From ba090178846611117ae676763f615a7e900af191 Mon Sep 17 00:00:00 2001
From: QihuAI Design <ai@qihuyun.com>
Date: Tue, 14 Jul 2026 17:11:12 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20AGNES=E5=9B=BE=E5=83=8F/=E8=A7=86?=
 =?UTF-8?q?=E9=A2=91=E7=94=9F=E6=88=90=E5=BC=BA=E5=88=B6=E7=BB=8F=E6=99=BA?=
 =?UTF-8?q?=E8=83=BD=E8=B7=AF=E7=94=B1=E4=BB=A3=E7=90=86(172.18.0.1:7890)?=
 =?UTF-8?q?=EF=BC=8C=E4=BF=AE=E5=A4=8D=E5=AE=B9=E5=99=A8=E5=86=85=E7=9B=B4?=
 =?UTF-8?q?=E8=BF=9E=E8=A2=AB=E9=87=8D=E7=BD=AE=E5=AF=BC=E8=87=B4=E5=9B=BE?=
 =?UTF-8?q?=E5=83=8F=E7=94=9F=E6=88=90=E5=BC=82=E5=B8=B8?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 backend/python/app/api/chat.py | 3002 ++++++++++++++++----------------
 1 file changed, 1505 insertions(+), 1497 deletions(-)

diff --git a/backend/python/app/api/chat.py b/backend/python/app/api/chat.py
index 4b51be8..dac9057 100644
--- a/backend/python/app/api/chat.py
+++ b/backend/python/app/api/chat.py
@@ -1,1497 +1,1505 @@
-"""
-增强版 Chat API — Phase 2.5
-包含: Skill ID统一(MarketApp映射), 真实Skill执行, Agent工作流注入,
-      多轮上下文保持, BM25 KB检索, Function Calling循环增强
-"""
-from fastapi import APIRouter, Depends, HTTPException, Query
-from sqlalchemy.orm import Session
-from sqlalchemy import and_, desc, asc, func
-from pydantic import BaseModel, Field
-from typing import Optional
-import uuid
-from datetime import datetime, timezone, timedelta
-import re
-import json
-import asyncio
-import math
-import httpx
-import os
-import logging
-
-from ..database import get_db, SessionLocal
-from ..models import (
-    User, Conversation, Message, AIModel, ModelProvider, TokenUsageLog,
-    KnowledgeBase, KnowledgeDocument, KnowledgeDocumentChunk, EvolutionSkill,
-    EvolutionTrajectory, MarketApp, AppPurchase, AppInstallation,
-    gen_uuid
-)
-from ..auth import get_current_user
-from ..middleware.log_service import LogService
-from ..capability_registry import resolve_auto_model
-
-router = APIRouter()
-
-logger = logging.getLogger(__name__)
-
-
-def detect_media_intent(text: str) -> str | None:
-    """根据用户输入判断是否为 图像/视频/语音 生成请求，返回对应能力或 None。
-
-    用于 Auto 模式：用户未指定具体模型时，依据意图自动匹配平台各类能力的默认模型。
-    疑问句（含「吗/呢/怎么/能否」或结尾问号）不触发，避免「你能生成视频吗」被误判。
-    """
-    if not text:
-        return None
-    t = text.lower().strip()
-    # 疑问句不触发媒体生成
-    if t.endswith(("？", "?")) or any(q in t for q in ("吗", "呢", "怎么", "如何", "能否", "是否可以", "可以吗", "会不会", "why", "how")):
-        return None
-    video_kw = ("视频", "短片", "影片", "vide", "动画视频", "一段视频", "做个视频", "生成视频",
-                "视频生成", "来段视频", "拍个视频", "做段视频", "视频剪辑")
-    image_kw = ("图片", "图像", "插画", "海报", "照片", "头像", "壁纸", "漫画", "表情包", "logo", "图标",
-                "imag", "画一张", "画张", "画一只", "画只", "画个", "画幅",
-                "来张", "来一张", "来幅", "发张图", "发个图",
-                "生成图片", "生成一张", "生成一幅", "生成图", "图像生成",
-                "做一张", "做张图", "做张", "出张图", "出一张图",
-                "帮我画", "给我画", "请画", "绘制", "作画", "ai绘画", "ai画图",
-                "画个图", "画张图", "画一幅图", "一张图", "一幅画", "生成一幅")
-    audio_kw = ("语音", "配音", "朗读", "播报", "念一段", "有声", "tts", "音频", "生成语音",
-                "读一下", "念一下", "语音合成")
-    if any(k in t for k in video_kw):
-        return "video"
-    if any(k in t for k in image_kw):
-        return "image"
-    if any(k in t for k in audio_kw):
-        return "audio"
-    return None
-
-# CST timezone (UTC+8) for consistent time display
-CST_TZ = timezone(timedelta(hours=8))
-
-
-def cst_now():
-    """Get current time in CST timezone"""
-    return datetime.now(CST_TZ)
-
-
-def _html_escape(s: str) -> str:
-    return (str(s).replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")
-            .replace('"', "&quot;").replace("'", "&#39;"))
-
-
-async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str:
-    """图像/视频生成模型统一入口：按 capabilities 调用 AGNES 等媒体生成 API。
-    返回可直接在 AI 聊天中渲染的富文本(含 <img>/<video> 标签)。"""
-    provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
-    api_key = ai_model.api_key or (provider.api_key if provider else None)
-    if not api_key:
-        return "⚠️ 该媒体模型缺少 API Key，请在后台「大模型库」中配置。"
-
-    caps = [c.lower() for c in (ai_model.capabilities or [])]
-    headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}
-    timeout = ai_model.timeout or 120
-
-    # ============ 图像生成 ============
-    if "image" in caps:
-        api_url = ai_model.endpoint
-        if not api_url:
-            api_url = f"{provider.api_base_url.rstrip('/')}/v1/images/generations" if provider else ""
-        if not api_url:
-            return "⚠️ 该图像模型未配置 Endpoint。"
-        payload = {
-            "model": ai_model.name,
-            "prompt": prompt,
-            "size": "1024x768",
-            "extra_body": {"response_format": "url"},
-        }
-        try:
-            async with httpx.AsyncClient(timeout=timeout) as client:
-                resp = await client.post(api_url, headers=headers, json=payload)
-            if resp.status_code != 200:
-                return f"⚠️ 图像生成失败 (HTTP {resp.status_code}): {resp.text[:200]}"
-            data = resp.json()
-            items = data.get("data") or []
-            if not items:
-                return f"⚠️ 图像生成失败：响应无图片数据 {str(data)[:200]}"
-            item = items[0]
-            url = item.get("url") or item.get("b64_json")
-            if not url:
-                return f"⚠️ 图像生成失败：未返回图片地址 {str(data)[:200]}"
-            if not url.startswith("http"):
-                url = f"data:image/png;base64,{url}"
-            return (f'<img class="md-img" src="{url}" alt="生成图片"/>\n\n'
-                    f'**提示词：** {_html_escape(prompt)}')
-        except Exception as e:
-            return f"⚠️ 图像生成异常：{str(e)[:200]}"
-
-    # ============ 视频生成 (异步任务 + 轮询) ============
-    if "video" in caps:
-        api_url = ai_model.endpoint
-        if not api_url:
-            api_url = f"{provider.api_base_url.rstrip('/')}/v1/videos" if provider else ""
-        if not api_url:
-            return "⚠️ 该视频模型未配置 Endpoint。"
-        base = api_url.split("/v1")[0] if "/v1" in api_url else api_url
-        payload = {
-            "model": ai_model.name,
-            "prompt": prompt,
-            "height": 768,
-            "width": 1152,
-            "num_frames": 121,
-            "frame_rate": 24,
-        }
-        try:
-            async with httpx.AsyncClient(timeout=timeout) as client:
-                resp = await client.post(api_url, headers=headers, json=payload)
-                if resp.status_code != 200:
-                    return f"⚠️ 视频生成失败 (HTTP {resp.status_code}): {resp.text[:200]}"
-                data = resp.json()
-                video_id = data.get("video_id") or data.get("task_id") or data.get("id")
-                if not video_id:
-                    return f"⚠️ 视频生成失败：未返回任务ID {str(data)[:200]}"
-                # 轮询结果(最多约 5 分钟)
-                result_url = None
-                err_msg = None
-                for _ in range(60):
-                    await asyncio.sleep(5)
-                    r2 = await client.get(f"{base}/agnesapi?video_id={video_id}", headers=headers)
-                    if r2.status_code != 200:
-                        continue
-                    rd = r2.json()
-                    st = rd.get("status")
-                    if st == "completed":
-                        result_url = rd.get("url")
-                        break
-                    if st == "failed":
-                        err_msg = rd.get("error")
-                        break
-                if result_url:
-                    return (f'<video class="md-video" controls src="{result_url}"></video>\n\n'
-                            f'**提示词：** {_html_escape(prompt)}')
-                return f"⚠️ 视频生成失败：{err_msg or '生成超时，请稍后在对话中重试'}"
-        except Exception as e:
-            return f"⚠️ 视频生成异常：{str(e)[:200]}"
-
-    return "⚠️ 未知的媒体模型类型。"
-
-
-
-# ============================================================
-# Request Models
-# ============================================================
-
-class CreateConversationRequest(BaseModel):
-    title: str = Field(default="新对话")
-    model_id: Optional[str] = None
-
-
-class UpdateConversationRequest(BaseModel):
-    title: Optional[str] = None
-    status: Optional[str] = None  # active, archived
-    pinned: Optional[bool] = None
-    agent_ids: Optional[list] = None
-    skill_ids: Optional[list] = None
-    knowledge_base_ids: Optional[list] = None
-
-
-class SendMessageRequest(BaseModel):
-    content: str
-    model_id: Optional[str] = None
-    skill_ids: Optional[list] = None
-    agent_ids: Optional[list] = None
-    knowledge_base_ids: Optional[list] = None
-    use_tools: Optional[bool] = False
-
-
-# ============================================================
-# Skill Resolution: MarketApp → Skill Definition
-# ============================================================
-
-def _resolve_market_skills(db: Session, skill_ids: list, user_id: str) -> list:
-    """
-    Stage A.1: 统一Skill ID空间
-    前端传MarketApp.id, 需要映射为可用的技能定义。
-    返回: [{id, name, description, skill_content, category, func_name}]
-    """
-    if not skill_ids:
-        return []
-
-    # 1. 从MarketApp查询 (前端传的是MarketApp.id)
-    market_skills = db.query(MarketApp).filter(
-        MarketApp.id.in_(skill_ids),
-        MarketApp.type == "skill"
-    ).all()
-
-    results = []
-    for ms in market_skills:
-        skill_content = ms.skill_content or ""
-        # 尝试从skill_content解析SKILL.md格式
-        # SKILL.md格式: ---\nname: ...\ndescription: ...\n---\n内容...
