "",
+        url: str = "",
+        snippet: str = "",
+        published_date: Optional[str] = None,
+        source_tool: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        """注册（或按规范化 URL 去重返回已有）一个网页来源 annotation。"""
+        if not url:
+            return {}
+        norm = normalize_url(url)
+        existing_id = self._id_by_url.get(norm)
+        if existing_id is not None:
+            return self._by_id[existing_id]
+
+        cid = make_url_citation_id(url)
+        # 哈希碰撞兜底（同一 URL 已命中则上面已返回；不同 URL 撞 hash 时加后缀）
+        while cid in self._by_id and normalize_url(self._by_id[cid].get("url", "")) != norm:
+            cid += "x"
+
+        ann: Dict[str, Any] = {
+            "id": cid,
+            "type": "url_citation",
+            "title": title or url,
+            "url": url,
+            "domain": domain_of(url),
+            "snippet": (snippet or "")[:SNIPPET_MAX_CHARS],
+        }
+        if published_date:
+            ann["published_date"] = published_date
+        if source_tool:
+            ann["source_tool"] = source_tool
+        self._by_id[cid] = ann
+        self._id_by_url[norm] = cid
+        return ann
+
+    def resolve(self, citation_id: str) -> Optional[Dict[str, Any]]:
+        return self._by_id.get(citation_id)
+
+
+def get_registry(terminal: Any) -> CitationRegistry:
+    """取 terminal 实例上的会话级 registry（懒创建）。"""
+    reg = getattr(terminal, "_citation_registry", None)
+    if reg is None:
+        reg = CitationRegistry()
+        terminal._citation_registry = reg
+    return reg
+
+
+def parse_marker_content(content: str) -> List[str]:
+    """拆分 marker 内容为引用 token 列表（支持中英文逗号分隔）。"""
+    parts = re.split(r"[,，]", content or "")
+    return [p.strip() for p in parts if p.strip()]
+
+
+def extract_citation_refs(text: str) -> List[str]:
+    """提取正文中全部完整 marker 的引用 token（保持出现顺序，去重）。"""
+    refs: List[str] = []
+    for m in CITATION_MARK_RE.finditer(text or ""):
+        for token in parse_marker_content(m.group(2)):
+            if token not in refs:
+                refs.append(token)
+    return refs
+
+
+def _parse_file_ref(token: str) -> Optional[Dict[str, Any]]:
+    """解析文件引用 token（裸相对路径，可带 #L120-148 / #p7 locator）。非法返回 None。"""
+    body = token.strip()
+    if not body:
+        return None
+    m = _LOCATOR_RE.match(body)
+    if m:
+        ref: Dict[str, Any] = {"path": m.group("path")}
+        if m.group("page"):
+            ref["page"] = int(m.group("page"))
+        if m.group("line_start"):
+            ref["line_start"] = int(m.group("line_start"))
+            if m.group("line_end"):
+                ref["line_end"] = int(m.group("line_end"))
+        return ref
+    return {"path": body}
+
+
+def _build_file_annotation(ref: Dict[str, Any], token: str, workspace_root: str) -> Optional[Dict[str, Any]]:
+    """校验文件引用：路径必须落在工作区内且文件存在；富化 file_name/size。
+
+    annotation id 与 marker 内 token 完全一致，前端按 id 精确匹配。
+    """
+    # 注意不能用 lstrip("./")：它是字符集语义，会把 .astrion 这类隐藏目录的前导点吃掉
+    rel_path = (ref.get("path") or "").strip()
+    if rel_path.startswith("./"):
+        rel_path = rel_path[2:]
+    if not rel_path or rel_path.startswith("/"):
+        return None
+    try:
+        root = Path(workspace_root).resolve()
+        target = (root / rel_path).resolve()
+        # 边界检查：必须在工作区内
+        if root != target and root not in target.parents:
+            return None
+        if not target.is_file():
+            return None
+        stat = target.stat()
+    except Exception:
+        return None
+
+    ann: Dict[str, Any] = {
+        "id": token,
+        "type": "file_citation",
+        "file_path": rel_path,
+        "file_name": target.name,
+        "size": stat.st_size,
+    }
+    snippet: Optional[str] = None
+    # 图片文件的 snippet 无意义（且 svg 这类文本型图片会把标记源码当摘要），
+    # 前端弹层按扩展名直接渲染 <img>
+    if target.suffix.lower() not in _IMAGE_EXTS:
+        snippet = _read_file_snippet(target, ref)
+    if snippet:
+        ann["snippet"] = snippet
+    if ref.get("page") is not None:
+        ann["page"] = ref["page"]
+    if ref.get("line_start") is not None:
+        ann["line_start"] = ref["line_start"]
+        if ref.get("line_end") is not None:
+            ann["line_end"] = ref["line_end"]
+    return ann
+
+
+def finalize_message_citations(
+    text: str,
+    registry: Optional[CitationRegistry],
+    workspace_root: str,
+) -> Tuple[str, List[Dict[str, Any]]]:
+    """assistant 消息落库前处理：剥离无效 marker，返回 (clean_text, used_annotations)。
+
+    - src_xxx：registry 查得到才保留，否则剥离 marker；
+    - file: 引用：路径在工作区内且文件存在才保留；
+    - 一个 marker 里部分有效时，只保留有效 token 重建 marker；全部无效则整体移除。
+    """
+    if not text or ("【cite:" not in text and "【file:" not in text):
+        return text, []
+
+    used: List[Dict[str, Any]] = []
+    seen_ids = set()
+
+    def _keep(ann: Optional[Dict[str, Any]]) -> Optional[str]:
+        if not ann:
+            return None
+        ann_id = ann.get("id")
+        if ann_id and ann_id not in seen_ids:
+            seen_ids.add(ann_id)
+            used.append(ann)
+        return ann_id
+
+    def _replace(m: re.Match) -> str:
+        kind = m.group(1)
+        tokens = parse_marker_content(m.group(2))
+        kept: List[str] = []
+        for token in tokens:
+            if kind == "cite":
+                # 网页来源：src_xxx 且 registry 查得到才保留
+                if not _SRC_ID_RE.match(token):
+                    continue
+                ann = registry.resolve(token) if registry else None
+                if _keep(ann):
+                    kept.append(token)
+            else:
+                # 文件来源：路径在工作区内且存在才保留
+                ref = _parse_file_ref(token)
+                if not ref:
+                    continue
+                ann = _build_file_annotation(ref, token, workspace_root)
+                if _keep(ann):
+                    kept.append(token)
+        if not kept:
+            return ""
+        return "【" + kind + ":" + ",".join(kept) + "】"
+
+    clean = CITATION_MARK_RE.sub(_replace, text)
+    return clean, used
diff --git a/modules/search_engine.py b/modules/search_engine.py
index b9af6880..80711778 100644
--- a/modules/search_engine.py
+++ b/modules/search_engine.py
@@ -6,6 +6,7 @@ from typing import Dict, Optional, Any, List
 from datetime import datetime
 from pathlib import Path
 import re
+from urllib.parse import urlparse
 try:
     from config import TAVILY_API_KEY, SEARCH_MAX_RESULTS, OUTPUT_FORMATS, DATA_DIR
 except ImportError:
@@ -150,10 +151,12 @@ class SearchEngine:
         
         # 处理每个搜索结果
         for idx, result in enumerate(raw_data.get("results", []), 1):
+            url = result.get("url", "")
             formatted_result = {
                 "index": idx,
                 "title": result.get("title", "无标题"),
-                "url": result.get("url", ""),
+                "url": url,
+                "domain": urlparse(url).netloc.lower() if url else "",
                 "content": result.get("content", ""),
                 "score": result.get("score", 0),
                 "published_date": result.get("published_date", "")
@@ -161,6 +164,48 @@ class SearchEngine:
             formatted["results"].append(formatted_result)
         
         return formatted
+
+    def build_summary_text(
+        self,
+        query: str,
+        results: List[Dict[str, Any]],
+        filters: Dict[str, Any],
+        timestamp: str
+    ) -> str:
+        """构建给模型看的搜索摘要文本。
+
+        若结果项带 citation_id（tools_execution 注册 citation 后回填），
+        标题行会带 [src_xxx] 前缀，供模型在行内引用中使用。
+        """
+        summary_lines = [
+            f"🔍 搜索查询: {query}",
+            f"📅 搜索时间: {timestamp}"
+        ]
+        
+        filter_notes = self._summarize_filters(filters or {})
+        if filter_notes:
+            summary_lines.append(filter_notes)
+        summary_lines.append("")
+        
+        # 添加搜索结果
+        if results:
+            summary_lines.append("📊 搜索结果:")
+            
+            for result in results:
+                cid = result.get("citation_id")
+                title_line = f"\n{result['index']}. [{cid}] {result['title']}" if cid else f"\n{result['index']}. {result['title']}"
+                summary_lines.extend([
+                    title_line,
+                    f"   🔗 {result['url']}",
+                    f"   📄 {result['content'][:200]}..." if len(result['content']) > 200 else f"   📄 {result['content']}",
+                ])
+                
+                if result.get("published_date"):
+                    summary_lines.append(f"   📅 发布时间: {result['published_date']}")
+        else:
+            summary_lines.append("未找到相关结果")
+        
+        return "\n".join(summary_lines)
     
     async def search_with_summary(
         self,
@@ -203,36 +248,15 @@ class SearchEngine:
                 "summary": ""
             }
         