-        parsed = _parse_skill_content(skill_content)
-
-        func_name = ms.name.replace(" ", "_").replace("-", "_").lower()
-        results.append({
-            "id": ms.id,
-            "market_app_id": ms.id,
-            "name": ms.name,
-            "description": ms.description or parsed.get("description", ""),
-            "skill_content": skill_content,
-            "category": parsed.get("category", "general"),
-            "skill_type": parsed.get("skill_type", "general"),
-            "func_name": func_name,
-            "parameters": parsed.get("parameters", {"input": {"type": "string", "description": "用户输入内容"}}),
-            "instructions": parsed.get("instructions", ""),
-        })
-
-    # 2. 如果有未匹配的ID, 尝试查EvolutionSkill (兼容旧的进化技能ID)
-    matched_ids = {r["id"] for r in results}
-    unmatched = [sid for sid in skill_ids if sid not in matched_ids]
-
-    if unmatched:
-        evo_skills = db.query(EvolutionSkill).filter(
-            EvolutionSkill.id.in_(unmatched),
-            EvolutionSkill.user_id == user_id
-        ).all()
-
-        for es in evo_skills:
-            # 读取SKILL.md内容
-            skill_md = ""
-            if es.skill_md_path and os.path.exists(es.skill_md_path):
-                try:
-                    with open(es.skill_md_path, "r", encoding="utf-8") as f:
-                        skill_md = f.read()
-                except Exception:
-                    skill_md = ""
-            elif es.description:
-                skill_md = es.description
-
-            parsed = _parse_skill_content(skill_md)
-            func_name = es.skill_name.replace(" ", "_").replace("-", "_").lower()
-
-            results.append({
-                "id": es.id,
-                "market_app_id": None,
-                "name": es.skill_name,
-                "description": es.description or parsed.get("description", ""),
-                "skill_content": skill_md,
-                "category": es.category or parsed.get("category", "general"),
-                "skill_type": parsed.get("skill_type", es.category or "general"),
-                "func_name": func_name,
-                "parameters": parsed.get("parameters", {"input": {"type": "string", "description": "用户输入内容"}}),
-                "instructions": parsed.get("instructions", ""),
-            })
-
-    return results
-
-
-def _resolve_market_agents(db: Session, agent_ids: list, user_id: str) -> list:
-    """
-    Stage B.1: 加载Agent配置
-    前端传MarketApp.id(agent类型), 需要解析AgentConfig。
-    返回: [{id, name, config, skills, knowledge_base_ids}]
-    """
-    if not agent_ids:
-        return []
-
-    agents = db.query(MarketApp).filter(
-        MarketApp.id.in_(agent_ids),
-        MarketApp.type == "agent"
-    ).all()
-
-    results = []
-    for ag in agents:
-        # AgentConfig在skill_content字段(JSON序列化)
-        agent_config = {}
-        try:
-            if ag.skill_content:
-                agent_config = json.loads(ag.skill_content)
-        except Exception:
-            agent_config = {"role": ag.description or "AI助手", "instructions": ag.description or ""}
-
-        # 提取配置字段
-        role_prompt = agent_config.get("role", ag.description or "AI助手")
-        instructions = agent_config.get("instructions", "")
-        agent_skill_ids = agent_config.get("skills", [])  # 关联的skill MarketApp IDs
-        agent_kb_ids = agent_config.get("knowledge_base_ids", [])  # 关联的知识库IDs
-        agent_model_id = agent_config.get("model_id", None)  # 指定使用的模型
-        tools = agent_config.get("tools", [])  # MCP server IDs(暂不用)
-
-        results.append({
-            "id": ag.id,
-            "name": ag.name,
-            "role_prompt": role_prompt,
-            "instructions": instructions,
-            "skill_ids": agent_skill_ids,
-            "knowledge_base_ids": agent_kb_ids,
-            "model_id": agent_model_id,
-            "tools": tools,
-            "description": ag.description or "",
-        })
-
-    return results
-
-
-def _parse_skill_content(content: str) -> dict:
-    """
-    解析SKILL.md格式内容, 提取结构化信息。
-    SKILL.md格式:
-    ---
-    name: 技能名
-    description: 描述
-    category: 分类
-    skill_type: 类型
-    parameters: JSON或简单定义
-    ---
-    技能指令/规则内容
-    """
-    result = {
-        "name": "",
-        "description": "",
-        "category": "general",
-        "skill_type": "general",
-        "parameters": {},
-        "instructions": "",
-    }
-
-    if not content:
-        return result
-
-    # 尝试解析YAML front matter
-    if content.strip().startswith("---"):
-        parts = content.split("---")
-        if len(parts) >= 3:
-            front_matter = parts[1].strip()
-            body = parts[2].strip() if len(parts) > 2 else ""
-
-            # 简单YAML解析 (不用pyyaml库)
-            for line in front_matter.split("\n"):
-                line = line.strip()
-                if ":" in line:
-                    key, val = line.split(":", 1)
-                    key = key.strip().lower()
-                    val = val.strip()
-                    if key == "name":
-                        result["name"] = val
-                    elif key == "description":
-                        result["description"] = val
-                    elif key == "category":
-                        result["category"] = val
-                    elif key == "skill_type":
-                        result["skill_type"] = val
-                    elif key == "parameters":
-                        try:
-                            result["parameters"] = json.loads(val)
-                        except Exception:
-                            # 简单格式: input:string,query:string
-                            params = {}
-                            for p in val.split(","):
-                                p = p.strip()
-                                if ":" in p:
-                                    pname, ptype = p.split(":", 1)
-                                    params[pname.strip()] = {"type": ptype.strip(), "description": pname.strip()}
-                            result["parameters"] = params
-
-            result["instructions"] = body
-        else:
-            result["instructions"] = content
-    else:
-        # 没有front matter, 全部当作instructions
-        result["instructions"] = content
-        # 尝试从第一行提取name
-        first_line = content.split("\n")[0].strip()
-        if first_line and len(first_line) < 50:
-            result["name"] = first_line
-
-    return result
-
-
-# ============================================================
-# BM25 Text Search (Stage C.2)
-# ============================================================
-
-def _bm25_search(query: str, documents: list, top_k: int = 5) -> list:
-    """
-    简易BM25检索 — 替代朴素的文档拼接。
-    对文档列表进行关键词匹配+TF-IDF排序, 返回最相关的top_k段。
-    """
-    if not documents or not query:
-        return documents[:top_k]
-
-    # 中文简单分词(按字符+标点)
-    def tokenize(text: str) -> list:
-        # 移除标点, 拆分为2-4字词组
-        text = re.sub(r'[^\w\u4e00-\u9fff]', ' ', text.lower())
-        tokens = []
-        words = text.split()
-        for w in words:
-            if len(w) >= 2:
-                tokens.append(w)
-            # 也生成2-gram
-            if len(w) >= 4:
-                for i in range(len(w) - 2):
-                    tokens.append(w[i:i+2])
-        return tokens
-
-    query_tokens = tokenize(query)
-    if not query_tokens:
-        return documents[:top_k]
-
-    scored_docs = []
-    avg_doc_len = sum(len(tokenize(d.get("content", ""))) for d in documents) / max(len(documents), 1)
-
-    # BM25参数
-    k1 = 1.5
-    b = 0.75
-
-    # 计算每个token的IDF
-    df = {}
-    for token in query_tokens:
-        for doc in documents:
-            if token in tokenize(doc.get("content", "")):
-                df[token] = df.get(token, 0) + 1
-
-    N = len(documents)
-    idf = {}
-    for token in query_tokens:
-        idf[token] = math.log((N - df.get(token, 0) + 0.5) / (df.get(token, 0) + 0.5) + 1)
-
-    # 计算每个文档的BM25得分
-    for doc in documents:
-        doc_tokens = tokenize(doc.get("content", ""))
-        doc_len = len(doc_tokens)
-        score = 0
-
-        for token in query_tokens:
-            tf = doc_tokens.count(token)
-            numerator = tf * (k1 + 1)
-            denominator = tf + k1 * (1 - b + b * doc_len / max(avg_doc_len, 1))
-            score += idf.get(token, 0) * numerator / denominator
-
-        scored_docs.append((score, doc))
-
-    # 按得分排序
-    scored_docs.sort(key=lambda x: x[0], reverse=True)
-    return [d for s, d in scored_docs[:top_k]]
-
-
-def _retrieve_kb_context(db: Session, kb_ids: list, query: str, max_chars: int = 3000) -> str:
-    """
-    Stage C: 增强版知识库检索 — 使用BM25排序替代朴素拼接
-    1. 获取每个KB的文档
-    2. 构建文档列表(content+title+metadata)
-    3. BM25检索最相关的段落
-    4. 拼接为上下文注入system prompt
-    """
-    if not kb_ids:
-        return ""
-
-    kbs = db.query(KnowledgeBase).filter(KnowledgeBase.id.in_(kb_ids)).all()
-    if not kbs:
-        return ""
-
-    # 收集所有文档段落
-    # Stage C.1: 优先使用 KnowledgeDocumentChunk 表的预分段数据
-    all_segments = []
-    for kb in kbs:
-        # 先查chunk表
-        chunks = db.query(KnowledgeDocumentChunk).filter(
-            KnowledgeDocumentChunk.knowledge_base_id == kb.id
-        ).order_by(KnowledgeDocumentChunk.chunk_index).limit(30).all()
-
-        if chunks:
-            # 使用预分段数据
-            for chunk in chunks:
-                if chunk.content and len(chunk.content.strip()) >= 20:
-                    all_segments.append({
-                        "content": "[{}/{}/段{}]: {}".format(
-                            kb.name, chunk.document_id[:8], chunk.chunk_index, chunk.content[:500]
-                        ),
-                        "raw_content": chunk.content,
-                        "kb_name": kb.name,
-                    })
-        else:
-            # 退化: 使用原始文档分段
-            docs = db.query(KnowledgeDocument).filter(
-                KnowledgeDocument.knowledge_base_id == kb.id,
-                KnowledgeDocument.status == "completed"
-            ).order_by(KnowledgeDocument.updated_at.desc()).limit(20).all()
-
-            for doc in docs:
-                content = doc.content or ""
-                if not content:
-                    continue
-
-                # 如果文档较长, 分段处理 (每段~800字)
-                if len(content) > 800:
-                    paragraphs = re.split(r'\n\n+', content)
-                    for pi, para in enumerate(paragraphs):
-                        if len(para.strip()) >= 20:
-                            all_segments.append({
-                                "content": "[{}/{}/段{}]: {}".format(
-                                    kb.name, doc.name, pi+1, para.strip()[:500]
-                                ),
-                                "raw_content": para.strip(),
-                                "kb_name": kb.name,
-                                "doc_name": doc.name,
-                            })
-                else:
-                    all_segments.append({
-                        "content": "[{}/{}]: {}".format(kb.name, doc.name, content[:500]),
-                        "raw_content": content,
-                        "kb_name": kb.name,
-                        "doc_name": doc.name,
-                    })
-
-    if not all_segments:
-        # 退化: 使用原始朴素拼接
-        kb_names = ", ".join([k.name for k in kbs])
-        return f"参考知识库: {kb_names}"
-
-    # BM25检索最相关的段落
-    top_segments = _bm25_search(query, all_segments, top_k=8)
-
-    # 拼接上下文
-    context_parts = [seg["content"] for seg in top_segments]
-    context = "\n".join(context_parts)
-
-    # 限制总长度
-    if len(context) > max_chars:
-        context = context[:max_chars] + "...(内容截断)"
-
-    return context
-
-
-# ============================================================
-# Dynamic System Prompt Builder (Stage B)
-# ============================================================
-
-def _build_system_prompt(
-    skills: list,
-    agents: list,
-    kb_context: str,
-    user_content: str,
-) -> tuple:
-    """
-    动态组装system prompt + function definitions。
-    返回: (system_prompt_str, function_definitions_list)
-    """
-    system_parts = ["你是AIFlow平台的AI助手，具备专业的知识储备和工具调用能力。"]
-
-    function_definitions = []
-
-    # Stage B.1: Agent角色注入
-    if agents:
-        for agent in agents:
-            role = agent.get("role_prompt", "")
-            instructions = agent.get("instructions", "")
-            if role:
-                system_parts.append(f"当前角色设定: {role}")
-            if instructions:
-                system_parts.append(f"角色行为指引: {instructions}")
-
-    # Stage A.2: Skill工具注入
-    if skills:
-        skill_descs = ", ".join([f"{s['name']}({s.get('category', 'general')})" for s in skills])
-        system_parts.append(f"可用技能工具: {skill_descs}")
-        system_parts.append("重要规则: 当用户请求涉及上述技能工具的处理范围时, 你必须优先通过function calling调用对应工具, 而不是直接回复文字。调用后将获得工具执行结果, 基于结果生成最终回复。")
-
-        # 为每个skill构建function definition
-        for s in skills:
-            # 从skill_content提取指令注入system prompt
-            instructions = s.get("instructions", "")
-            if instructions and len(instructions) < 800:
-                system_parts.append(f"技能「{s['name']}」的详细指令: {instructions}")
-
-            # 构建参数schema
-            params = s.get("parameters", {})
-            if isinstance(params, dict) and len(params) > 0:
-                properties = {}
-                required = []
-                for pname, pdef in params.items():
-                    if isinstance(pdef, dict):
-                        properties[pname] = {
-                            "type": pdef.get("type", "string"),
-                            "description": pdef.get("description", f"传递给{s['name']}的{pname}参数"),
-                        }
-                        if pdef.get("required", True):
-                            required.append(pname)
-                    else:
-                        properties[pname] = {"type": "string", "description": f"传递给{s['name']}的{pname}参数"}
-                        required.append(pname)
-
-                if not properties:
-                    properties = {"input": {"type": "string", "description": f"需要传递给{s['name']}的完整用户输入"}}
-                    required = ["input"]
-            else:
-                properties = {"input": {"type": "string", "description": f"需要传递给{s['name']}的完整用户输入"}}
-                required = ["input"]
-
-            desc = s.get("description", "")
-            category = s.get("category", "general")
-            func_desc = f"调用技能「{s['name']}」处理用户请求。使用场景: {desc}。当用户请求涉及{category}类任务时优先调用此工具。"
-
-            function_definitions.append({
-                "name": s["func_name"],
-                "description": func_desc,
-                "parameters": {
-                    "type": "object",
-                    "properties": properties,
-                    "required": required,
-                },
-            })
-
-    # KB context injection
-    if kb_context:
-        system_parts.append(f"参考以下知识库内容回答用户问题:\n{kb_context}")
-
-    system_prompt = "\n".join(system_parts)
-    return system_prompt, function_definitions
-
-
-# ============================================================
-# Skill Tool Execution (Stage A.2 rewritten)
-# ============================================================
-
-def _execute_skill_tool(
-    db: Session,
-    func_name: str,
-    args: dict,
-    skills: list,
-    user_content: str,
-) -> str:
-    """
-    Stage A.2: 真实Skill执行引擎 — 替代mock实现
-    1. 查找匹配的skill定义
-    2. 读取skill_content(SKILL.md)中的指令
-    3. 根据指令类型组装执行prompt
-    4. 返回结构化的执行结果给LLM
-    """
-    # 查找匹配的skill
-    matched_skill = None
-    for s in skills:
-        expected_name = s["func_name"]
-        if func_name == expected_name or func_name == f"skill_{s['id'][:8]}":
-            matched_skill = s
-            break
-
-    if not matched_skill:
-        # 退化: 返回通用结果
-        return f"工具 '{func_name}' 执行结果: 已处理输入内容, 但未找到对应的技能定义。请基于通用知识回复用户。"
-
-    skill_name = matched_skill["name"]
-    instructions = matched_skill.get("instructions", "")
-    input_text = args.get("input", user_content)
-
-    # 根据skill_content中的指令生成真实执行结果
-    # 核心思路: 将skill的instructions作为"执行逻辑",
-    # 把input_text传入, 让结果包含指令要求的处理格式
-
-    if instructions:
-        # 有明确指令的技能 — 按指令结构化输出
-        result_parts = [
-            f"技能「{skill_name}」执行结果:",
-            f"---",
-        ]
-
-        # 尝试从指令中提取输出格式要求
-        # 如果指令包含"输出"或"格式"关键字, 遵循其要求
-        output_format = ""
-        if "输出格式" in instructions or "output format" in instructions.lower():
-            # 提取格式描述
-            for line in instructions.split("\n"):
-                if "输出" in line or "格式" in line or "format" in line.lower():
-                    output_format += line + "\n"
-
-        # 根据skill_type增强处理
-        skill_type = matched_skill.get("skill_type", "general")
-
-        if skill_type == "writer" or "写作" in skill_name or "写" in instructions:
-            result_parts.append(f"基于技能「{skill_name}」的写作指令处理用户输入:")
-            result_parts.append(f"主题: {input_text[:200]}")
-            result_parts.append(f"根据技能规则: {instructions[:300]}")
-            result_parts.append(f"处理完成 — 请基于上述技能指令和输入内容, 生成符合技能要求的写作输出。")
-        elif skill_type == "analyzer" or "分析" in skill_name or "分析" in instructions:
-            result_parts.append(f"基于技能「{skill_name}」的分析指令处理用户输入:")
-            result_parts.append(f"分析对象: {input_text[:200]}")
-            result_parts.append(f"分析维度(按技能规则): {instructions[:300]}")
-            result_parts.append(f"关键发现: 识别到{max(1, len(input_text)//50)}个分析要点")
-            result_parts.append(f"处理完成 — 请基于上述技能指令, 对输入内容进行深入分析并给出结构化的分析报告。")
-        elif skill_type == "translator" or "翻译" in skill_name or "翻译" in instructions:
-            result_parts.append(f"基于技能「{skill_name}」的翻译指令处理用户输入:")
-            result_parts.append(f"原文: {input_text[:200]}")
-            result_parts.append(f"翻译规则: {instructions[:300]}")
-            result_parts.append(f"处理完成 — 请基于技能翻译规则, 将输入内容翻译为目标语言。")
-        elif skill_type == "summarizer" or "总结" in skill_name or "摘要" in instructions:
-            result_parts.append(f"基于技能「{skill_name}」的摘要指令处理用户输入:")
-            result_parts.append(f"原文长度: {len(input_text)} 字符")
-            result_parts.append(f"摘要规则: {instructions[:300]}")
-            result_parts.append(f"处理完成 — 请基于技能摘要规则, 对输入内容进行精准摘要。")
-        else:
-            # 通用技能 — 传递指令让LLM执行
-            result_parts.append(f"基于技能「{skill_name}」的指令处理用户请求:")
-            result_parts.append(f"用户输入: {input_text[:200]}")
-            result_parts.append(f"技能指令: {instructions[:500]}")
-            if output_format:
-                result_parts.append(f"期望输出格式: {output_format[:200]}")
-            result_parts.append(f"处理完成 — 请严格按照技能指令的规则和要求, 处理用户输入并生成最终回复。")
-
-        result = "\n".join(result_parts)
-    else:
-        # 无明确指令的技能 — 提供基本信息让LLM自行处理
-        desc = matched_skill.get("description", "")
-        result = (
-            f"技能「{skill_name}」已激活。\n"
-            f"技能描述: {desc}\n"
-            f"用户输入: {input_text[:200]}\n"
-            f"请基于技能「{skill_name}」的专业能力处理上述输入。"
-        )
-
-    # 更新使用计数(如果关联了MarketApp)
-    market_app_id = matched_skill.get("market_app_id")
-    if market_app_id:
-        try:
-            app = db.query(MarketApp).filter(MarketApp.id == market_app_id).first()
-            if app:
-                app.install_count = (app.install_count or 0) + 1
-        except Exception:
-            pass
-
-    return result
-
-
-# ============================================================
-# Conversation Context Persistence (Stage B.2)
-# ============================================================
-
-def _get_conversation_context(db: Session, conv_id: str) -> dict:
-    """
-    Stage B.2: 获取对话的持久化上下文
-    对话级别的agent/skill/kb选择保存在Conversation字段中。
-    """
-    conv = db.query(Conversation).filter(Conversation.id == conv_id).first()
-    if not conv:
-        return {}
-
-    # Conversation目前没有存储agent/skill/kb的JSON字段
-    # 我们从第一条消息中获取(如果存在), 或返回空
-    # 后续可通过数据库迁移添加Conversation.context字段
-
-    # 查询该对话中最早一条含有skill_ids/agent_ids/knowledge_ids的消息
-    first_tool_msg = db.query(Message).filter(
-        Message.conversation_id == conv_id,
-        Message.role == "user",
-    ).order_by(asc(Message.created_at)).first()
-
-    context = {}
-    if first_tool_msg:
-        if first_tool_msg.skill_ids:
-            context["skill_ids"] = first_tool_msg.skill_ids
-        if first_tool_msg.agent_ids:
-            context["agent_ids"] = first_tool_msg.agent_ids
-        if first_tool_msg.knowledge_ids:
-            context["knowledge_base_ids"] = first_tool_msg.knowledge_ids
-
-    return context
-
-
-# ============================================================
-# API Endpoints
-# ============================================================
-
-@router.get("/conversations")
-def list_conversations(
-    status: Optional[str] = Query(None),
-    search: Optional[str] = Query(None),
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    query = db.query(Conversation).filter(Conversation.user_id == current_user.id)
-
-    if status:
-        query = query.filter(Conversation.status == status)
-    else:
-        query = query.filter(Conversation.status != "archived")
-
-    if search:
-        query = query.filter(Conversation.title.ilike(f"%{search}%"))
-
-    query = query.order_by(desc(Conversation.pinned), desc(Conversation.updated_at))
-
-    conversations = query.all()
-    result = []
-    for conv in conversations:
-        result.append({
-            "id": conv.id,
-            "title": conv.title,
-            "model_id": conv.model_id,
-            "status": conv.status,
-            "pinned": conv.pinned,
-            "knowledge_graph_count": conv.knowledge_graph_count or 0,
-            "source": getattr(conv, "source", None),
-            "created_at": str(conv.created_at) if conv.created_at else None,
-            "updated_at": str(conv.updated_at) if conv.updated_at else None,
-        })
-
-    return {"code": 0, "data": result, "message": "ok"}
-
-
-@router.post("/conversations")
-def create_conversation(
-    req: CreateConversationRequest,
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    conv = Conversation(
-        id=gen_uuid(),
-        user_id=current_user.id,
-        title=req.title,
-        model_id=req.model_id,
-        status="active",
-        pinned=False,
-        knowledge_graph_count=0,
-        created_at=cst_now(),
-        updated_at=cst_now(),
-    )
-    db.add(conv)
-    db.commit()
-    db.refresh(conv)
-
-    return {
-        "code": 0,
-        "data": {
-            "id": conv.id,
-            "title": conv.title,
-            "model_id": conv.model_id,
-            "status": conv.status,
-            "pinned": conv.pinned,
-            "knowledge_graph_count": conv.knowledge_graph_count,
-            "created_at": str(conv.created_at) if conv.created_at else None,
-            "updated_at": str(conv.updated_at) if conv.updated_at else None,
-        },
-        "message": "创建成功",
-    }
-
-
-@router.put("/conversations/{conv_id}")
-def update_conversation(
-    conv_id: str,
-    req: UpdateConversationRequest,
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    conv = db.query(Conversation).filter(
-        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
-    ).first()
-    if not conv:
-        raise HTTPException(status_code=404, detail="对话不存在")
-
-    if req.title is not None:
-        conv.title = req.title
-    if req.status is not None:
-        conv.status = req.status
-    if req.pinned is not None:
-        conv.pinned = req.pinned
-    # Stage B.2: 保存agent/skill/kb选择到对话上下文
-    # (存储在第一条system消息的knowledge_ids/agent_ids/skill_ids字段中)
-    conv.updated_at = cst_now()
-
-    db.commit()
-    db.refresh(conv)
-
-    return {"code": 0, "data": {
-        "id": conv.id, "title": conv.title, "status": conv.status, "pinned": conv.pinned,
-        "updated_at": str(conv.updated_at),
-    }, "message": "更新成功"}
-
-
-@router.delete("/conversations/{conv_id}")
-def delete_conversation(
-    conv_id: str,
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    conv = db.query(Conversation).filter(
-        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
-    ).first()
-    if not conv:
-        raise HTTPException(status_code=404, detail="对话不存在")
-
-    conv.status = "archived"
-    conv.updated_at = cst_now()
-    db.commit()
-
-    return {"code": 0, "message": "删除成功"}
-
-
-@router.get("/conversations/{conv_id}/messages")
-def get_messages(
-    conv_id: str,
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    conv = db.query(Conversation).filter(
-        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
-    ).first()
-    if not conv:
-        raise HTTPException(status_code=404, detail="对话不存在")
-
-    messages = (
-        db.query(Message)
-        .filter(Message.conversation_id == conv_id)
-        .order_by(asc(Message.created_at), asc(Message.id))
-        .all()
-    )
-
-    result = []
-    for msg in messages:
-        result.append({
-            "id": msg.id,
-            "conversation_id": msg.conversation_id,
-            "role": msg.role,
-            "content": msg.content,
-            "token_used": msg.token_used or 0,
-            "model_id": msg.model_id,
-            "used_skills": msg.used_skills or [],
-            "created_at": str(msg.created_at) if msg.created_at else None,
-        })
-
-    return {"code": 0, "data": result, "message": "ok"}
-
-
-@router.post("/conversations/{conv_id}/messages")
-async def send_message(
-    conv_id: str,
-    req: SendMessageRequest,
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    conv = db.query(Conversation).filter(
-        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
-    ).first()
-    if not conv:
-        raise HTTPException(status_code=404, detail="对话不存在")
-
-    # Stage B.2: 多轮上下文保持 — 如果当前请求没有skill/agent/kb,
-    # 从对话历史中继承上一次的选择
-    effective_skill_ids = req.skill_ids or []
-    effective_agent_ids = req.agent_ids or []
-    effective_kb_ids = req.knowledge_base_ids or []
-    effective_use_tools = req.use_tools or False
-
-    if not effective_skill_ids and not effective_agent_ids and not effective_kb_ids:
-        # 从对话上下文继承
-        prev_context = _get_conversation_context(db, conv_id)
-        if prev_context:
-            effective_skill_ids = prev_context.get("skill_ids", [])
-            effective_agent_ids = prev_context.get("agent_ids", [])
-            effective_kb_ids = prev_context.get("knowledge_base_ids", [])
-            if effective_skill_ids or effective_agent_ids:
-                effective_use_tools = True
-
-    # 保存用户消息(含skill/agent/kb选择用于上下文保持)
-    user_msg = Message(
-        id=gen_uuid(),
-        conversation_id=conv_id,
-        role="user",
-        content=req.content,
-        token_used=len(req.content) // 2,
-        model_id=req.model_id or conv.model_id,
-        skill_ids=req.skill_ids if req.skill_ids else None,
-        agent_ids=req.agent_ids if req.agent_ids else None,
-        knowledge_ids=req.knowledge_base_ids if req.knowledge_base_ids else None,
-        created_at=cst_now(),
-    )
-    db.add(user_msg)
-
-    # ===== 核心聊天引擎 =====
-    user_id = current_user.id
-    model_id = req.model_id or conv.model_id
-    prompt_tokens = 0
-    completion_tokens = 0
-    total_tokens = 0
-    ai_content = ""
-
-    ai_model = None
-    skills = []  # 初始化避免UnboundLocalError
-    agents = []  # 初始化避免UnboundLocalError
-    # 解析模型：支持 Auto 模式（model_id='auto' 或未指定时，按能力自动匹配平台默认模型）
-    if model_id and model_id != "auto":
-        ai_model = db.query(AIModel).filter(AIModel.id == model_id).first()
-    if not ai_model:
-        # Auto / 未指定 → 按用户输入意图匹配对应能力的默认模型，否则回退文本对话
-        intent = detect_media_intent(req.content)
-        resolved = resolve_auto_model(intent, db) if intent else None
-        ai_model = resolved or resolve_auto_model("text", db)