-        # 构建摘要
-        summary_lines = [
-            f"🔍 搜索查询: {query}",
-            f"📅 搜索时间: {results['timestamp']}"
-        ]
-        
-        filter_notes = self._summarize_filters(results.get("filters", {}))
-        if filter_notes:
-            summary_lines.append(filter_notes)
-        summary_lines.append("")
-        
-        # 添加搜索结果
-        if results["results"]:
-            summary_lines.append("📊 搜索结果:")
-            
-            for result in results["results"]:
-                summary_lines.extend([
-                    f"\n{result['index']}. {result['title']}",
-                    f"   🔗 {result['url']}",
-                    f"   📄 {result['content'][:200]}..." if len(result['content']) > 200 else f"   📄 {result['content']}",
-                ])
-                
-                if result.get("published_date"):
-                    summary_lines.append(f"   📅 发布时间: {result['published_date']}")
-        else:
-            summary_lines.append("未找到相关结果")
-        
         return {
             "success": True,
-            "summary": "\n".join(summary_lines),
+            "summary": self.build_summary_text(
+                query,
+                results["results"],
+                results.get("filters", {}),
+                results["timestamp"]
+            ),
+            "timestamp": results["timestamp"],
             "filters": results.get("filters", {}),
             "query": results.get("query", query),
             "results": results.get("results", []),
diff --git a/prompts/main_system.txt b/prompts/main_system.txt
index e9bb0182..4aeccd1c 100644
--- a/prompts/main_system.txt
+++ b/prompts/main_system.txt
@@ -111,6 +111,29 @@
 
 注意：路径必须是工作区内已存在的文件。输出前确保文件已通过 `write_file` 或其他方式创建。
 
+### 3.1.6 行内引用（Citation）
+
+当你使用网页搜索结果、文件内容或其他外部资料回答时，必须在对应事实陈述之后添加引用标记，前端会渲染为可点击的来源胶囊。用户明确要求「引用」某个文件/来源时，同样必须输出标记——用加粗、行内代码或表格提到文件名都不算引用。
+
+标记格式：
+- 网页来源：【cite:src_xxx】（src_xxx 必须来自本次对话工具结果中明确提供的引用 ID，如 `[src_xxx]` 前缀或「来源 ID: src_xxx」）
+- 文件来源：【file:工作区相对路径】，例如 【file:AGENTS.md】；指向具体行段：【file:AGENTS.md#L120-148】
+- 图片等二进制文件同样用 file 标记引用（如 【file:.astrion/user_upload/xxx.png】），前端会渲染缩略预览
+- 一个陈述引用多个来源：同类逗号并列【cite:src_xxx,src_yyy】/【file:report.pdf,notes.md】；网页与文件混排相邻写两个标记，例如 【cite:src_xxx】【file:report.pdf】
+
+引用要求：
+1. 只引用实际支持当前陈述的来源，不引用无关来源
+2. 引用紧跟在相关句子或段落末尾，不要统一放在回答结尾或集中列「参考资料」（除非用户要求）
+3. 没有使用外部资料时不要添加引用，也不要逐句标注
+4. 只能使用工具结果明确提供的 src_xxx 和工作区内真实存在的文件相对路径，禁止编造；标记内禁止写 URL
+5. 代码块和行内代码中的标记不会被渲染，不要在代码示例里放引用
+
+示例：
+- 该功能在 v2.3 版本引入。【cite:src_abc123def4】
+- 两个项目的实现方案一致。【cite:src_abc123def4,src_xyz987abcd】
+- 配置项说明见部署文档。【file:docs/deploy.md#L45-60】
+- 这一结论同时有网页和本地文档支持。【cite:src_abc123def4】【file:docs/deploy.md】
+
 ### 3.2 文件操作
 
 - `write_file`：写入文件（`append` 控制覆盖/追加）
diff --git a/prompts/main_system_vl.txt b/prompts/main_system_vl.txt
index caa012d7..5f61a9ce 100644
--- a/prompts/main_system_vl.txt
+++ b/prompts/main_system_vl.txt
@@ -111,6 +111,29 @@
 
 注意：路径必须是工作区内已存在的文件。输出前确保文件已通过 `write_file` 或其他方式创建。
 
+### 3.1.6 行内引用（Citation）
+
+当你使用网页搜索结果、文件内容或其他外部资料回答时，必须在对应事实陈述之后添加引用标记，前端会渲染为可点击的来源胶囊。用户明确要求「引用」某个文件/来源时，同样必须输出标记——用加粗、行内代码或表格提到文件名都不算引用。
+
+标记格式：
+- 网页来源：【cite:src_xxx】（src_xxx 必须来自本次对话工具结果中明确提供的引用 ID，如 `[src_xxx]` 前缀或「来源 ID: src_xxx」）
+- 文件来源：【file:工作区相对路径】，例如 【file:AGENTS.md】；指向具体行段：【file:AGENTS.md#L120-148】
+- 图片等二进制文件同样用 file 标记引用（如 【file:.astrion/user_upload/xxx.png】），前端会渲染缩略预览
+- 一个陈述引用多个来源：同类逗号并列【cite:src_xxx,src_yyy】/【file:report.pdf,notes.md】；网页与文件混排相邻写两个标记，例如 【cite:src_xxx】【file:report.pdf】
+
+引用要求：
+1. 只引用实际支持当前陈述的来源，不引用无关来源
+2. 引用紧跟在相关句子或段落末尾，不要统一放在回答结尾或集中列「参考资料」（除非用户要求）
+3. 没有使用外部资料时不要添加引用，也不要逐句标注
+4. 只能使用工具结果明确提供的 src_xxx 和工作区内真实存在的文件相对路径，禁止编造；标记内禁止写 URL
+5. 代码块和行内代码中的标记不会被渲染，不要在代码示例里放引用
+
+示例：
+- 该功能在 v2.3 版本引入。【cite:src_abc123def4】
+- 两个项目的实现方案一致。【cite:src_abc123def4,src_xyz987abcd】
+- 配置项说明见部署文档。【file:docs/deploy.md#L45-60】
+- 这一结论同时有网页和本地文档支持。【cite:src_abc123def4】【file:docs/deploy.md】
+
 ### 3.2 文件操作
 
 - `write_file`：写入文件（`append` 控制覆盖/追加）
diff --git a/prompts/multi_agent/master.txt b/prompts/multi_agent/master.txt
index f43a245d..97fac92c 100644
--- a/prompts/multi_agent/master.txt
+++ b/prompts/multi_agent/master.txt
@@ -283,6 +283,29 @@ id: ask_fse_001
 
 注意：路径必须是工作区内已存在的文件。输出前确保文件已通过 `write_file` 或其他方式创建。
 
+### 行内引用（Citation）
+
+你是最终向用户汇报的人。当你使用网页搜索结果、文件内容或其他外部资料回答时，必须在对应事实陈述之后添加引用标记，前端会渲染为可点击的来源胶囊。用户明确要求「引用」某个文件/来源时，同样必须输出标记——用加粗、行内代码或表格提到文件名都不算引用。
+
+标记格式：
+- 网页来源：【cite:src_xxx】（src_xxx 必须来自本次对话工具结果中明确提供的引用 ID，如 `[src_xxx]` 前缀或「来源 ID: src_xxx」；子智能体汇报中附带的来源 ID 同样可用）
+- 文件来源：【file:工作区相对路径】，例如 【file:AGENTS.md】；指向具体行段：【file:AGENTS.md#L120-148】
+- 图片等二进制文件同样用 file 标记引用（如 【file:.astrion/user_upload/xxx.png】），前端会渲染缩略预览
+- 一个陈述引用多个来源：同类逗号并列【cite:src_xxx,src_yyy】/【file:report.pdf,notes.md】；网页与文件混排相邻写两个标记，例如 【cite:src_xxx】【file:report.pdf】
+
+引用要求：
+1. 只引用实际支持当前陈述的来源，不引用无关来源
+2. 引用紧跟在相关句子或段落末尾，不要统一放在回答结尾或集中列「参考资料」（除非用户要求）
+3. 没有使用外部资料时不要添加引用，也不要逐句标注
+4. 只能使用工具结果或子智能体汇报中明确提供的 src_xxx 和工作区内真实存在的文件相对路径，禁止编造；标记内禁止写 URL
+5. 代码块和行内代码中的标记不会被渲染，不要在代码示例里放引用
+
+示例：
+- 该功能在 v2.3 版本引入。【cite:src_abc123def4】
+- 两个项目的实现方案一致。【cite:src_abc123def4,src_xyz987abcd】
+- 配置项说明见部署文档。【file:docs/deploy.md#L45-60】
+- 这一结论同时有网页和本地文档支持。【cite:src_abc123def4】【file:docs/deploy.md】
+
 ## 关于显示名
 
 - 主智能体固定显示名：`Team Leader`
diff --git a/server/chat_flow_task_main.py b/server/chat_flow_task_main.py
index f7a5914d..30945162 100644
--- a/server/chat_flow_task_main.py
+++ b/server/chat_flow_task_main.py
@@ -2034,6 +2034,8 @@ async def handle_task_with_sender(
     # 统计和限制变量
     total_iterations = 0
     total_tool_calls = 0
+    # 行内引用：任务级已用来源累积（task_complete 时带给前端实时渲染），按 id 去重
+    task_citations: Dict[str, dict] = {}
     consecutive_same_tool = defaultdict(int)
     last_tool_name = ""
     auto_fix_attempts = 0
@@ -2312,6 +2314,23 @@ async def handle_task_with_sender(
         
         assistant_content = "\n".join(assistant_content_parts) if assistant_content_parts else ""
         
+        # 行内引用：校验 marker、剥离无效引用，已用来源挂到消息 metadata 一并落库
+        round_citations = []
+        if assistant_content and ("【cite:" in assistant_content or "【file:" in assistant_content):
+            try:
+                from modules.citations import finalize_message_citations, get_registry
+                assistant_content, round_citations = finalize_message_citations(
+                    assistant_content,
+                    get_registry(web_terminal),
+                    str(getattr(web_terminal, "project_path", "") or ""),
+                )
+                for ann in round_citations:
+                    if ann.get("id"):
+                        task_citations[ann["id"]] = ann
+            except Exception as _cite_exc:
+                debug_log(f"[Citations] 行内引用处理失败，保留原文: {_cite_exc}")
+                round_citations = []
+        
         # 添加到消息历史（用于API继续对话，不保存到文件）
         assistant_message = {
             "role": "assistant",
@@ -2328,7 +2347,8 @@ async def handle_task_with_sender(
                 "assistant",
                 assistant_content,
                 tool_calls=tool_calls if tool_calls else None,
-                reasoning_content=current_thinking or ""
+                reasoning_content=current_thinking or "",
+                metadata={"citations": round_citations} if round_citations else None
             )
         
         # 为下一轮迭代重置流状态标志，但保留 full_response 供上面保存使用
@@ -2757,4 +2777,6 @@ async def handle_task_with_sender(
         # 前端就应保持运行态并继续轮询。
         'has_running_multi_agent': has_running_multi_agent or has_pending_ma_messages,
         'pending_runtime_guidance_messages': pending_runtime_guidance_messages,
+        # 行内引用：本任务全部已用来源，前端挂到当前 assistant 消息渲染 chip
+        'citations': list(task_citations.values()),
     })
diff --git a/static/src/App.vue b/static/src/App.vue
index ad909257..33fc65d3 100644
--- a/static/src/App.vue
+++ b/static/src/App.vue
@@ -404,6 +404,7 @@
         <FilePreviewPanel
           v-if="!isMobileViewport"
         />
+        <CitationPopover :host-mode="versioningHostMode" />
         <div
           v-if="!isMobileViewport && (terminalPanelOpen || gitChangesPanelOpen)"
           class="resize-handle resize-handle--right-panels"
@@ -843,6 +844,7 @@ import VideoPicker from './components/overlay/VideoPicker.vue';
 import ImageLightbox from './components/overlay/ImageLightbox.vue';
 import QuickDock from './components/chat/quickdock/QuickDock.vue';
 import FilePreviewPanel from './components/chat/quickdock/FilePreviewPanel.vue';
+import CitationPopover from './components/chat/CitationPopover.vue';
 import { useTutorialStore } from './stores/tutorial';
 import { usePersonalizationStore } from './stores/personalization';
 
diff --git a/static/src/app/methods/history.ts b/static/src/app/methods/history.ts
index 0abb37e2..96bc8c30 100644
--- a/static/src/app/methods/history.ts
+++ b/static/src/app/methods/history.ts
@@ -318,6 +318,26 @@ export const historyMethods = {
           };
         }
 
+        // 行内引用：工具循环的多条 assistant 持久化消息会合并成同一条前端消息，
+        // citations 需随合并按 id 去重汇聚，否则 chip 无数据可渲染
+        const msgCitations = message.metadata?.citations;
+        if (Array.isArray(msgCitations) && msgCitations.length) {
+          const merged = Array.isArray(currentAssistantMessage.metadata?.citations)
+            ? [...currentAssistantMessage.metadata.citations]
+            : [];
+          const seen = new Set(merged.map((c) => c?.id));
+          msgCitations.forEach((c) => {
+            if (c && c.id && !seen.has(c.id)) {
+              seen.add(c.id);
+              merged.push(c);
+            }
+          });
+          currentAssistantMessage.metadata = {
+            ...(currentAssistantMessage.metadata || {}),
+            citations: merged
+          };
+        }
+
         const content = message.content || '';
         const reasoningText = (message.reasoning_content || '').trim();
 
diff --git a/static/src/app/methods/taskPolling/lifecycle.ts b/static/src/app/methods/taskPolling/lifecycle.ts
index 7b3ac38a..79ac1640 100644
--- a/static/src/app/methods/taskPolling/lifecycle.ts
+++ b/static/src/app/methods/taskPolling/lifecycle.ts
@@ -446,6 +446,18 @@ export const lifecycleMethods = {
       debugLog('[TaskPolling] 任务完成');
     }
 