-
-    if ai_model:
-        # ====== 图像/视频生成模型分支 ======
-        caps = [c.lower() for c in (ai_model.capabilities or [])] if ai_model else []
-        if ai_model and ("image" in caps or "video" in caps):
-            try:
-                ai_content = await _generate_media(db, ai_model, req.content, current_user.id)
-            except Exception as e:
-                ai_content = f"⚠️ 媒体生成失败：{str(e)[:200]}"
-        elif ai_model and ai_model.endpoint:
-            provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
-            api_key = ai_model.api_key or (provider.api_key if provider else None)
-            api_url = ai_model.endpoint
-
-            if not api_url and provider and provider.api_base_url:
-                api_url = f"{provider.api_base_url.rstrip('/')}/v1/chat/completions"
-
-            if api_key and api_url:
-                try:
-                    # Stage A+B: 解析Skills和Agents
-                    skills = _resolve_market_skills(db, effective_skill_ids, user_id)
-                    agents = _resolve_market_agents(db, effective_agent_ids, user_id)
-
-                    # Stage B.1: Agent可能自带skill_ids和kb_ids
-                    agent_extra_skill_ids = []
-                    agent_extra_kb_ids = []
-                    agent_model_override = None
-                    for agent in agents:
-                        if agent.get("skill_ids"):
-                            agent_extra_skill_ids.extend(agent["skill_ids"])
-                        if agent.get("knowledge_base_ids"):
-                            agent_extra_kb_ids.extend(agent["knowledge_base_ids"])
-                        if agent.get("model_id"):
-                            agent_model_override = agent["model_id"]
-
-                    # 合并Agent自带的skills和KBs
-                    all_skill_ids = list(set(effective_skill_ids + agent_extra_skill_ids))
-                    all_kb_ids = list(set(effective_kb_ids + agent_extra_kb_ids))
-
-                    if agent_extra_skill_ids or agent_extra_kb_ids:
-                        # 重新解析合并后的skills
-                        skills = _resolve_market_skills(db, all_skill_ids, user_id)
-                        effective_use_tools = True
-
-                    # Stage C: 增强版KB检索
-                    kb_context = _retrieve_kb_context(db, all_kb_ids, req.content)
-
-                    # Stage B: 动态组装system prompt + function definitions
-                    system_prompt, function_definitions = _build_system_prompt(
-                        skills=skills,
-                        agents=agents,
-                        kb_context=kb_context,
-                        user_content=req.content,
-                    )
-
-                    # 构建消息列表
-                    messages = [{"role": "system", "content": system_prompt}]
-
-                    # 加载对话历史(最多10条)
-                    history = db.query(Message).filter(
-                        Message.conversation_id == conv_id
-                    ).order_by(asc(Message.created_at)).limit(10).all()
-                    for h in history:
-                        if h.role == "user" or h.role == "assistant":
-                            messages.append({"role": h.role, "content": h.content})
-
-                    messages.append({"role": "user", "content": req.content})
-
-                    # Stage B.3: Function Calling循环(最多5轮)
-                    max_tool_rounds = 5
-                    for round_num in range(max_tool_rounds):
-                        request_body = {
-                            "model": ai_model.name,
-                            "messages": messages,
-                            "max_tokens": 2048,
-                        }
-                        if function_definitions and effective_use_tools:
-                            request_body["tools"] = [
-                                {"type": "function", "function": fd}
-                                for fd in function_definitions
-                            ]
-                            request_body["tool_choice"] = "auto"
-
-                        async with httpx.AsyncClient(timeout=ai_model.timeout or 120) as client:
-                            resp = await client.post(
-                                api_url,
-                                headers={
-                                    "Authorization": f"Bearer {api_key}",
-                                    "Content-Type": "application/json"
-                                },
-                                json=request_body
-                            )
-                            if resp.status_code != 200:
-                                if round_num == 0:
-                                    ai_content = f"AI模型调用失败 (HTTP {resp.status_code}), 请检查模型配置。"
-                                break
-
-                            data = resp.json()
-                            choice = data.get("choices", [{}])[0]
-                            message = choice.get("message", {})
-
-                            # 记录token使用量
-                            usage = data.get("usage", {})
-                            prompt_tokens += usage.get("prompt_tokens", 0)
-                            completion_tokens += usage.get("completion_tokens", 0)
-                            total_tokens += usage.get("total_tokens", 0)
-
-                            # 检查是否有tool_calls
-                            tool_calls = message.get("tool_calls", [])
-                            if tool_calls and effective_use_tools:
-                                # 将assistant消息(含tool_calls)加入上下文
-                                messages.append({
-                                    "role": "assistant",
-                                    "content": message.get("content") or "",
-                                    "tool_calls": tool_calls,
-                                })
-
-                                # Stage A.2: 真实Skill执行
-                                for tc in tool_calls:
-                                    func_name = tc.get("function", {}).get("name", "")
-                                    func_args_str = tc.get("function", {}).get("arguments", "{}")
-                                    try:
-                                        func_args = json.loads(func_args_str)
-                                    except Exception:
-                                        func_args = {"input": func_args_str}
-
-                                    tool_result = _execute_skill_tool(
-                                        db, func_name, func_args, skills, req.content
-                                    )
-                                    messages.append({
-                                        "role": "tool",
-                                        "tool_call_id": tc.get("id", f"call_{round_num}"),
-                                        "content": tool_result,
-                                    })
-
-                                # 继续下一轮让模型处理tool结果
-                                continue
-                            else:
-                                # 普通回复
-                                ai_content = message.get("content", "")
-                                if not ai_content and round_num > 0:
-                                    ai_content = "AI模型已处理工具调用结果, 但返回了空响应。"
-                                break
-
-                    if not ai_content:
-                        ai_content = "AI模型返回了空响应, 请稍后重试。"
-
-                except Exception as e:
-                    logger.warning(f"AI API调用异常: {e}")
-                    ai_content = "AI服务暂时不可用, 请稍后重试或检查模型API配置。"
-            else:
-                ai_content = "该模型缺少API密钥或端点配置, 请在后台管理-大模型库中配置API Key和Endpoint。"
-        else:
-            ai_content = "请选择一个已配置的AI模型进行对话。"
-    else:
-        ai_content = "请先在后台管理-大模型库配置默认对话模型（或聊天中选择模型），Auto 模式下将自动使用平台默认模型。"
-
-    ai_token_count = len(ai_content) // 2
-    ai_msg = Message(
-        id=gen_uuid(),
-        conversation_id=conv_id,
-        role="assistant",
-        content=ai_content,
-        token_used=ai_token_count,
-        model_id=ai_model.id if ai_model else (req.model_id or conv.model_id),
-        skill_ids=req.skill_ids if req.skill_ids else None,
-        agent_ids=req.agent_ids if req.agent_ids else None,
-        knowledge_ids=req.knowledge_base_ids if req.knowledge_base_ids else None,
-        created_at=cst_now(),
-    )
-    db.add(ai_msg)
-
-    # 更新used_skills
-    used_skills = []
-    if skills:
-        used_skills = [s["name"] for s in skills]
-        ai_msg.used_skills = used_skills
-
-    # 更新对话统计
-    conv.knowledge_graph_count = (conv.knowledge_graph_count or 0) + 1
-    conv.updated_at = cst_now()
-
-    # 记录Token使用
-    token_usage = TokenUsageLog(
-        id=gen_uuid(),
-        user_id=current_user.id,
-        model_id=ai_model.id if ai_model else (req.model_id or conv.model_id),
-        model_name=ai_model.name if ai_model else None,
-        conversation_id=conv_id,
-        prompt_tokens=prompt_tokens or len(req.content) // 2,
-        completion_tokens=completion_tokens or ai_token_count,
-        total_tokens=total_tokens or (len(req.content) // 2 + ai_token_count),
-        created_at=cst_now(),
-    )
-    db.add(token_usage)
-
-    # 进化轨迹记录
-    try:
-        existing = db.query(EvolutionTrajectory).filter(
-            EvolutionTrajectory.conversation_id == conv_id
-        ).first()
-        if not existing:
-            trajectory = EvolutionTrajectory(
-                id=gen_uuid(),
-                user_id=current_user.id,
-                conversation_id=conv_id,
-                status="discovered",
-                token_used=ai_token_count,
-                created_at=cst_now(),
-            )
-            db.add(trajectory)
-            db.flush()
-
-            try:
-                from app.services.evolution_service import EvolutionWatcherManager
-                watcher = EvolutionWatcherManager()
-                status = watcher.get_user_status(current_user.id, lambda: SessionLocal())
-                if not status.get("scheduler_running"):
-                    watcher.start_user_evolution(current_user.id, lambda: SessionLocal())
-                    logger.info(f"[Evolution] 自动启动用户 {current_user.id} 的进化Pipeline")
-            except Exception as evolution_err:
-                logger.warning(f"[Evolution] 自动启动Pipeline失败: {evolution_err}")
-    except Exception:
-        pass
-
-    db.commit()
-    db.refresh(user_msg)
-    db.refresh(ai_msg)
-
-    return {
-        "code": 0,
-        "data": {
-            "user_message": {
-                "id": user_msg.id, "role": user_msg.role, "content": user_msg.content,
-                "token_used": user_msg.token_used, "created_at": str(user_msg.created_at),
-            },
-            "assistant_message": {
-                "id": ai_msg.id, "role": ai_msg.role, "content": ai_msg.content,
-                "token_used": ai_msg.token_used, "created_at": str(ai_msg.created_at),
-                "used_skills": used_skills,
-            },
-        },
-        "message": "发送成功",
-    }
-
-
-@router.post("/conversations/{conv_id}/clear")
-def clear_conversation(
-    conv_id: str,
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    conv = db.query(Conversation).filter(
-        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
-    ).first()
-    if not conv:
-        raise HTTPException(status_code=404, detail="对话不存在")
-
-    db.query(Message).filter(
-        Message.conversation_id == conv_id
-    ).delete()
-
-    conv.knowledge_graph_count = 0
-    conv.updated_at = cst_now()
-    db.commit()
-
-    return {"code": 0, "message": "对话上下文已清空"}
-
-
-def _build_knowledge_graph(db: Session, kb: KnowledgeBase) -> dict:
-    """从知识库文档构建知识图谱节点和边，返回 graph_data。"""
-    documents = db.query(KnowledgeDocument).filter(
-        and_(
-            KnowledgeDocument.knowledge_base_id == kb.id,
-            KnowledgeDocument.status == "completed",
-        )
-    ).all()
-
-    graph_nodes = []
-    graph_edges = []
-
-    doc_entity_types = [
-        {"type": "concept", "label": "核心概念"},
-        {"type": "entity", "label": "关键实体"},
-        {"type": "attribute", "label": "属性特征"},
-        {"type": "relation", "label": "关联关系"},
-    ]
-
-    for idx, doc in enumerate(documents):
-        doc_node_id = f"doc-{doc.id[:8]}"
-        graph_nodes.append({
-            "id": doc_node_id, "label": doc.filename, "type": "document", "size": 18, "group": "document",
-        })
-
-        filename_base = doc.filename.rsplit(".", 1)[0]
-        words = filename_base.replace("_", " ").replace("-", " ").split()
-
-        for wi, word in enumerate(words):
-            if len(word) >= 2:
-                entity_type = doc_entity_types[wi % len(doc_entity_types)]
-                node_id = f"ent-{doc.id[:8]}-{wi}"
-                graph_nodes.append({
-                    "id": node_id, "label": word, "type": entity_type["type"],
-                    "size": 10 + (wi % 8), "group": entity_type["type"],
-                })
-                graph_edges.append({
-                    "source": doc_node_id, "target": node_id,
-                    "label": entity_type["label"], "weight": 0.5 + (0.5 / (wi + 1)),
-                })
-
-    for i in range(len(documents)):
-        for j in range(i + 1, len(documents)):
-            fi = documents[i].filename.lower().rsplit(".", 1)[0]
-            fj = documents[j].filename.lower().rsplit(".", 1)[0]
-            words_i = set(fi.replace("_", " ").replace("-", " ").split())
-            words_j = set(fj.replace("_", " ").replace("-", " ").split())
-            common = words_i & words_j
-            if common and len(common) > 0:
-                graph_edges.append({
-                    "source": f"doc-{documents[i].id[:8]}",
-                    "target": f"doc-{documents[j].id[:8]}",
-                    "label": f"共有关键词: {', '.join(list(common)[:3])}",
-                    "weight": min(0.9, 0.3 * len(common)),
-                })
-
-    graph_data = {
-        "nodes": graph_nodes,
-        "edges": graph_edges,
-        "stats": {
-            "total_nodes": len(graph_nodes), "total_edges": len(graph_edges),
-            "document_count": len(documents),
-        },
-    }
-
-    kb.knowledge_graph_count = len(graph_nodes)
-    kb.updated_at = cst_now()
-    return graph_data
-
-
-@router.post("/conversations/{conv_id}/to-knowledge-graph")
-def convert_to_knowledge_graph(
-    conv_id: str,
-    db: Session = Depends(get_db),
-    current_user: User = Depends(get_current_user),
-):
-    """将对话内容转换为知识图谱（幂等：同一对话仅生成一次知识库）"""
-    conv = db.query(Conversation).filter(
-        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
-    ).first()
-    if not conv:
-        raise HTTPException(status_code=404, detail="对话不存在")
-
-    messages = (
-        db.query(Message)
-        .filter(Message.conversation_id == conv_id)
-        .order_by(asc(Message.created_at), asc(Message.id))
-        .all()
-    )
-
-    if not messages:
-        raise HTTPException(status_code=400, detail="对话中没有消息, 无法生成知识图谱")
-
-    # 幂等：若该对话已生成过知识图谱且知识库仍存在，则直接刷新返回
-    if conv.knowledge_graph_base_id:
-        existing_kb = db.query(KnowledgeBase).filter(
-            KnowledgeBase.id == conv.knowledge_graph_base_id
-        ).first()
-        if existing_kb:
-            graph_data = _build_knowledge_graph(db, existing_kb)
-            conv.knowledge_graph_count = 1
-            conv.updated_at = cst_now()
-            db.commit()
-            return {
-                "code": 0,
-                "data": {
-                    "knowledge_base_id": existing_kb.id,
-                    "knowledge_base_name": existing_kb.name,
-                    "graph": graph_data,
-                    "knowledge_graph_count": 1,
-                    "already_exists": True,
-                },
-                "message": "知识图谱已存在, 已刷新展示",
-            }
-
-    kb_name = f"对话图谱-{conv.title}"
-    kb = KnowledgeBase(
-        id=gen_uuid(),
-        user_id=current_user.id,
-        name=kb_name,
-        description=f"从对话「{conv.title}」自动生成的知识图谱",
-        directory_path="/",
-        document_count=0,
-        created_at=cst_now(),
-        updated_at=cst_now(),
-    )
-    db.add(kb)
-    db.flush()
-
-    doc_count = 0
-    for idx, msg in enumerate(messages):
-        role_label = "用户" if msg.role == "user" else "AI"
-        preview = re.sub(r'[\\n\\r\\t]+', ' ', msg.content[:40]).strip()
-        filename = f"msg_{idx+1:03d}_{role_label}_{preview}.txt"
-
-        doc = KnowledgeDocument(
-            id=gen_uuid(),
-            knowledge_base_id=kb.id,
-            filename=filename,
-            file_type="txt",
-            file_size=len(msg.content.encode('utf-8')),
-            content=msg.content,
-            status="completed",
-            chunk_count=max(1, len(msg.content) // 500 + 1),
-            created_at=cst_now(),
-            updated_at=cst_now(),
-        )
-        db.add(doc)
-        doc_count += 1
-
-    kb.document_count = doc_count
-    db.flush()
-
-    # 构建知识图谱节点和边
-    graph_data = _build_knowledge_graph(db, kb)
-    conv.knowledge_graph_base_id = kb.id
-    conv.knowledge_graph_count = 1
-    conv.updated_at = cst_now()
-
-    db.commit()
-
-    LogService.log_operation(
-        user_id=current_user.id, username=current_user.username, action="convert_to_graph",
-        target="conversation", target_id=conv_id,
-        detail=f"将对话 {conv.title} 转换为知识图谱 {kb.name}",
-    )
-
-    return {
-        "code": 0,
-        "data": {
-            "knowledge_base_id": kb.id,
-            "knowledge_base_name": kb.name,
-            "graph": graph_data,
-            "knowledge_graph_count": 1,
-            "already_exists": False,
-        },
-        "message": f"知识图谱构建完成",
-    }
-
-
-@router.get("/models")
-def get_models(db: Session = Depends(get_db)):
-    models = db.query(AIModel).filter(AIModel.is_enabled == True).all()
-    result = []
-    for m in models:
-        provider = db.query(ModelProvider).filter(ModelProvider.id == m.provider_id).first()
-        result.append({
-            "id": m.id,
-            "name": m.name,
-            "display_name": getattr(m, "display_name", None) or m.name,
-            "provider_id": m.provider_id,
-            "provider_name": provider.name if provider else "",
-            "capabilities": m.capabilities or [],
-            "status": m.status,
-            "price_per_million_tokens": float(m.price_per_million_tokens or 0),
-            "is_enabled": m.is_enabled,
-            "is_default": bool(m.is_default) if hasattr(m, 'is_default') else False,
-            "total_calls": m.total_calls or 0,
-            "total_tokens": m.total_tokens or 0,
-            "created_at": str(m.created_at) if m.created_at else None,
-        })
-    result.sort(key=lambda x: (0 if x["is_default"] else 1, x["name"]))
-    return {"code": 0, "data": result, "message": "ok"}
-
+"""
+增强版 Chat API — Phase 2.5
+包含: Skill ID统一(MarketApp映射), 真实Skill执行, Agent工作流注入,
+      多轮上下文保持, BM25 KB检索, Function Calling循环增强
+"""
+from fastapi import APIRouter, Depends, HTTPException, Query
+from sqlalchemy.orm import Session
+from sqlalchemy import and_, desc, asc, func
+from pydantic import BaseModel, Field
+from typing import Optional
+import uuid
+from datetime import datetime, timezone, timedelta
+import re
+import json
+import asyncio
+import math
+import httpx
+import os
+import logging
+
+from ..database import get_db, SessionLocal
+from ..models import (
+    User, Conversation, Message, AIModel, ModelProvider, TokenUsageLog,
+    KnowledgeBase, KnowledgeDocument, KnowledgeDocumentChunk, EvolutionSkill,
+    EvolutionTrajectory, MarketApp, AppPurchase, AppInstallation,
+    gen_uuid
+)
+from ..auth import get_current_user
+from ..middleware.log_service import LogService
+from ..capability_registry import resolve_auto_model
+
+router = APIRouter()
+
+logger = logging.getLogger(__name__)
+
+
+def detect_media_intent(text: str) -> str | None:
+    """根据用户输入判断是否为 图像/视频/语音 生成请求，返回对应能力或 None。
+
+    用于 Auto 模式：用户未指定具体模型时，依据意图自动匹配平台各类能力的默认模型。
+    疑问句（含「吗/呢/怎么/能否」或结尾问号）不触发，避免「你能生成视频吗」被误判。
+    """
+    if not text:
+        return None
+    t = text.lower().strip()
+    # 疑问句不触发媒体生成
+    if t.endswith(("？", "?")) or any(q in t for q in ("吗", "呢", "怎么", "如何", "能否", "是否可以", "可以吗", "会不会", "why", "how")):
+        return None
+    video_kw = ("视频", "短片", "影片", "vide", "动画视频", "一段视频", "做个视频", "生成视频",
+                "视频生成", "来段视频", "拍个视频", "做段视频", "视频剪辑")
+    image_kw = ("图片", "图像", "插画", "海报", "照片", "头像", "壁纸", "漫画", "表情包", "logo", "图标",
+                "imag", "画一张", "画张", "画一只", "画只", "画个", "画幅",
+                "来张", "来一张", "来幅", "发张图", "发个图",
+                "生成图片", "生成一张", "生成一幅", "生成图", "图像生成",
+                "做一张", "做张图", "做张", "出张图", "出一张图",
+                "帮我画", "给我画", "请画", "绘制", "作画", "ai绘画", "ai画图",
+                "画个图", "画张图", "画一幅图", "一张图", "一幅画", "生成一幅")
+    audio_kw = ("语音", "配音", "朗读", "播报", "念一段", "有声", "tts", "音频", "生成语音",
+                "读一下", "念一下", "语音合成")
+    if any(k in t for k in video_kw):
+        return "video"
+    if any(k in t for k in image_kw):
+        return "image"
+    if any(k in t for k in audio_kw):
+        return "audio"
+    return None
+
+# CST timezone (UTC+8) for consistent time display
+CST_TZ = timezone(timedelta(hours=8))
+
+
+def cst_now():
+    """Get current time in CST timezone"""
+    return datetime.now(CST_TZ)
+
+
+def _html_escape(s: str) -> str:
+    return (str(s).replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")
+            .replace('"', "&quot;").replace("'", "&#39;"))
+
+
+def _outbound_proxy() -> str:
+    """出网代理地址。容器内直连 apihub.agnes-ai.com 会被对端重置
+    (httpx RemoteProtocolError: Server disconnected without sending a response)，
+    必须经 172.18.0.1:7890 智能路由代理(其对国内 AGNES 域名走直连)。"""
+    import os
+    return os.environ.get("HTTPS_PROXY") or os.environ.get("HTTP_PROXY") or "http://172.18.0.1:7890"
+
+
+async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str:
+    """图像/视频生成模型统一入口：按 capabilities 调用 AGNES 等媒体生成 API。
+    返回可直接在 AI 聊天中渲染的富文本(含 <img>/<video> 标签)。"""
+    provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
+    api_key = ai_model.api_key or (provider.api_key if provider else None)
+    if not api_key:
+        return "⚠️ 该媒体模型缺少 API Key，请在后台「大模型库」中配置。"
+
+    caps = [c.lower() for c in (ai_model.capabilities or [])]
+    headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}
+    timeout = ai_model.timeout or 120
+
+    # ============ 图像生成 ============
+    if "image" in caps:
+        api_url = ai_model.endpoint
+        if not api_url:
+            api_url = f"{provider.api_base_url.rstrip('/')}/v1/images/generations" if provider else ""
+        if not api_url:
+            return "⚠️ 该图像模型未配置 Endpoint。"
+        payload = {
+            "model": ai_model.name,
+            "prompt": prompt,
+            "size": "1024x768",
+            "extra_body": {"response_format": "url"},
+        }
+        try:
+            async with httpx.AsyncClient(timeout=timeout, proxy=_outbound_proxy()) as client:
+                resp = await client.post(api_url, headers=headers, json=payload)
+            if resp.status_code != 200:
+                return f"⚠️ 图像生成失败 (HTTP {resp.status_code}): {resp.text[:200]}"
+            data = resp.json()
+            items = data.get("data") or []
+            if not items:
+                return f"⚠️ 图像生成失败：响应无图片数据 {str(data)[:200]}"
+            item = items[0]
+            url = item.get("url") or item.get("b64_json")
+            if not url:
+                return f"⚠️ 图像生成失败：未返回图片地址 {str(data)[:200]}"
+            if not url.startswith("http"):
+                url = f"data:image/png;base64,{url}"
+            return (f'<img class="md-img" src="{url}" alt="生成图片"/>\n\n'
+                    f'**提示词：** {_html_escape(prompt)}')
+        except Exception as e:
+            return f"⚠️ 图像生成异常：{str(e)[:200]}"
+
+    # ============ 视频生成 (异步任务 + 轮询) ============
+    if "video" in caps:
+        api_url = ai_model.endpoint
+        if not api_url:
+            api_url = f"{provider.api_base_url.rstrip('/')}/v1/videos" if provider else ""
+        if not api_url:
+            return "⚠️ 该视频模型未配置 Endpoint。"
+        base = api_url.split("/v1")[0] if "/v1" in api_url else api_url
+        payload = {
+            "model": ai_model.name,
+            "prompt": prompt,
+            "height": 768,
+            "width": 1152,
+            "num_frames": 121,
+            "frame_rate": 24,
+        }
+        try:
+            async with httpx.AsyncClient(timeout=timeout, proxy=_outbound_proxy()) as client:
+                resp = await client.post(api_url, headers=headers, json=payload)
+                if resp.status_code != 200:
+                    return f"⚠️ 视频生成失败 (HTTP {resp.status_code}): {resp.text[:200]}"
+                data = resp.json()
+                video_id = data.get("video_id") or data.get("task_id") or data.get("id")
+                if not video_id:
+                    return f"⚠️ 视频生成失败：未返回任务ID {str(data)[:200]}"
+                # 轮询结果(最多约 5 分钟)
+                result_url = None
+                err_msg = None
+                for _ in range(60):
+                    await asyncio.sleep(5)
+                    r2 = await client.get(f"{base}/agnesapi?video_id={video_id}", headers=headers)
+                    if r2.status_code != 200:
+                        continue
+                    rd = r2.json()
+                    st = rd.get("status")
+                    if st == "completed":
+                        result_url = rd.get("url")
+                        break
+                    if st == "failed":
+                        err_msg = rd.get("error")
+                        break
+                if result_url:
+                    return (f'<video class="md-video" controls src="{result_url}"></video>\n\n'
+                            f'**提示词：** {_html_escape(prompt)}')
+                return f"⚠️ 视频生成失败：{err_msg or '生成超时，请稍后在对话中重试'}"
+        except Exception as e:
+            return f"⚠️ 视频生成异常：{str(e)[:200]}"
+
+    return "⚠️ 未知的媒体模型类型。"
+
+
+
+# ============================================================
+# Request Models
+# ============================================================
+
+class CreateConversationRequest(BaseModel):
+    title: str = Field(default="新对话")
+    model_id: Optional[str] = None
+
+
+class UpdateConversationRequest(BaseModel):
+    title: Optional[str] = None
+    status: Optional[str] = None  # active, archived
+    pinned: Optional[bool] = None
+    agent_ids: Optional[list] = None
+    skill_ids: Optional[list] = None
+    knowledge_base_ids: Optional[list] = None
+
+
+class SendMessageRequest(BaseModel):
+    content: str
+    model_id: Optional[str] = None
+    skill_ids: Optional[list] = None
+    agent_ids: Optional[list] = None
+    knowledge_base_ids: Optional[list] = None
+    use_tools: Optional[bool] = False
+
+
+# ============================================================
+# Skill Resolution: MarketApp → Skill Definition
+# ============================================================
+
+def _resolve_market_skills(db: Session, skill_ids: list, user_id: str) -> list:
+    """
+    Stage A.1: 统一Skill ID空间
+    前端传MarketApp.id, 需要映射为可用的技能定义。
+    返回: [{id, name, description, skill_content, category, func_name}]
+    """
+    if not skill_ids:
+        return []
+
+    # 1. 从MarketApp查询 (前端传的是MarketApp.id)
+    market_skills = db.query(MarketApp).filter(
+        MarketApp.id.in_(skill_ids),
+        MarketApp.type == "skill"
+    ).all()
+
+    results = []
+    for ms in market_skills:
+        skill_content = ms.skill_content or ""
+        # 尝试从skill_content解析SKILL.md格式
+        # SKILL.md格式: ---\nname: ...\ndescription: ...\n---\n内容...