+    // 行内引用：任务完成事件带回权威来源（轮询通道；socket 通道在 useLegacySocket 同款处理），
+    // 挂到最后一条 assistant 消息 metadata，MarkdownRenderer watcher 据此做 chip 裁决与富化
+    if (Array.isArray(data?.citations) && Array.isArray(this.messages) && this.messages.length) {
+      const lastMsg = this.messages[this.messages.length - 1];
+      if (lastMsg && lastMsg.role === 'assistant') {
+        lastMsg.metadata = {
+          ...(lastMsg.metadata || {}),
+          citations: data.citations
+        };
+      }
+    }
+
     // 同步处理状态更新
     this.streamingMessage = false;
     this.stopRequested = false;
diff --git a/static/src/components/chat/ChatArea.vue b/static/src/components/chat/ChatArea.vue
index e6d25d4f..ade1eb42 100644
--- a/static/src/components/chat/ChatArea.vue
+++ b/static/src/components/chat/ChatArea.vue
@@ -187,6 +187,9 @@
             <MinimalBlocks
               v-if="hasRenderableAssistantActions(msg.actions || [])"
               :actions="msg.actions || []"
+              :citations="citationsForMessage(msg)"
+              :citations-final="citationsFinalForMessage(msg)"
+              enable-citations
               :conversation-running="streamingMessage"
               :is-latest-message="index === latestMessageIndex"
               :icon-style="iconStyleSafe"
@@ -308,6 +311,9 @@
                     <MarkdownRenderer
                       :content="group.action.content || ''"
                       :is-streaming="group.action.streaming"
+                      :citations="citationsForMessage(msg)"
+                      :citations-final="citationsFinalForMessage(msg)"
+                      enable-citations
                     />
                   </div>
                 </div>
@@ -533,6 +539,9 @@
                     <MarkdownRenderer
                       :content="action.content || ''"
                       :is-streaming="action.streaming"
+                      :citations="citationsForMessage(msg)"
+                      :citations-final="citationsFinalForMessage(msg)"
+                      enable-citations
                     />
                   </div>
                 </div>
@@ -739,6 +748,7 @@ import ToolAction from '@/components/chat/actions/ToolAction.vue';
 import StackedBlocks from './StackedBlocks.vue';
 import MinimalBlocks from './MinimalBlocks.vue';
 import MarkdownRenderer from './MarkdownRenderer.vue';
+import { collectConversationCitations, type CitationAnnotation } from './citationChips';
 import EditSummaryCard from './EditSummaryCard.vue';
 import { usePersonalizationStore } from '@/stores/personalization';
 import { useUiStore } from '@/stores/ui';
@@ -778,6 +788,22 @@ const personalizationReady = computed(() => {
 const renderPending = computed(() => {
   return !!props.historyLoading || !personalizationReady.value;
 });
+
+// 行内引用：对话级来源收集（扫描所有消息的工具结果 result.citations）。
+// 工具完成先于模型输出 marker，chip 在输出瞬间即可解析渲染，不等任务完成。
+const collectedCitations = computed<CitationAnnotation[]>(() =>
+  collectConversationCitations(props.messages || [])
+);
+/** 该消息可用的引用来源：权威（metadata.citations，后端裁决富化）优先，否则用收集表 */
+function citationsForMessage(msg: any): CitationAnnotation[] | undefined {
+  const meta = msg?.metadata?.citations;
+  if (Array.isArray(meta)) return meta;
+  return collectedCitations.value.length ? collectedCitations.value : undefined;
+}
+/** metadata.citations 到达即为权威裁决（含空数组=全部无效），触发无效 chip 移除与富化 */
+function citationsFinalForMessage(msg: any): boolean {
+  return Array.isArray(msg?.metadata?.citations);
+}
 const blockDisplayMode = computed(() => {
   return personalization.experiments.blockDisplayMode || 'stacked';
 });
diff --git a/static/src/components/chat/CitationPopover.vue b/static/src/components/chat/CitationPopover.vue
new file mode 100644
index 00000000..ebc37060
--- /dev/null
+++ b/static/src/components/chat/CitationPopover.vue
@@ -0,0 +1,564 @@
+<template>
+  <teleport to="body">
+    <div
+      v-if="citationPopover.visible"
+      ref="popoverEl"
+      class="citation-popover"
+      :style="popoverStyle"
+      @mouseenter="keepCitationPopover"
+      @mouseleave="leaveCitationPopover"
+      @click.stop
+    >
+      <!-- 单来源：完整卡片 -->
+      <template v-if="single">
+        <div
+          class="pop-header"
+          :class="{ clickable: isFile(single) && canPreviewFile }"
+          @click="onHeaderClick(single)"
+        >
+          <span class="pop-icon" v-html="iconHtml(single)"></span>
+          <span class="pop-domain">{{ headerText(single) }}</span>
+        </div>
+        <div class="pop-scroll">
+          <div class="pop-body">
+            <div class="pop-title-row">
+              <div class="pop-title">{{ single.title || single.file_name || '' }}</div>
+              <span v-if="locatorText(single)" class="pop-locator">{{ locatorText(single) }}</span>
+            </div>
+            <div v-if="single.url" class="pop-url">{{ single.url }}</div>
+            <div v-if="isImageFile(single)" class="pop-image">
+              <img :src="fileContentUrl(single)" :alt="single.file_name || ''" loading="lazy" />
+            </div>
+            <div v-else-if="displaySnippet(single)" class="pop-snippet">“{{ displaySnippet(single) }}”</div>
+          </div>
+        </div>
+        <div v-if="hasFooterAction" class="pop-footer">
+          <a
+            v-if="!isFile(single) && single.url"
+            class="pop-open"
+            :href="single.url"
+            target="_blank"
+            rel="noopener"
+          >{{ t('chat.citationOpenSource') }} ↗</a>
+          <span
+            v-else-if="isFile(single) && hostMode"
+            class="pop-open"
+            @click="openOnComputer(single)"
+          >{{ t('quickdock.menuRevealInManager') }}</span>
+        </div>
+      </template>
+
+      <!-- 多来源：列表 -->
+      <template v-else>
+        <div class="pop-header">
+          <span class="pop-domain">{{ t('chat.citationSources', { n: citationPopover.annotations.length }) }}</span>
+        </div>
+        <div class="pop-scroll">
+          <div
+            v-for="ann in citationPopover.annotations"
+            :key="ann.id"
+            class="pop-list-row"
+            @click="onRowClick(ann)"
+          >
+            <span class="pop-list-icon" v-html="iconHtml(ann)"></span>
+            <span class="pop-list-text">
+              <span class="pop-list-title">
+                <span class="pop-list-title-text">{{ ann.title || ann.file_name || '' }}</span>
+                <span v-if="isFile(ann) && locatorText(ann)" class="pop-locator pop-locator--sm">{{ locatorText(ann) }}</span>
+              </span>
+              <span class="pop-list-domain">{{ rowSubText(ann) }}</span>
+            </span>
+          </div>
+        </div>
+      </template>
+    </div>
+  </teleport>
+</template>
+
+<script setup lang="ts">
+import { computed, onBeforeUnmount, onMounted, ref, watch, nextTick } from 'vue';
+import { t } from '@/locales';
+import { useQuickDockStore } from '@/stores/quickDock';
+import { useUiStore } from '@/stores/ui';
+import {
+  citationPopover,
+  keepCitationPopover,
+  leaveCitationPopover,
+  closeCitationPopover,
+  type CitationAnnotation,
+} from './citationChips';
+
+defineOptions({ name: 'CitationPopover' });
+
+// hostMode：文件 footer 的「在文件管理器中打开」仅宿主机模式可用（docker 无此能力）
+const props = defineProps<{ hostMode?: boolean }>();
+
+const quickDock = useQuickDockStore();
+const uiStore = useUiStore();
+
+const single = computed<CitationAnnotation | null>(() =>
+  citationPopover.annotations.length === 1 ? citationPopover.annotations[0] : null,
+);
+
+// 底部有动作才渲染 footer：网页=打开来源；文件=在电脑上直接打开（仅宿主机）
+const hasFooterAction = computed(() => {
+  const s = single.value;
+  if (!s) return false;
+  if (isFile(s)) return !!props.hostMode;
+  return !!s.url;
+});
+
+const popoverEl = ref<HTMLElement | null>(null);
+const popoverStyle = ref<Record<string, string>>({});
+
+// 右侧文件预览面板移动端不渲染（App.vue 门禁），此处同步门禁
+const canPreviewFile = computed(() => !uiStore.isMobileViewport);
+
+function isFile(ann: CitationAnnotation) {
+  return ann.type === 'file_citation';
+}
+
+const IMAGE_EXTS = new Set(['.png', '.jpg', '.jpeg', '.gif', '.webp', '.svg', '.bmp', '.ico', '.avif']);
+
+function isImageFile(ann: CitationAnnotation): boolean {
+  if (!isFile(ann)) return false;
+  const p = (ann.file_path || ann.file_name || '').toLowerCase();
+  const dot = p.lastIndexOf('.');
+  return dot >= 0 && IMAGE_EXTS.has(p.slice(dot));
+}
+
+/** /api/file/content 的 inline 白名单含 image/，可直接作为 <img> 源 */
+function fileContentUrl(ann: CitationAnnotation): string {
+  return `/api/file/content?path=${encodeURIComponent(ann.file_path || '')}`;
+}
+
+function shortDomain(domain?: string) {
+  return (domain || '').replace(/^www\./, '');
+}
+
+function escapeText(s: string): string {
+  return s
+    .replace(/&/g, '&amp;')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;');
+}
+
+const FILE_ICON_SVG =
+  '<svg class="chip-file-icon" viewBox="0 0 16 16" fill="none">' +
+  '<path d="M4 1.5h5.5L13 5v9.5H4z" stroke="currentColor" stroke-width="1.2" stroke-linejoin="round"/>' +
+  '<path d="M9.5 1.5V5H13" stroke="currentColor" stroke-width="1.2" stroke-linejoin="round"/></svg>';
+
+function iconHtml(ann: CitationAnnotation): string {
+  if (isFile(ann)) return FILE_ICON_SVG;
+  const d = shortDomain(ann.domain);
+  const letter = (d[0] || '?').toUpperCase();
+  return (
+    `<img class="chip-favicon" loading="lazy" alt="" ` +
+    `src="https://www.google.com/s2/favicons?domain=${encodeURIComponent(d)}&sz=64" ` +
+    `onerror="this.outerHTML='<span class=&quot;chip-letter&quot;>${letter}</span>'">`
+  );
+}
+
+function headerText(ann: CitationAnnotation): string {
+  return isFile(ann) ? (ann.file_path || ann.file_name || '') : shortDomain(ann.domain);
+}
+
+function locatorText(ann: CitationAnnotation): string {
+  const bits: string[] = [];
+  if (ann.page != null) bits.push(`p.${ann.page}`);
+  if (ann.line_start != null) {
+    bits.push(
+      ann.line_end != null && ann.line_end !== ann.line_start
+        ? `L${ann.line_start}–${ann.line_end}`
+        : `L${ann.line_start}`,
+    );
+  }
+  return bits.join(' · ');
+}
+
+function rowSubText(ann: CitationAnnotation): string {
+  // 行号徽章已提到标题行，副行固定显示路径（网页显示域名）
+  if (isFile(ann)) return ann.file_path || '';
+  return shortDomain(ann.domain);
+}
+
+function openFile(ann: CitationAnnotation) {
+  if (!ann.file_path || !canPreviewFile.value) return;
+  quickDock.openPreview(ann.file_path);
+  closeCitationPopover();
+}
+
+/* ---------- 文件内容片段懒加载 ----------
+ * 后端落库时已富化 snippet；但旧消息（富化前持久化）与流式期的临时 annotation
+ * 没有 snippet，这里在弹层打开时按 file_path 拉一次内容补齐。 */
+const snippetCache = ref<Record<string, string>>({});
+
+function displaySnippet(ann: CitationAnnotation): string {
+  return ann.snippet || snippetCache.value[ann.id] || '';
+}
+
+async function maybeFetchSnippet(ann: CitationAnnotation) {
+  if (!isFile(ann) || isImageFile(ann) || ann.snippet || !ann.file_path || snippetCache.value[ann.id]) return;
+  const path = ann.file_path;
+  try {
+    const resp = await fetch(`/api/file/content?path=${encodeURIComponent(path)}`);
+    if (!resp.ok) return;
+    let text = await resp.text();
+    // 二进制守卫：空字节或大量替换字符则放弃
+    if (text.includes('\x00')) return;
+    const bad = (text.match(/\uFFFD/g) || []).length;
+    if (bad > Math.max(4, text.length * 0.005)) return;
+    if (ann.line_start != null) {
+      const lines = text.split('\n');
+      const s = Math.max(1, ann.line_start);
+      const e = Math.min(lines.length, ann.line_end ?? s);
+      text = lines.slice(s - 1, e).join('\n');
+    }
+    // 保留换行（弹层 pre-line 渲染），仅压缩行内空白、去空行
+    text = text
+      .split('\n')
+      .map((l) => l.trim())
+      .filter(Boolean)
+      .join('\n')
+      .slice(0, 300);
+    if (text) {
+      snippetCache.value = { ...snippetCache.value, [ann.id]: text };
+    }
+  } catch {
+    /* 读取失败静默：弹层退化为仅路径/定位信息 */
+  }
+}
+
+watch(
+  () => [citationPopover.visible, single.value?.id] as const,
+  ([visible]) => {
+    if (visible && single.value) maybeFetchSnippet(single.value);
+  },
+);
+
+/** 文件卡片头部点击 → 复用右侧文件预览面板 */
+function onHeaderClick(ann: CitationAnnotation) {
+  if (isFile(ann)) openFile(ann);
+}
+
+/** 在电脑上直接打开：复用 QuickDock 的链路（系统默认应用 = 候选列表第一个） */
+async function openOnComputer(ann: CitationAnnotation) {
+  const path = ann.file_path;
+  if (!path) return;
+  try {
+    const resp = await fetch(`/api/project/file-open-apps?path=${encodeURIComponent(path)}`);
+    const payload = await resp.json().catch(() => ({}));
+    if (!resp.ok || !payload?.success) {
+      throw new Error(payload?.error || t('quickdock.revealDetectAppsFailed'));
+    }
+    const apps = Array.isArray(payload?.data?.apps) ? payload.data.apps : [];
+    if (!apps.length) {
+      throw new Error(t('quickdock.revealNoAppsFound'));
+    }
+    const openResp = await fetch('/api/project/open-file-with-app', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ path, app_id: apps[0].id }),
+    });
+    const openPayload = await openResp.json().catch(() => ({}));
+    if (!openResp.ok || !openPayload?.success) {
+      throw new Error(openPayload?.error || t('quickdock.revealOpenFileFailed'));
+    }
+  } catch (err: any) {
+    uiStore.pushToast({ message: err?.message || t('quickdock.revealOpenFileFailed'), type: 'error' });
+    return;
+  }
+  closeCitationPopover();
+}
+
+function onRowClick(ann: CitationAnnotation) {
+  if (isFile(ann)) {
+    openFile(ann);
+  } else if (ann.url) {
+    window.open(ann.url, '_blank', 'noopener');
+    closeCitationPopover();
+  }
+}
+
+/** 定位：优先放胶囊下方，空间不足翻到上方；水平方向视口内夹取 */
+function positionPopover() {
+  const anchor = citationPopover.anchor;
+  const el = popoverEl.value;
+  if (!anchor || !el) return;
+  const rect = anchor.getBoundingClientRect();
+  const width = 320;
+  const maxHeight = 300;
+  const height = Math.min(el.offsetHeight || 200, maxHeight);
+  let left = Math.min(Math.max(8, rect.left), window.innerWidth - width - 8);
+  let top = rect.bottom + 6;
+  if (top + height > window.innerHeight - 8) {
+    top = Math.max(8, rect.top - height - 6);
+  }
+  popoverStyle.value = { left: `${left}px`, top: `${top}px` };
+}
+
+watch(
+  () => citationPopover.visible,
+  async (visible) => {
+    if (visible) {
+      await nextTick();
+      positionPopover();
+    }
+  },
+);
+
+const onGlobalClick = (e: MouseEvent) => {
+  if (popoverEl.value && !popoverEl.value.contains(e.target as Node)) {
+    closeCitationPopover();
+  }
+};
+const onKeydown = (e: KeyboardEvent) => {
+  if (e.key === 'Escape') closeCitationPopover();
+};
+// 页面一旦滚动就收起（含固定态），符合全局交互约定
+const onScroll = () => closeCitationPopover();
+const onResize = () => closeCitationPopover();
+
+onMounted(() => {
+  document.addEventListener('click', onGlobalClick, true);
+  document.addEventListener('keydown', onKeydown);
+  window.addEventListener('scroll', onScroll, true);
+  window.addEventListener('resize', onResize);
+});
+
+onBeforeUnmount(() => {
+  document.removeEventListener('click', onGlobalClick, true);
+  document.removeEventListener('keydown', onKeydown);
+  window.removeEventListener('scroll', onScroll, true);
+  window.removeEventListener('resize', onResize);
+});
+</script>
+
+<style scoped>
+/* 实体面板：不透明、无 backdrop-filter；中性阴影；颜色全走语义 token */
+.citation-popover {
+  position: fixed;
+  z-index: 1000;
+  width: 320px;
+  max-height: 300px;
+  display: flex;
+  flex-direction: column;
+  background: var(--surface-card);
+  border: 1px solid var(--border-default);
+  border-radius: 10px;
+  box-shadow: var(--shadow-card);
+  overflow: hidden;
+}
+
+.pop-scroll {
+  overflow-y: auto;
+  min-height: 0;
+  scrollbar-width: none;
+}
+.pop-scroll::-webkit-scrollbar {
+  width: 0;
+  height: 0;
+}
+
+.pop-header {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  height: 34px;
+  padding: 0 12px;
+  flex: none;
+  border-bottom: 1px solid var(--border-default);
+  font-size: 11px;
+  color: var(--text-tertiary);
+}
+.pop-header.clickable {
+  cursor: pointer;
+}
+.pop-header.clickable:hover {
+  color: var(--text-secondary);
+}
+.pop-icon {
+  display: inline-flex;
+  align-items: center;
+  flex: none;
+}
+.pop-domain {
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+  font-weight: 500;
+}
+
+.pop-body {
+  padding: 10px 12px 12px;
+}
+.pop-title-row {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  margin-bottom: 4px;
+}
+.pop-title {
+  font-size: 13.5px;
+  font-weight: 600;
+  line-height: 1.4;
+  color: var(--text-primary);
+  min-width: 0;
+  display: -webkit-box;
+  -webkit-line-clamp: 2;
+  -webkit-box-orient: vertical;
+  overflow: hidden;
+}
+.pop-url {
+  font-size: 11.5px;
+  color: var(--text-tertiary);
+  white-space: nowrap;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  margin-bottom: 8px;
+}
+.pop-locator {
+  flex: none;
+  display: inline-flex;
+  align-items: center;
+  height: 20px;
+  padding: 0 8px;
+  border-radius: 999px;
+  background: var(--surface-muted);
+  font-size: 11px;
+  color: var(--text-secondary);
+}
+.pop-snippet {
+  font-size: 12.5px;
+  line-height: 1.55;
+  color: var(--text-secondary);
+  white-space: pre-line; /* 保留文件内容换行 */
+  display: -webkit-box;
+  -webkit-line-clamp: 4;
+  -webkit-box-orient: vertical;
+  overflow: hidden;
+}
+.pop-image {
+  margin-top: 6px;
+}
+.pop-image img {
+  display: block;
+  max-width: 100%;
+  max-height: 180px;
+  border-radius: 6px;
+}
+.pop-footer {
+  flex: none;
+  height: 38px;
+  padding: 0 12px;
+  border-top: 1px solid var(--border-default);
+  display: flex;
+  align-items: center;
+}
+.pop-open {
+  font-size: 12.5px;
+  color: var(--accent);
+  font-weight: 500;
+  display: inline-flex;
+  align-items: center;
+  gap: 4px;
+  text-decoration: none;
+  cursor: pointer;
+}
+.pop-open:hover {
+  color: var(--accent-hover);
+  text-decoration: underline;
+}
+
+.pop-list-row {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  height: 44px;
+  padding: 0 12px;
+  cursor: pointer;
+}
+.pop-list-row:hover {
+  background: var(--hover-bg);
+}
+.pop-list-row + .pop-list-row {
+  border-top: 1px solid var(--border-default);
+}
+.pop-list-icon {
+  flex: none;
+  width: 16px;
+  height: 16px;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+}
+.pop-list-icon :deep(.chip-favicon) {
+  width: 14px;
+  height: 14px;
+}
+.pop-list-icon :deep(.chip-file-icon) {
+  width: 14px;
+  height: 14px;
+  color: var(--text-tertiary); /* 显式颜色绑定，stroke=currentColor 才能生效 */
+}
+.pop-list-text {
+  min-width: 0;
+  flex: 1;
+  display: flex;
+  flex-direction: column;
+}
+.pop-list-title {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  min-width: 0;
+  font-size: 12.5px;
+  color: var(--text-primary);
+  line-height: 1.3;
+}
+.pop-list-title-text {
+  min-width: 0;
+  white-space: nowrap;
+  overflow: hidden;
+  text-overflow: ellipsis;
+}
+/* 列表行内的小型行号徽章（44px 行高内容纳得下） */
+.pop-locator--sm {
+  height: 16px;
+  padding: 0 6px;
+  font-size: 10px;
+}
+.pop-list-domain {
+  font-size: 11px;
+  color: var(--text-tertiary);
+  white-space: nowrap;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  line-height: 1.3;
+}
+
+.pop-header :deep(.chip-favicon) {
+  width: 12px;
+  height: 12px;
+  border-radius: 3px;
+}
+.pop-header :deep(.chip-file-icon) {
+  width: 12px;
+  height: 12px;
+  color: var(--text-tertiary);
+}
+.pop-header :deep(.chip-letter),
+.pop-list-icon :deep(.chip-letter) {
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  width: 14px;
+  height: 14px;
+  border-radius: 3px;
+  font-size: 9px;
+  font-weight: 700;
+  color: var(--on-accent);
+  background: var(--accent);
+}
+</style>
diff --git a/static/src/components/chat/MarkdownRenderer.vue b/static/src/components/chat/MarkdownRenderer.vue
index b6421af2..238c1837 100644
--- a/static/src/components/chat/MarkdownRenderer.vue
+++ b/static/src/components/chat/MarkdownRenderer.vue
@@ -32,12 +32,17 @@ import {
   type MarkdownSegment
 } from '@/composables/useMarkdownRenderer';
 import { chunkRenderedHtml, type HtmlChunk } from '@/utils/htmlChunks';
+import { enhanceCitationChips, type CitationAnnotation } from './citationChips';
 