+        parsed = _parse_skill_content(skill_content)
+
+        func_name = ms.name.replace(" ", "_").replace("-", "_").lower()
+        results.append({
+            "id": ms.id,
+            "market_app_id": ms.id,
+            "name": ms.name,
+            "description": ms.description or parsed.get("description", ""),
+            "skill_content": skill_content,
+            "category": parsed.get("category", "general"),
+            "skill_type": parsed.get("skill_type", "general"),
+            "func_name": func_name,
+            "parameters": parsed.get("parameters", {"input": {"type": "string", "description": "用户输入内容"}}),
+            "instructions": parsed.get("instructions", ""),
+        })
+
+    # 2. 如果有未匹配的ID, 尝试查EvolutionSkill (兼容旧的进化技能ID)
+    matched_ids = {r["id"] for r in results}
+    unmatched = [sid for sid in skill_ids if sid not in matched_ids]
+
+    if unmatched:
+        evo_skills = db.query(EvolutionSkill).filter(
+            EvolutionSkill.id.in_(unmatched),
+            EvolutionSkill.user_id == user_id
+        ).all()
+
+        for es in evo_skills:
+            # 读取SKILL.md内容
+            skill_md = ""
+            if es.skill_md_path and os.path.exists(es.skill_md_path):
+                try:
+                    with open(es.skill_md_path, "r", encoding="utf-8") as f:
+                        skill_md = f.read()
+                except Exception:
+                    skill_md = ""
+            elif es.description:
+                skill_md = es.description
+
+            parsed = _parse_skill_content(skill_md)
+            func_name = es.skill_name.replace(" ", "_").replace("-", "_").lower()
+
+            results.append({
+                "id": es.id,
+                "market_app_id": None,
+                "name": es.skill_name,
+                "description": es.description or parsed.get("description", ""),
+                "skill_content": skill_md,
+                "category": es.category or parsed.get("category", "general"),
+                "skill_type": parsed.get("skill_type", es.category or "general"),
+                "func_name": func_name,
+                "parameters": parsed.get("parameters", {"input": {"type": "string", "description": "用户输入内容"}}),
+                "instructions": parsed.get("instructions", ""),
+            })
+
+    return results
+
+
+def _resolve_market_agents(db: Session, agent_ids: list, user_id: str) -> list:
+    """
+    Stage B.1: 加载Agent配置
+    前端传MarketApp.id(agent类型), 需要解析AgentConfig。
+    返回: [{id, name, config, skills, knowledge_base_ids}]
+    """
+    if not agent_ids:
+        return []
+
+    agents = db.query(MarketApp).filter(
+        MarketApp.id.in_(agent_ids),
+        MarketApp.type == "agent"
+    ).all()
+
+    results = []
+    for ag in agents:
+        # AgentConfig在skill_content字段(JSON序列化)
+        agent_config = {}
+        try:
+            if ag.skill_content:
+                agent_config = json.loads(ag.skill_content)
+        except Exception:
+            agent_config = {"role": ag.description or "AI助手", "instructions": ag.description or ""}
+
+        # 提取配置字段
+        role_prompt = agent_config.get("role", ag.description or "AI助手")
+        instructions = agent_config.get("instructions", "")
+        agent_skill_ids = agent_config.get("skills", [])  # 关联的skill MarketApp IDs
+        agent_kb_ids = agent_config.get("knowledge_base_ids", [])  # 关联的知识库IDs
+        agent_model_id = agent_config.get("model_id", None)  # 指定使用的模型
+        tools = agent_config.get("tools", [])  # MCP server IDs(暂不用)
+
+        results.append({
+            "id": ag.id,
+            "name": ag.name,
+            "role_prompt": role_prompt,
+            "instructions": instructions,
+            "skill_ids": agent_skill_ids,
+            "knowledge_base_ids": agent_kb_ids,
+            "model_id": agent_model_id,
+            "tools": tools,
+            "description": ag.description or "",
+        })
+
+    return results
+
+
+def _parse_skill_content(content: str) -> dict:
+    """
+    解析SKILL.md格式内容, 提取结构化信息。
+    SKILL.md格式:
+    ---
+    name: 技能名
+    description: 描述
+    category: 分类
+    skill_type: 类型
+    parameters: JSON或简单定义
+    ---
+    技能指令/规则内容
+    """
+    result = {
+        "name": "",
+        "description": "",
+        "category": "general",
+        "skill_type": "general",
+        "parameters": {},
+        "instructions": "",
+    }
+
+    if not content:
+        return result
+
+    # 尝试解析YAML front matter
+    if content.strip().startswith("---"):
+        parts = content.split("---")
+        if len(parts) >= 3:
+            front_matter = parts[1].strip()
+            body = parts[2].strip() if len(parts) > 2 else ""
+
+            # 简单YAML解析 (不用pyyaml库)
+            for line in front_matter.split("\n"):
+                line = line.strip()
+                if ":" in line:
+                    key, val = line.split(":", 1)
+                    key = key.strip().lower()
+                    val = val.strip()
+                    if key == "name":
+                        result["name"] = val
+                    elif key == "description":
+                        result["description"] = val
+                    elif key == "category":
+                        result["category"] = val
+                    elif key == "skill_type":
+                        result["skill_type"] = val
+                    elif key == "parameters":
+                        try:
+                            result["parameters"] = json.loads(val)
+                        except Exception:
+                            # 简单格式: input:string,query:string
+                            params = {}
+                            for p in val.split(","):
+                                p = p.strip()
+                                if ":" in p:
+                                    pname, ptype = p.split(":", 1)
+                                    params[pname.strip()] = {"type": ptype.strip(), "description": pname.strip()}
+                            result["parameters"] = params
+
+            result["instructions"] = body
+        else:
+            result["instructions"] = content
+    else:
+        # 没有front matter, 全部当作instructions
+        result["instructions"] = content
+        # 尝试从第一行提取name
+        first_line = content.split("\n")[0].strip()
+        if first_line and len(first_line) < 50:
+            result["name"] = first_line
+
+    return result
+
+
+# ============================================================
+# BM25 Text Search (Stage C.2)
+# ============================================================
+
+def _bm25_search(query: str, documents: list, top_k: int = 5) -> list:
+    """
+    简易BM25检索 — 替代朴素的文档拼接。
+    对文档列表进行关键词匹配+TF-IDF排序, 返回最相关的top_k段。
+    """
+    if not documents or not query:
+        return documents[:top_k]
+
+    # 中文简单分词(按字符+标点)
+    def tokenize(text: str) -> list:
+        # 移除标点, 拆分为2-4字词组
+        text = re.sub(r'[^\w\u4e00-\u9fff]', ' ', text.lower())
+        tokens = []
+        words = text.split()
+        for w in words:
+            if len(w) >= 2:
+                tokens.append(w)
+            # 也生成2-gram
+            if len(w) >= 4:
+                for i in range(len(w) - 2):
+                    tokens.append(w[i:i+2])
+        return tokens
+
+    query_tokens = tokenize(query)
+    if not query_tokens:
+        return documents[:top_k]
+
+    scored_docs = []
+    avg_doc_len = sum(len(tokenize(d.get("content", ""))) for d in documents) / max(len(documents), 1)
+
+    # BM25参数
+    k1 = 1.5
+    b = 0.75
+
+    # 计算每个token的IDF
+    df = {}
+    for token in query_tokens:
+        for doc in documents:
+            if token in tokenize(doc.get("content", "")):
+                df[token] = df.get(token, 0) + 1
+
+    N = len(documents)
+    idf = {}
+    for token in query_tokens:
+        idf[token] = math.log((N - df.get(token, 0) + 0.5) / (df.get(token, 0) + 0.5) + 1)
+
+    # 计算每个文档的BM25得分
+    for doc in documents:
+        doc_tokens = tokenize(doc.get("content", ""))
+        doc_len = len(doc_tokens)
+        score = 0
+
+        for token in query_tokens:
+            tf = doc_tokens.count(token)
+            numerator = tf * (k1 + 1)
+            denominator = tf + k1 * (1 - b + b * doc_len / max(avg_doc_len, 1))
+            score += idf.get(token, 0) * numerator / denominator
+
+        scored_docs.append((score, doc))
+
+    # 按得分排序
+    scored_docs.sort(key=lambda x: x[0], reverse=True)
+    return [d for s, d in scored_docs[:top_k]]
+
+
+def _retrieve_kb_context(db: Session, kb_ids: list, query: str, max_chars: int = 3000) -> str:
+    """
+    Stage C: 增强版知识库检索 — 使用BM25排序替代朴素拼接
+    1. 获取每个KB的文档
+    2. 构建文档列表(content+title+metadata)
+    3. BM25检索最相关的段落
+    4. 拼接为上下文注入system prompt
+    """
+    if not kb_ids:
+        return ""
+
+    kbs = db.query(KnowledgeBase).filter(KnowledgeBase.id.in_(kb_ids)).all()
+    if not kbs:
+        return ""
+
+    # 收集所有文档段落
+    # Stage C.1: 优先使用 KnowledgeDocumentChunk 表的预分段数据
+    all_segments = []
+    for kb in kbs:
+        # 先查chunk表
+        chunks = db.query(KnowledgeDocumentChunk).filter(
+            KnowledgeDocumentChunk.knowledge_base_id == kb.id
+        ).order_by(KnowledgeDocumentChunk.chunk_index).limit(30).all()
+
+        if chunks:
+            # 使用预分段数据
+            for chunk in chunks:
+                if chunk.content and len(chunk.content.strip()) >= 20:
+                    all_segments.append({
+                        "content": "[{}/{}/段{}]: {}".format(
+                            kb.name, chunk.document_id[:8], chunk.chunk_index, chunk.content[:500]
+                        ),
+                        "raw_content": chunk.content,
+                        "kb_name": kb.name,
+                    })
+        else:
+            # 退化: 使用原始文档分段
+            docs = db.query(KnowledgeDocument).filter(
+                KnowledgeDocument.knowledge_base_id == kb.id,
+                KnowledgeDocument.status == "completed"
+            ).order_by(KnowledgeDocument.updated_at.desc()).limit(20).all()
+
+            for doc in docs:
+                content = doc.content or ""
+                if not content:
+                    continue
+
+                # 如果文档较长, 分段处理 (每段~800字)
+                if len(content) > 800:
+                    paragraphs = re.split(r'\n\n+', content)
+                    for pi, para in enumerate(paragraphs):
+                        if len(para.strip()) >= 20:
+                            all_segments.append({
+                                "content": "[{}/{}/段{}]: {}".format(
+                                    kb.name, doc.name, pi+1, para.strip()[:500]
+                                ),
+                                "raw_content": para.strip(),
+                                "kb_name": kb.name,
+                                "doc_name": doc.name,
+                            })
+                else:
+                    all_segments.append({
+                        "content": "[{}/{}]: {}".format(kb.name, doc.name, content[:500]),
+                        "raw_content": content,
+                        "kb_name": kb.name,
+                        "doc_name": doc.name,
+                    })
+
+    if not all_segments:
+        # 退化: 使用原始朴素拼接
+        kb_names = ", ".join([k.name for k in kbs])
+        return f"参考知识库: {kb_names}"
+
+    # BM25检索最相关的段落
+    top_segments = _bm25_search(query, all_segments, top_k=8)
+
+    # 拼接上下文
+    context_parts = [seg["content"] for seg in top_segments]
+    context = "\n".join(context_parts)
+
+    # 限制总长度
+    if len(context) > max_chars:
+        context = context[:max_chars] + "...(内容截断)"
+
+    return context
+
+
+# ============================================================
+# Dynamic System Prompt Builder (Stage B)
+# ============================================================
+
+def _build_system_prompt(
+    skills: list,
+    agents: list,
+    kb_context: str,
+    user_content: str,
+) -> tuple:
+    """
+    动态组装system prompt + function definitions。
+    返回: (system_prompt_str, function_definitions_list)
+    """
+    system_parts = ["你是AIFlow平台的AI助手，具备专业的知识储备和工具调用能力。"]
+
+    function_definitions = []
+
+    # Stage B.1: Agent角色注入
+    if agents:
+        for agent in agents:
+            role = agent.get("role_prompt", "")
+            instructions = agent.get("instructions", "")
+            if role:
+                system_parts.append(f"当前角色设定: {role}")
+            if instructions:
+                system_parts.append(f"角色行为指引: {instructions}")
+
+    # Stage A.2: Skill工具注入
+    if skills:
+        skill_descs = ", ".join([f"{s['name']}({s.get('category', 'general')})" for s in skills])
+        system_parts.append(f"可用技能工具: {skill_descs}")
+        system_parts.append("重要规则: 当用户请求涉及上述技能工具的处理范围时, 你必须优先通过function calling调用对应工具, 而不是直接回复文字。调用后将获得工具执行结果, 基于结果生成最终回复。")
+
+        # 为每个skill构建function definition
+        for s in skills:
+            # 从skill_content提取指令注入system prompt
+            instructions = s.get("instructions", "")
+            if instructions and len(instructions) < 800:
+                system_parts.append(f"技能「{s['name']}」的详细指令: {instructions}")
+
+            # 构建参数schema
+            params = s.get("parameters", {})
+            if isinstance(params, dict) and len(params) > 0:
+                properties = {}
+                required = []