 defineOptions({ name: 'MarkdownRenderer' });
 
 const props = defineProps<{
   content: string;
   isStreaming?: boolean;
+  citations?: CitationAnnotation[];
+  /** citations 是否为后端裁决后的权威版本（message.metadata.citations 已到达） */
+  citationsFinal?: boolean;
+  enableCitations?: boolean;
 }>();
 
 const containerRef = ref<HTMLElement | null>(null);
@@ -47,7 +52,8 @@ const segments = computed(() => parseMarkdownSegments(props.content || '', props
 
 function renderText(text: string) {
   // 透传流式标志：show_html 卡片在流式期间需要 partial 渲染（实时渲染/渲染中占位）
-  return renderMarkdownText(text, props.isStreaming);
+  // enableCitations：仅 assistant 正文开启引用渲染，其他场景【cite:】按原文显示
+  return renderMarkdownText(text, props.isStreaming, props.enableCitations);
 }
 
 // 分段级分块缓存：segments 每个 token 都是新对象，按 key 缓存避免对
@@ -83,6 +89,22 @@ function renderMath() {
 onMounted(renderMath);
 onUpdated(renderMath);
 watch(() => props.content, renderMath, { immediate: true });
+
+// 行内引用：渲染后扫描 chip 占位 span，填充内容并接管交互。
+// chip 在输出瞬间即解析（工具结果查表 / file token 自解析）；
+// citationsFinal 到达时做权威裁决（移除无效 chip）与富化。
+function enhanceCitations() {
+  if (!props.enableCitations) return;
+  nextTick(() => {
+    if (containerRef.value) {
+      enhanceCitationChips(containerRef.value, props.citations, { final: props.citationsFinal });
+    }
+  });
+}
+
+onMounted(enhanceCitations);
+onUpdated(enhanceCitations);
+watch(() => [props.citations, props.citationsFinal], enhanceCitations);
 </script>
 
 <style scoped>
diff --git a/static/src/components/chat/MinimalBlocks.vue b/static/src/components/chat/MinimalBlocks.vue
index 430cab02..4cc79347 100644
--- a/static/src/components/chat/MinimalBlocks.vue
+++ b/static/src/components/chat/MinimalBlocks.vue
@@ -121,7 +121,7 @@
         :class="{ 'streaming-text': group.streaming }"
       >
         <div class="text-content" :class="{ 'streaming-text': group.streaming }">
-          <MarkdownRenderer :content="group.content || ''" :is-streaming="group.streaming" />
+          <MarkdownRenderer :content="group.content || ''" :is-streaming="group.streaming" :citations="citations" :citations-final="citationsFinal" :enable-citations="enableCitations" />
         </div>
       </div>
       <div v-else-if="group.type === 'system'" class="summary-group sub-agent-system-group">
@@ -159,6 +159,7 @@ const personalizationStore = usePersonalizationStore();
 const heightLimited = computed(() => personalizationStore.form.minimal_expand_height_limited);
 import { getRandomLoader } from './loaders/index';
 import MarkdownRenderer from './MarkdownRenderer.vue';
+import type { CitationAnnotation } from './citationChips';
 
 interface Action {
   id?: string;
@@ -196,6 +197,9 @@ interface BlockGroup {
 
 const props = defineProps<{
   actions: Action[];
+  citations?: CitationAnnotation[];
+  citationsFinal?: boolean;
+  enableCitations?: boolean;
   conversationRunning?: boolean;
   isLatestMessage?: boolean;
   iconStyle?: (name: string) => any;
diff --git a/static/src/components/chat/citationChips.ts b/static/src/components/chat/citationChips.ts
new file mode 100644
index 00000000..ecb35585
--- /dev/null
+++ b/static/src/components/chat/citationChips.ts
@@ -0,0 +1,275 @@
+/**
+ * 行内引用（Inline Citations）：chip 增强与 popover 状态控制。
+ *
+ * 数据流：消息文本中的【cite:...】marker 经 useMarkdownRenderer 转换为
+ * <span class="md-citation-chip" data-citation-ids="...">，本模块在
+ * MarkdownRenderer 渲染后扫描容器，用 message.metadata.citations 填充
+ * chip 内容并接管交互（悬停延迟开 / 移出延迟关 / 点击固定 / 滚动收起）。
+ */
+import { reactive } from 'vue';
+
+export interface CitationAnnotation {
+  id: string;
+  type: 'url_citation' | 'file_citation';
+  // url 来源
+  title?: string;
+  url?: string;
+  domain?: string;
+  published_date?: string;
+  // 文件来源
+  file_path?: string;
+  file_name?: string;
+  size?: number;
+  // 可选定位
+  page?: number;
+  line_start?: number;
+  line_end?: number;
+  snippet?: string;
+}
+
+/** popover 全局状态（单例；同一时间只存在一个引用弹层） */
+export const citationPopover = reactive({
+  visible: false,
+  pinned: false,
+  anchor: null as HTMLElement | null,
+  annotations: [] as CitationAnnotation[],
+});
+
+const HOVER_OPEN_DELAY = 250; // 悬停多久后打开
+const HOVER_CLOSE_DELAY = 300; // 移出多久后关闭（留出移到弹层上的余量）
+
+/** chip 元素上挂的当前 annotations（final 到达后可被富化替换，触发器事件时读取） */
+type ChipWithAnnotations = HTMLElement & { _citationAnnotations?: CitationAnnotation[] };
+
+let openTimer: ReturnType<typeof setTimeout> | undefined;
+let closeTimer: ReturnType<typeof setTimeout> | undefined;
+
+function clearPopoverTimers() {
+  clearTimeout(openTimer);
+  clearTimeout(closeTimer);
+  openTimer = undefined;
+  closeTimer = undefined;
+}
+
+export function openCitationPopover(anchor: HTMLElement, annotations: CitationAnnotation[], pinned: boolean) {
+  closeCitationPopover();
+  citationPopover.anchor = anchor;
+  citationPopover.annotations = annotations;
+  citationPopover.pinned = pinned;
+  citationPopover.visible = true;
+  anchor.classList.add('is-open');
+}
+
+export function closeCitationPopover() {
+  clearPopoverTimers();
+  if (citationPopover.anchor) {
+    citationPopover.anchor.classList.remove('is-open');
+  }
+  citationPopover.visible = false;
+  citationPopover.pinned = false;
+  citationPopover.anchor = null;
+  citationPopover.annotations = [];
+}
+
+function schedulePopoverClose() {
+  clearTimeout(closeTimer);
+  closeTimer = setTimeout(() => {
+    if (!citationPopover.pinned) closeCitationPopover();
+  }, HOVER_CLOSE_DELAY);
+}
+
+/** popover 自身 hover 进入时取消关闭（CitationPopover 组件调用） */
+export function keepCitationPopover() {
+  clearTimeout(closeTimer);
+}
+
+/** popover 自身 hover 离开时按延迟关闭（未固定时） */
+export function leaveCitationPopover() {
+  if (!citationPopover.pinned) schedulePopoverClose();
+}
+
+function attachChipTriggers(chip: ChipWithAnnotations) {
+  // 事件时从元素上读最新 annotations（final 富化替换后无需重绑监听器）
+  const current = () => chip._citationAnnotations || [];
+  chip.addEventListener('mouseenter', () => {
+    clearTimeout(closeTimer);
+    if (citationPopover.anchor === chip && citationPopover.visible) return;
+    clearTimeout(openTimer);
+    openTimer = setTimeout(() => openCitationPopover(chip, current(), false), HOVER_OPEN_DELAY);
+  });
+  chip.addEventListener('mouseleave', () => {
+    clearTimeout(openTimer);
+    if (citationPopover.visible && citationPopover.anchor === chip && !citationPopover.pinned) {
+      schedulePopoverClose();
+    }
+  });
+  chip.addEventListener('click', (e) => {
+    e.stopPropagation();
+    clearPopoverTimers();
+    if (citationPopover.anchor === chip && citationPopover.visible && citationPopover.pinned) {
+      closeCitationPopover(); // 再点已固定的弹层 → 收起
+    } else {
+      openCitationPopover(chip, current(), true); // 点击 = 固定
+    }
+  });
+}
+
+/** 从文件引用 token（如 docs/x.md#L1-10）直接解析出临时 annotation：
+ *  marker 自带路径与定位，chip 在输出瞬间即可渲染，无需等后端富化。 */
+export function fileAnnotationFromToken(token: string): CitationAnnotation | null {
+  const raw = (token || '').trim();
+  if (!raw || /^src_/i.test(raw)) return null;
+  // 兼容 #L1-10 与 GitHub 风格 #L1-L10 / #p7
+  const m = raw.match(/^(.*?)#(?:L(\d+)(?:-L?(\d+))?|p(\d+))$/i);
+  let path = (m ? m[1] : raw).trim();
+  if (path.startsWith('./')) path = path.slice(2);
+  if (!path) return null;
+  const ann: CitationAnnotation = {
+    id: token,
+    type: 'file_citation',
+    file_path: path,
+    file_name: path.split('/').pop() || path,
+  };
+  if (m) {
+    if (m[4]) ann.page = parseInt(m[4], 10);
+    if (m[2]) {
+      ann.line_start = parseInt(m[2], 10);
+      if (m[3]) ann.line_end = parseInt(m[3], 10);
+    }
+  }
+  return ann;
+}
+
+/** 扫描对话消息中的工具结果，收集 web_search / extract_webpage 注册的来源 annotation。
+ *  工具完成先于模型输出 marker，因此流式期间即可解析 cite:src_xxx。 */
+export function collectConversationCitations(messages: any[]): CitationAnnotation[] {
+  const map = new Map<string, CitationAnnotation>();
+  for (const msg of messages || []) {
+    for (const action of msg?.actions || []) {
+      const list = action?.tool?.result?.citations;
+      if (!Array.isArray(list)) continue;
+      for (const ann of list) {
+        if (ann && typeof ann.id === 'string' && !map.has(ann.id)) {
+          map.set(ann.id, ann);
+        }
+      }
+    }
+  }
+  return [...map.values()];
+}
+
+function shortDomain(domain?: string): string {
+  return (domain || '').replace(/^www\./, '');
+}
+
+function chipLabel(ann: CitationAnnotation): string {
+  return ann.type === 'file_citation' ? (ann.file_name || ann.file_path || '') : shortDomain(ann.domain);
+}
+
+function faviconHtml(domain: string): string {
+  const d = shortDomain(domain);
+  // 站点图标加载失败时退化为域名首字符
+  const letter = (d[0] || '?').toUpperCase();
+  return (
+    `<img class="chip-favicon" loading="lazy" alt="" ` +
+    `src="https://www.google.com/s2/favicons?domain=${encodeURIComponent(d)}&sz=64" ` +
+    `onerror="this.outerHTML='<span class=&quot;chip-letter&quot;>${letter}</span>'">`
+  );
+}
+
+const FILE_ICON_SVG =
+  '<svg class="chip-file-icon" viewBox="0 0 16 16" fill="none">' +
+  '<path d="M4 1.5h5.5L13 5v9.5H4z" stroke="currentColor" stroke-width="1.2" stroke-linejoin="round"/>' +
+  '<path d="M9.5 1.5V5H13" stroke="currentColor" stroke-width="1.2" stroke-linejoin="round"/></svg>';
+
+function chipIconHtml(ann: CitationAnnotation): string {
+  return ann.type === 'file_citation' ? FILE_ICON_SVG : faviconHtml(ann.domain || '');
+}
+
+/**
+ * 扫描容器内的 chip 占位 span，填充内容并接管交互。
+ *
+ * 两阶段解析：
+ * - 即时（输出瞬间）：cite 来源从对话工具结果收集的 citations 查表；
+ *   file 来源 token 自解析（marker 自带路径/locator，无需等待后端）。
+ * - 权威（final=true，message.metadata.citations 到达）：仅以权威表为准，
+ *   查不到的 chip 移除（后端已剥离无效 marker）；已渲染 chip 用富化数据
+ *   替换引用（popover 可读 snippet/size），DOM 不重建。
+ */
+export function enhanceCitationChips(
+  container: HTMLElement,
+  citations: CitationAnnotation[] | undefined,
+  opts: { final?: boolean } = {},
+) {
+  const chips = container.querySelectorAll<ChipWithAnnotations>('.md-citation-chip');
+  if (!chips.length) return;
+  const final = !!opts.final;
+  const map = new Map((citations || []).map((c) => [c.id, c]));
+
+  chips.forEach((chip) => {
+    const ids = (chip.dataset.citationIds || '')
+      .split(',')
+      .map((s) => s.trim())
+      .filter(Boolean);
+
+    // 已渲染的 chip：final 到达时换权威富化数据（供 popover 读取），DOM 不动
+    if (chip.dataset.citeEnhanced === '1' && chip.dataset.citeLoaded === '1') {
+      if (final) {
+        const enriched = ids
+          .map((id) => map.get(id))
+          .filter((a): a is CitationAnnotation => !!a);
+        if (enriched.length) chip._citationAnnotations = enriched;
+      }
+      return;
+    }
+
+    let annotations = ids
+      .map((id) => map.get(id))
+      .filter((a): a is CitationAnnotation => !!a);
+
+    // 非 final：file token 自解析作为临时数据（查表未命中时），让 chip 即时渲染
+    if (!annotations.length && !final) {
+      const derived = ids
+        .map((token) => fileAnnotationFromToken(token))
+        .filter((a): a is CitationAnnotation => !!a);
+      if (derived.length) annotations = derived;
+    }
+
+    if (!annotations.length) {
+      if (final) {
+        // 权威裁决：后端已剥离该 marker（幻觉 id / 文件不存在），移除 chip
+        chip.remove();
+        return;
+      }
+      // 流式占位：中性胶囊，暂不可交互
+      if (chip.dataset.citeEnhanced !== '1') {
+        chip.dataset.citeEnhanced = '1';
+        chip.classList.add('is-pending');
+        chip.innerHTML = '<span class="chip-pending-dot"></span>';
+      }
+      return;
+    }
+
+    const icons = annotations.slice(0, 2).map(chipIconHtml).join('');
+    const more =
+      annotations.length > 1 ? `<span class="chip-count">+${annotations.length - 1}</span>` : '';
+    chip.innerHTML =
+      `<span class="chip-icons">${icons}</span>` +
+      `<span class="chip-label">${escapeText(chipLabel(annotations[0]))}</span>` +
+      more;
+    chip.classList.remove('is-pending');
+    chip.classList.add(annotations[0].type === 'file_citation' ? 'is-file' : 'is-url');
+    chip.dataset.citeEnhanced = '1';
+    chip.dataset.citeLoaded = '1';
+    chip._citationAnnotations = annotations;
+    attachChipTriggers(chip);
+  });
+}
+
+function escapeText(s: string): string {
+  return s
+    .replace(/&/g, '&amp;')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;');
+}
diff --git a/static/src/composables/useLegacySocket.ts b/static/src/composables/useLegacySocket.ts
index f58e7fe9..ee69c2e0 100644
--- a/static/src/composables/useLegacySocket.ts
+++ b/static/src/composables/useLegacySocket.ts
@@ -1802,6 +1802,18 @@ export async function initializeLegacySocket(ctx: any) {
 
       resetPendingToolEvents();
 
+      // 行内引用：任务完成时后端带回已用来源，挂到最后一条 assistant 消息的 metadata，
+      // MarkdownRenderer 的 citations watcher 会触发 chip 增强（流式期间的占位胶囊在此转正）
+      if (Array.isArray(data?.citations) && Array.isArray(ctx.messages) && ctx.messages.length) {
+        const lastMsg = ctx.messages[ctx.messages.length - 1];
+        if (lastMsg && lastMsg.role === 'assistant') {
+          lastMsg.metadata = {
+            ...(lastMsg.metadata || {}),
+            citations: data.citations
+          };
+        }
+      }
+
       // 任务完成后立即更新Token统计（关键修复）
       if (ctx.currentConversationId) {
         ctx.updateCurrentContextTokens();
diff --git a/static/src/composables/useMarkdownRenderer.ts b/static/src/composables/useMarkdownRenderer.ts
index 7e5aa581..a6e9bbbd 100644
--- a/static/src/composables/useMarkdownRenderer.ts
+++ b/static/src/composables/useMarkdownRenderer.ts
@@ -313,6 +313,45 @@ function transformShowFileBlocks(raw: string) {
   return output;
 }
 