+                for pname, pdef in params.items():
+                    if isinstance(pdef, dict):
+                        properties[pname] = {
+                            "type": pdef.get("type", "string"),
+                            "description": pdef.get("description", f"传递给{s['name']}的{pname}参数"),
+                        }
+                        if pdef.get("required", True):
+                            required.append(pname)
+                    else:
+                        properties[pname] = {"type": "string", "description": f"传递给{s['name']}的{pname}参数"}
+                        required.append(pname)
+
+                if not properties:
+                    properties = {"input": {"type": "string", "description": f"需要传递给{s['name']}的完整用户输入"}}
+                    required = ["input"]
+            else:
+                properties = {"input": {"type": "string", "description": f"需要传递给{s['name']}的完整用户输入"}}
+                required = ["input"]
+
+            desc = s.get("description", "")
+            category = s.get("category", "general")
+            func_desc = f"调用技能「{s['name']}」处理用户请求。使用场景: {desc}。当用户请求涉及{category}类任务时优先调用此工具。"
+
+            function_definitions.append({
+                "name": s["func_name"],
+                "description": func_desc,
+                "parameters": {
+                    "type": "object",
+                    "properties": properties,
+                    "required": required,
+                },
+            })
+
+    # KB context injection
+    if kb_context:
+        system_parts.append(f"参考以下知识库内容回答用户问题:\n{kb_context}")
+
+    system_prompt = "\n".join(system_parts)
+    return system_prompt, function_definitions
+
+
+# ============================================================
+# Skill Tool Execution (Stage A.2 rewritten)
+# ============================================================
+
+def _execute_skill_tool(
+    db: Session,
+    func_name: str,
+    args: dict,
+    skills: list,
+    user_content: str,
+) -> str:
+    """
+    Stage A.2: 真实Skill执行引擎 — 替代mock实现
+    1. 查找匹配的skill定义
+    2. 读取skill_content(SKILL.md)中的指令
+    3. 根据指令类型组装执行prompt
+    4. 返回结构化的执行结果给LLM
+    """
+    # 查找匹配的skill
+    matched_skill = None
+    for s in skills:
+        expected_name = s["func_name"]
+        if func_name == expected_name or func_name == f"skill_{s['id'][:8]}":
+            matched_skill = s
+            break
+
+    if not matched_skill:
+        # 退化: 返回通用结果
+        return f"工具 '{func_name}' 执行结果: 已处理输入内容, 但未找到对应的技能定义。请基于通用知识回复用户。"
+
+    skill_name = matched_skill["name"]
+    instructions = matched_skill.get("instructions", "")
+    input_text = args.get("input", user_content)
+
+    # 根据skill_content中的指令生成真实执行结果
+    # 核心思路: 将skill的instructions作为"执行逻辑",
+    # 把input_text传入, 让结果包含指令要求的处理格式
+
+    if instructions:
+        # 有明确指令的技能 — 按指令结构化输出
+        result_parts = [
+            f"技能「{skill_name}」执行结果:",
+            f"---",
+        ]
+
+        # 尝试从指令中提取输出格式要求
+        # 如果指令包含"输出"或"格式"关键字, 遵循其要求
+        output_format = ""
+        if "输出格式" in instructions or "output format" in instructions.lower():
+            # 提取格式描述
+            for line in instructions.split("\n"):
+                if "输出" in line or "格式" in line or "format" in line.lower():
+                    output_format += line + "\n"
+
+        # 根据skill_type增强处理
+        skill_type = matched_skill.get("skill_type", "general")
+
+        if skill_type == "writer" or "写作" in skill_name or "写" in instructions:
+            result_parts.append(f"基于技能「{skill_name}」的写作指令处理用户输入:")
+            result_parts.append(f"主题: {input_text[:200]}")
+            result_parts.append(f"根据技能规则: {instructions[:300]}")
+            result_parts.append(f"处理完成 — 请基于上述技能指令和输入内容, 生成符合技能要求的写作输出。")
+        elif skill_type == "analyzer" or "分析" in skill_name or "分析" in instructions:
+            result_parts.append(f"基于技能「{skill_name}」的分析指令处理用户输入:")
+            result_parts.append(f"分析对象: {input_text[:200]}")
+            result_parts.append(f"分析维度(按技能规则): {instructions[:300]}")
+            result_parts.append(f"关键发现: 识别到{max(1, len(input_text)//50)}个分析要点")
+            result_parts.append(f"处理完成 — 请基于上述技能指令, 对输入内容进行深入分析并给出结构化的分析报告。")
+        elif skill_type == "translator" or "翻译" in skill_name or "翻译" in instructions:
+            result_parts.append(f"基于技能「{skill_name}」的翻译指令处理用户输入:")
+            result_parts.append(f"原文: {input_text[:200]}")
+            result_parts.append(f"翻译规则: {instructions[:300]}")
+            result_parts.append(f"处理完成 — 请基于技能翻译规则, 将输入内容翻译为目标语言。")
+        elif skill_type == "summarizer" or "总结" in skill_name or "摘要" in instructions:
+            result_parts.append(f"基于技能「{skill_name}」的摘要指令处理用户输入:")
+            result_parts.append(f"原文长度: {len(input_text)} 字符")
+            result_parts.append(f"摘要规则: {instructions[:300]}")
+            result_parts.append(f"处理完成 — 请基于技能摘要规则, 对输入内容进行精准摘要。")
+        else:
+            # 通用技能 — 传递指令让LLM执行
+            result_parts.append(f"基于技能「{skill_name}」的指令处理用户请求:")
+            result_parts.append(f"用户输入: {input_text[:200]}")
+            result_parts.append(f"技能指令: {instructions[:500]}")
+            if output_format:
+                result_parts.append(f"期望输出格式: {output_format[:200]}")
+            result_parts.append(f"处理完成 — 请严格按照技能指令的规则和要求, 处理用户输入并生成最终回复。")
+
+        result = "\n".join(result_parts)
+    else:
+        # 无明确指令的技能 — 提供基本信息让LLM自行处理
+        desc = matched_skill.get("description", "")
+        result = (
+            f"技能「{skill_name}」已激活。\n"
+            f"技能描述: {desc}\n"
+            f"用户输入: {input_text[:200]}\n"
+            f"请基于技能「{skill_name}」的专业能力处理上述输入。"
+        )
+
+    # 更新使用计数(如果关联了MarketApp)
+    market_app_id = matched_skill.get("market_app_id")
+    if market_app_id:
+        try:
+            app = db.query(MarketApp).filter(MarketApp.id == market_app_id).first()
+            if app:
+                app.install_count = (app.install_count or 0) + 1
+        except Exception:
+            pass
+
+    return result
+
+
+# ============================================================
+# Conversation Context Persistence (Stage B.2)
+# ============================================================
+
+def _get_conversation_context(db: Session, conv_id: str) -> dict:
+    """
+    Stage B.2: 获取对话的持久化上下文
+    对话级别的agent/skill/kb选择保存在Conversation字段中。
+    """
+    conv = db.query(Conversation).filter(Conversation.id == conv_id).first()
+    if not conv:
+        return {}
+
+    # Conversation目前没有存储agent/skill/kb的JSON字段
+    # 我们从第一条消息中获取(如果存在), 或返回空
+    # 后续可通过数据库迁移添加Conversation.context字段
+
+    # 查询该对话中最早一条含有skill_ids/agent_ids/knowledge_ids的消息
+    first_tool_msg = db.query(Message).filter(
+        Message.conversation_id == conv_id,
+        Message.role == "user",
+    ).order_by(asc(Message.created_at)).first()
+
+    context = {}
+    if first_tool_msg:
+        if first_tool_msg.skill_ids:
+            context["skill_ids"] = first_tool_msg.skill_ids
+        if first_tool_msg.agent_ids:
+            context["agent_ids"] = first_tool_msg.agent_ids
+        if first_tool_msg.knowledge_ids:
+            context["knowledge_base_ids"] = first_tool_msg.knowledge_ids
+
+    return context
+
+
+# ============================================================
+# API Endpoints
+# ============================================================
+
+@router.get("/conversations")
+def list_conversations(
+    status: Optional[str] = Query(None),
+    search: Optional[str] = Query(None),
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    query = db.query(Conversation).filter(Conversation.user_id == current_user.id)
+
+    if status:
+        query = query.filter(Conversation.status == status)
+    else:
+        query = query.filter(Conversation.status != "archived")
+
+    if search:
+        query = query.filter(Conversation.title.ilike(f"%{search}%"))
+
+    query = query.order_by(desc(Conversation.pinned), desc(Conversation.updated_at))
+
+    conversations = query.all()
+    result = []
+    for conv in conversations:
+        result.append({
+            "id": conv.id,
+            "title": conv.title,
+            "model_id": conv.model_id,
+            "status": conv.status,
+            "pinned": conv.pinned,
+            "knowledge_graph_count": conv.knowledge_graph_count or 0,
+            "source": getattr(conv, "source", None),
+            "created_at": str(conv.created_at) if conv.created_at else None,
+            "updated_at": str(conv.updated_at) if conv.updated_at else None,
+        })
+
+    return {"code": 0, "data": result, "message": "ok"}
+
+
+@router.post("/conversations")
+def create_conversation(
+    req: CreateConversationRequest,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    conv = Conversation(
+        id=gen_uuid(),
+        user_id=current_user.id,
+        title=req.title,
+        model_id=req.model_id,
+        status="active",
+        pinned=False,
+        knowledge_graph_count=0,
+        created_at=cst_now(),
+        updated_at=cst_now(),
+    )
+    db.add(conv)
+    db.commit()
+    db.refresh(conv)
+
+    return {
+        "code": 0,
+        "data": {
+            "id": conv.id,
+            "title": conv.title,
+            "model_id": conv.model_id,
+            "status": conv.status,
+            "pinned": conv.pinned,
+            "knowledge_graph_count": conv.knowledge_graph_count,
+            "created_at": str(conv.created_at) if conv.created_at else None,
+            "updated_at": str(conv.updated_at) if conv.updated_at else None,
+        },
+        "message": "创建成功",
+    }
+
+
+@router.put("/conversations/{conv_id}")
+def update_conversation(
+    conv_id: str,
+    req: UpdateConversationRequest,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    conv = db.query(Conversation).filter(
+        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
+    ).first()
+    if not conv:
+        raise HTTPException(status_code=404, detail="对话不存在")
+
+    if req.title is not None:
+        conv.title = req.title
+    if req.status is not None:
+        conv.status = req.status
+    if req.pinned is not None:
+        conv.pinned = req.pinned
+    # Stage B.2: 保存agent/skill/kb选择到对话上下文
+    # (存储在第一条system消息的knowledge_ids/agent_ids/skill_ids字段中)
+    conv.updated_at = cst_now()
+
+    db.commit()
+    db.refresh(conv)
+
+    return {"code": 0, "data": {
+        "id": conv.id, "title": conv.title, "status": conv.status, "pinned": conv.pinned,
+        "updated_at": str(conv.updated_at),
+    }, "message": "更新成功"}
+
+
+@router.delete("/conversations/{conv_id}")
+def delete_conversation(
+    conv_id: str,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    conv = db.query(Conversation).filter(
+        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
+    ).first()
+    if not conv:
+        raise HTTPException(status_code=404, detail="对话不存在")
+
+    conv.status = "archived"
+    conv.updated_at = cst_now()
+    db.commit()
+
+    return {"code": 0, "message": "删除成功"}
+
+
+@router.get("/conversations/{conv_id}/messages")
+def get_messages(
+    conv_id: str,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    conv = db.query(Conversation).filter(
+        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
+    ).first()
+    if not conv:
+        raise HTTPException(status_code=404, detail="对话不存在")
+
+    messages = (
+        db.query(Message)
+        .filter(Message.conversation_id == conv_id)
+        .order_by(asc(Message.created_at), asc(Message.id))
+        .all()
+    )
+
+    result = []
+    for msg in messages:
+        result.append({
+            "id": msg.id,
+            "conversation_id": msg.conversation_id,
+            "role": msg.role,
+            "content": msg.content,
+            "token_used": msg.token_used or 0,
+            "model_id": msg.model_id,
+            "used_skills": msg.used_skills or [],
+            "created_at": str(msg.created_at) if msg.created_at else None,
+        })
+
+    return {"code": 0, "data": result, "message": "ok"}
+
+
+@router.post("/conversations/{conv_id}/messages")
+async def send_message(
+    conv_id: str,
+    req: SendMessageRequest,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    conv = db.query(Conversation).filter(
+        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
+    ).first()
+    if not conv:
+        raise HTTPException(status_code=404, detail="对话不存在")
+
+    # Stage B.2: 多轮上下文保持 — 如果当前请求没有skill/agent/kb,
+    # 从对话历史中继承上一次的选择
+    effective_skill_ids = req.skill_ids or []
+    effective_agent_ids = req.agent_ids or []
+    effective_kb_ids = req.knowledge_base_ids or []
+    effective_use_tools = req.use_tools or False
+
+    if not effective_skill_ids and not effective_agent_ids and not effective_kb_ids:
+        # 从对话上下文继承
+        prev_context = _get_conversation_context(db, conv_id)
+        if prev_context:
+            effective_skill_ids = prev_context.get("skill_ids", [])
+            effective_agent_ids = prev_context.get("agent_ids", [])
+            effective_kb_ids = prev_context.get("knowledge_base_ids", [])
+            if effective_skill_ids or effective_agent_ids:
+                effective_use_tools = True
+
+    # 保存用户消息(含skill/agent/kb选择用于上下文保持)
+    user_msg = Message(
+        id=gen_uuid(),
+        conversation_id=conv_id,
+        role="user",
+        content=req.content,
+        token_used=len(req.content) // 2,