+/**
+ * 预处理行内引用 marker（【cite:src_xxx】网页 / 【file:相对路径】文件），转换为 citation chip 占位 span。
+ * - 仅转换完整闭合的 marker；流式期间尾部未闭合的【…暂时隐藏，避免原始字符闪烁
+ * - 行内代码内的 marker 保持原样（fenced code block 已在 segment 拆分阶段剥离）
+ * - marker 内容进 data 属性前做属性转义，防注入
+ */
+function transformCitationMarkers(raw: string, isStreaming: boolean): string {
+  if (!raw || raw.indexOf('【') === -1) return raw;
+
+  // 保护行内代码：`...` 内的 marker 原样展示
+  const INLINE_CODE_PLACEHOLDER = '__CITE_PROTECT_INLINE_CODE_';
+  const inlineCodes: string[] = [];
+  let text = raw.replace(/`[^`\n]+`/g, (match) => {
+    inlineCodes.push(match);
+    return `${INLINE_CODE_PLACEHOLDER}${inlineCodes.length - 1}__`;
+  });
+
+  text = text.replace(/【(?:cite|file):([^】]+)】/g, (_m, ids: string) => {
+    const clean = ids
+      .split(/[,，]/)
+      .map((s) => s.trim())
+      .filter(Boolean)
+      .map((s) => s.replace(/&/g, '&amp;').replace(/"/g, '&quot;').replace(/</g, '&lt;').replace(/>/g, '&gt;'))
+      .join(',');
+    if (!clean) return '';
+    return `<span class="md-citation-chip" data-citation-ids="${clean}"></span>`;
+  });
+
+  if (isStreaming) {
+    // 尾部未闭合 marker（【cite:... 甚至只输入了一半的【）暂时隐藏
+    text = text.replace(/【[^】]*$/, '');
+  }
+
+  // 还原行内代码
+  text = text.replace(new RegExp(`${INLINE_CODE_PLACEHOLDER}(\\d+)__`, 'g'), (_m, i: string) => inlineCodes[Number(i)]);
+
+  return text;
+}
+
 /**
  * 预处理 LaTeX 数学公式，避免被 markdown parser 拆段。
  * $$...$$ 转换为块级占位符，$...$ 转换为行内占位符，
@@ -532,7 +571,7 @@ const sanitizedSchema: Record<string, any> = {
   attributes: {
     ...(defaultSchema.attributes || {}),
     div: [...((defaultSchema.attributes || {}).div || []), 'className', 'data-md-table-scroll'],
-    span: [...((defaultSchema.attributes || {}).span || []), 'className', 'dataLatex', 'dataDisplay', 'dataMathRendered', 'data-latex', 'data-display', 'data-math-rendered'],
+    span: [...((defaultSchema.attributes || {}).span || []), 'className', 'dataLatex', 'dataDisplay', 'dataMathRendered', 'data-latex', 'data-display', 'data-math-rendered', 'data-citation-ids', 'dataCitationIds'],
     a: [
       'ariaDescribedBy', 'ariaLabel', 'ariaLabelledBy',
       'dataFootnoteBackref', 'dataFootnoteRef',
@@ -719,15 +758,17 @@ export function parseMarkdownSegments(text: string, isStreaming = false): Markdo
   return segments;
 }
 
-export function renderMarkdownText(text: string, isStreaming = false): string {
+export function renderMarkdownText(text: string, isStreaming = false, enableCitations = false): string {
   if (!text) return '';
 
   // isStreaming 必须透传：流式期间未闭合的 show_html 需要编码成 data-partial 占位
   // （js=off 实时渲染 / js=on 显示"渲染中"），否则原始标签文本会直接散落到消息里
+  // enableCitations：仅 assistant 正文开启；用户消息/预览等静态文本里的【cite:】原样显示
+  const withCustomBlocks = transformShowFileBlocks(
+    transformShowImageBlocks(transformShowHtmlBlocks(text, isStreaming))
+  );
   const safeText = transformMathBlocks(
-    transformShowFileBlocks(
-      transformShowImageBlocks(transformShowHtmlBlocks(text, isStreaming))
-    )
+    enableCitations ? transformCitationMarkers(withCustomBlocks, isStreaming) : withCustomBlocks
   );
 
   let html = '';
diff --git a/static/src/locales/en-US/chat.ts b/static/src/locales/en-US/chat.ts
index edc3af2c..af6d45be 100644
--- a/static/src/locales/en-US/chat.ts
+++ b/static/src/locales/en-US/chat.ts
@@ -37,6 +37,10 @@ export default {
   thinking: 'Thinking',
   thinkingRunning: 'Thinking...',
 
+  // —— Inline citations (citation chip / popover) ——
+  citationSources: '{n} sources',
+  citationOpenSource: 'Open source',
+
   // —— File append (append / append_payload) ——
   targetFile: 'target file',
   appendDone: 'File append complete',
diff --git a/static/src/locales/zh-CN/chat.ts b/static/src/locales/zh-CN/chat.ts
index d0c69285..47aab241 100644
--- a/static/src/locales/zh-CN/chat.ts
+++ b/static/src/locales/zh-CN/chat.ts
@@ -47,6 +47,10 @@ export default {
   thinking: '思考过程',
   thinkingRunning: '正在思考...',
 
+  // —— 行内引用（citation chip / popover） ——
+  citationSources: '{n} 个来源',
+  citationOpenSource: '打开来源',
+
   // —— 文件追加（append / append_payload） ——
   targetFile: '目标文件',
   appendDone: '文件追加完成',
diff --git a/static/src/styles/components/chat/_chat-area.scss b/static/src/styles/components/chat/_chat-area.scss
index 32c43c78..d21d221d 100644
--- a/static/src/styles/components/chat/_chat-area.scss
+++ b/static/src/styles/components/chat/_chat-area.scss
@@ -2463,6 +2463,106 @@ a.md-download-link {
   }
 }
 
+// ===== 行内引用胶囊（citation chip） =====
+// 交互约定：hover / 展开态只变色，不动边框；弹层交互见 CitationPopover.vue
+.md-citation-chip {
+  display: inline-flex;
+  align-items: center;
+  gap: 4px;
+  height: 20px;
+  padding: 0 7px;
+  margin: 0 2px;
+  border-radius: 999px;
+  font-size: 12px;
+  line-height: 1;
+  white-space: nowrap;
+  vertical-align: middle; // 几何中心对齐「基线 + x字高/2」——与文字完全上下居中，且 20px 盒仍在 25.5px 行框内不撑行
+  background: var(--surface-muted);
+  border: 1px solid var(--border-default);
+  color: var(--text-secondary);
+  cursor: pointer;
+  user-select: none;
+  transition: background 0.15s, color 0.15s;
+
+  &:hover,
+  &.is-open {
+    background: var(--hover-bg);
+    color: var(--text-primary);
+  }
+
+  &.is-pending {
+    cursor: default;
+  }
+
+  .chip-icons {
+    display: inline-flex;
+    align-items: center;
+    flex: none;
+  }
+
+  .chip-favicon {
+    width: 12px;
+    height: 12px;
+    border-radius: 3px;
+    flex: none;
+  }
+
+  .chip-favicon + .chip-favicon {
+    margin-left: -5px;
+    border: 1px solid var(--surface-muted);
+  }
+
+  .chip-file-icon {
+    width: 12px;
+    height: 12px;
+    flex: none;
+    color: var(--text-tertiary);
+  }
+
+  .chip-letter {
+    display: inline-flex;
+    align-items: center;
+    justify-content: center;
+    width: 12px;
+    height: 12px;
+    border-radius: 3px;
+    flex: none;
+    font-size: 8px;
+    font-weight: 700;
+    color: var(--on-accent);
+    background: var(--accent);
+  }
+
+  .chip-label {
+    max-width: 140px;
+    overflow: hidden;
+    text-overflow: ellipsis;
+    font-weight: 500;
+    // 行高 1 会把 j/g/y 等字母的下沉笔画裁掉；放大行高让 descender 落入盒内
+    // （chip 总高 20px 由自身 height 锁定，行高放大不撑高）
+    line-height: 1.4;
+  }
+
+  .chip-count {
+    color: var(--text-tertiary);
+    font-weight: 500;
+    line-height: 1.4;
+  }
+
+  &:hover .chip-count,
+  &.is-open .chip-count {
+    color: var(--text-secondary);
+  }
+
+  .chip-pending-dot {
+    width: 4px;
+    height: 4px;
+    border-radius: 50%;
+    background: var(--text-tertiary);
+    margin: 0 4px;
+  }
+}
+
 // ===== 数学公式样式 =====
 .math-block {
   display: block;
diff --git a/utils/tool_result_formatter/web_media.py b/utils/tool_result_formatter/web_media.py
index cf5357f4..7349c52a 100644
--- a/utils/tool_result_formatter/web_media.py
+++ b/utils/tool_result_formatter/web_media.py
@@ -33,7 +33,9 @@ def _format_extract_webpage(result_data: Dict[str, Any]) -> str:
     content = result_data.get("content") or ""
     length = len(content)
     truncated_flag = result_data.get("truncated") or False
-    header = f"提取完成：{url}，长度 {length} 字符。"
+    citation_id = result_data.get("citation_id")
+    citation_note = f"（来源 ID: {citation_id}）" if citation_id else ""
+    header = f"提取完成：{url}{citation_note}，长度 {length} 字符。"
     if not content:
         return f"{header} 内容为空。"
     # 为模型保留完整正文，避免 800 字预览导致上下文缺失