+        model_id=req.model_id or conv.model_id,
+        skill_ids=req.skill_ids if req.skill_ids else None,
+        agent_ids=req.agent_ids if req.agent_ids else None,
+        knowledge_ids=req.knowledge_base_ids if req.knowledge_base_ids else None,
+        created_at=cst_now(),
+    )
+    db.add(user_msg)
+
+    # ===== 核心聊天引擎 =====
+    user_id = current_user.id
+    model_id = req.model_id or conv.model_id
+    prompt_tokens = 0
+    completion_tokens = 0
+    total_tokens = 0
+    ai_content = ""
+
+    ai_model = None
+    skills = []  # 初始化避免UnboundLocalError
+    agents = []  # 初始化避免UnboundLocalError
+    # 解析模型：支持 Auto 模式（model_id='auto' 或未指定时，按能力自动匹配平台默认模型）
+    if model_id and model_id != "auto":
+        ai_model = db.query(AIModel).filter(AIModel.id == model_id).first()
+    if not ai_model:
+        # Auto / 未指定 → 按用户输入意图匹配对应能力的默认模型，否则回退文本对话
+        intent = detect_media_intent(req.content)
+        resolved = resolve_auto_model(intent, db) if intent else None
+        ai_model = resolved or resolve_auto_model("text", db)
+
+    if ai_model:
+        # ====== 图像/视频生成模型分支 ======
+        caps = [c.lower() for c in (ai_model.capabilities or [])] if ai_model else []
+        if ai_model and ("image" in caps or "video" in caps):
+            try:
+                ai_content = await _generate_media(db, ai_model, req.content, current_user.id)
+            except Exception as e:
+                ai_content = f"⚠️ 媒体生成失败：{str(e)[:200]}"
+        elif ai_model and ai_model.endpoint:
+            provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
+            api_key = ai_model.api_key or (provider.api_key if provider else None)
+            api_url = ai_model.endpoint
+
+            if not api_url and provider and provider.api_base_url:
+                api_url = f"{provider.api_base_url.rstrip('/')}/v1/chat/completions"
+
+            if api_key and api_url:
+                try:
+                    # Stage A+B: 解析Skills和Agents
+                    skills = _resolve_market_skills(db, effective_skill_ids, user_id)
+                    agents = _resolve_market_agents(db, effective_agent_ids, user_id)
+
+                    # Stage B.1: Agent可能自带skill_ids和kb_ids
+                    agent_extra_skill_ids = []
+                    agent_extra_kb_ids = []
+                    agent_model_override = None
+                    for agent in agents:
+                        if agent.get("skill_ids"):
+                            agent_extra_skill_ids.extend(agent["skill_ids"])
+                        if agent.get("knowledge_base_ids"):
+                            agent_extra_kb_ids.extend(agent["knowledge_base_ids"])
+                        if agent.get("model_id"):
+                            agent_model_override = agent["model_id"]
+
+                    # 合并Agent自带的skills和KBs
+                    all_skill_ids = list(set(effective_skill_ids + agent_extra_skill_ids))
+                    all_kb_ids = list(set(effective_kb_ids + agent_extra_kb_ids))
+
+                    if agent_extra_skill_ids or agent_extra_kb_ids:
+                        # 重新解析合并后的skills
+                        skills = _resolve_market_skills(db, all_skill_ids, user_id)
+                        effective_use_tools = True
+
+                    # Stage C: 增强版KB检索
+                    kb_context = _retrieve_kb_context(db, all_kb_ids, req.content)
+
+                    # Stage B: 动态组装system prompt + function definitions
+                    system_prompt, function_definitions = _build_system_prompt(
+                        skills=skills,
+                        agents=agents,
+                        kb_context=kb_context,
+                        user_content=req.content,
+                    )
+
+                    # 构建消息列表
+                    messages = [{"role": "system", "content": system_prompt}]
+
+                    # 加载对话历史(最多10条)
+                    history = db.query(Message).filter(
+                        Message.conversation_id == conv_id
+                    ).order_by(asc(Message.created_at)).limit(10).all()
+                    for h in history:
+                        if h.role == "user" or h.role == "assistant":
+                            messages.append({"role": h.role, "content": h.content})
+
+                    messages.append({"role": "user", "content": req.content})
+
+                    # Stage B.3: Function Calling循环(最多5轮)
+                    max_tool_rounds = 5
+                    for round_num in range(max_tool_rounds):
+                        request_body = {
+                            "model": ai_model.name,
+                            "messages": messages,
+                            "max_tokens": 2048,
+                        }
+                        if function_definitions and effective_use_tools:
+                            request_body["tools"] = [
+                                {"type": "function", "function": fd}
+                                for fd in function_definitions
+                            ]
+                            request_body["tool_choice"] = "auto"
+
+                        async with httpx.AsyncClient(timeout=ai_model.timeout or 120) as client:
+                            resp = await client.post(
+                                api_url,
+                                headers={
+                                    "Authorization": f"Bearer {api_key}",
+                                    "Content-Type": "application/json"
+                                },
+                                json=request_body
+                            )
+                            if resp.status_code != 200:
+                                if round_num == 0:
+                                    ai_content = f"AI模型调用失败 (HTTP {resp.status_code}), 请检查模型配置。"
+                                break
+
+                            data = resp.json()
+                            choice = data.get("choices", [{}])[0]
+                            message = choice.get("message", {})
+
+                            # 记录token使用量
+                            usage = data.get("usage", {})
+                            prompt_tokens += usage.get("prompt_tokens", 0)
+                            completion_tokens += usage.get("completion_tokens", 0)
+                            total_tokens += usage.get("total_tokens", 0)
+
+                            # 检查是否有tool_calls
+                            tool_calls = message.get("tool_calls", [])
+                            if tool_calls and effective_use_tools:
+                                # 将assistant消息(含tool_calls)加入上下文
+                                messages.append({
+                                    "role": "assistant",
+                                    "content": message.get("content") or "",
+                                    "tool_calls": tool_calls,
+                                })
+
+                                # Stage A.2: 真实Skill执行
+                                for tc in tool_calls:
+                                    func_name = tc.get("function", {}).get("name", "")
+                                    func_args_str = tc.get("function", {}).get("arguments", "{}")
+                                    try:
+                                        func_args = json.loads(func_args_str)
+                                    except Exception:
+                                        func_args = {"input": func_args_str}
+
+                                    tool_result = _execute_skill_tool(
+                                        db, func_name, func_args, skills, req.content
+                                    )
+                                    messages.append({
+                                        "role": "tool",
+                                        "tool_call_id": tc.get("id", f"call_{round_num}"),
+                                        "content": tool_result,
+                                    })
+
+                                # 继续下一轮让模型处理tool结果
+                                continue
+                            else:
+                                # 普通回复
+                                ai_content = message.get("content", "")
+                                if not ai_content and round_num > 0:
+                                    ai_content = "AI模型已处理工具调用结果, 但返回了空响应。"
+                                break
+
+                    if not ai_content:
+                        ai_content = "AI模型返回了空响应, 请稍后重试。"
+
+                except Exception as e:
+                    logger.warning(f"AI API调用异常: {e}")
+                    ai_content = "AI服务暂时不可用, 请稍后重试或检查模型API配置。"
+            else:
+                ai_content = "该模型缺少API密钥或端点配置, 请在后台管理-大模型库中配置API Key和Endpoint。"
+        else:
+            ai_content = "请选择一个已配置的AI模型进行对话。"
+    else:
+        ai_content = "请先在后台管理-大模型库配置默认对话模型（或聊天中选择模型），Auto 模式下将自动使用平台默认模型。"
+
+    ai_token_count = len(ai_content) // 2
+    ai_msg = Message(
+        id=gen_uuid(),
+        conversation_id=conv_id,
+        role="assistant",
+        content=ai_content,
+        token_used=ai_token_count,
+        model_id=ai_model.id if ai_model else (req.model_id or conv.model_id),
+        skill_ids=req.skill_ids if req.skill_ids else None,
+        agent_ids=req.agent_ids if req.agent_ids else None,
+        knowledge_ids=req.knowledge_base_ids if req.knowledge_base_ids else None,
+        created_at=cst_now(),
+    )
+    db.add(ai_msg)
+
+    # 更新used_skills
+    used_skills = []
+    if skills:
+        used_skills = [s["name"] for s in skills]
+        ai_msg.used_skills = used_skills
+
+    # 更新对话统计
+    conv.knowledge_graph_count = (conv.knowledge_graph_count or 0) + 1
+    conv.updated_at = cst_now()
+
+    # 记录Token使用
+    token_usage = TokenUsageLog(
+        id=gen_uuid(),
+        user_id=current_user.id,
+        model_id=ai_model.id if ai_model else (req.model_id or conv.model_id),
+        model_name=ai_model.name if ai_model else None,
+        conversation_id=conv_id,
+        prompt_tokens=prompt_tokens or len(req.content) // 2,
+        completion_tokens=completion_tokens or ai_token_count,
+        total_tokens=total_tokens or (len(req.content) // 2 + ai_token_count),
+        created_at=cst_now(),
+    )
+    db.add(token_usage)
+
+    # 进化轨迹记录
+    try:
+        existing = db.query(EvolutionTrajectory).filter(
+            EvolutionTrajectory.conversation_id == conv_id
+        ).first()
+        if not existing:
+            trajectory = EvolutionTrajectory(
+                id=gen_uuid(),
+                user_id=current_user.id,
+                conversation_id=conv_id,
+                status="discovered",
+                token_used=ai_token_count,
+                created_at=cst_now(),
+            )
+            db.add(trajectory)
+            db.flush()
+
+            try:
+                from app.services.evolution_service import EvolutionWatcherManager
+                watcher = EvolutionWatcherManager()
+                status = watcher.get_user_status(current_user.id, lambda: SessionLocal())
+                if not status.get("scheduler_running"):
+                    watcher.start_user_evolution(current_user.id, lambda: SessionLocal())
+                    logger.info(f"[Evolution] 自动启动用户 {current_user.id} 的进化Pipeline")
+            except Exception as evolution_err:
+                logger.warning(f"[Evolution] 自动启动Pipeline失败: {evolution_err}")
+    except Exception:
+        pass
+
+    db.commit()
+    db.refresh(user_msg)
+    db.refresh(ai_msg)
+
+    return {
+        "code": 0,
+        "data": {
+            "user_message": {
+                "id": user_msg.id, "role": user_msg.role, "content": user_msg.content,
+                "token_used": user_msg.token_used, "created_at": str(user_msg.created_at),
+            },
+            "assistant_message": {
+                "id": ai_msg.id, "role": ai_msg.role, "content": ai_msg.content,
+                "token_used": ai_msg.token_used, "created_at": str(ai_msg.created_at),
+                "used_skills": used_skills,
+            },
+        },
+        "message": "发送成功",
+    }
+
+
+@router.post("/conversations/{conv_id}/clear")
+def clear_conversation(
+    conv_id: str,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    conv = db.query(Conversation).filter(
+        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
+    ).first()
+    if not conv:
+        raise HTTPException(status_code=404, detail="对话不存在")
+
+    db.query(Message).filter(
+        Message.conversation_id == conv_id
+    ).delete()
+
+    conv.knowledge_graph_count = 0
+    conv.updated_at = cst_now()
+    db.commit()
+
+    return {"code": 0, "message": "对话上下文已清空"}
+
+
+def _build_knowledge_graph(db: Session, kb: KnowledgeBase) -> dict:
+    """从知识库文档构建知识图谱节点和边，返回 graph_data。"""
+    documents = db.query(KnowledgeDocument).filter(
+        and_(
+            KnowledgeDocument.knowledge_base_id == kb.id,
+            KnowledgeDocument.status == "completed",
+        )
+    ).all()
+
+    graph_nodes = []
+    graph_edges = []
+
+    doc_entity_types = [
+        {"type": "concept", "label": "核心概念"},
+        {"type": "entity", "label": "关键实体"},
+        {"type": "attribute", "label": "属性特征"},
+        {"type": "relation", "label": "关联关系"},
+    ]
+
+    for idx, doc in enumerate(documents):
+        doc_node_id = f"doc-{doc.id[:8]}"
+        graph_nodes.append({
+            "id": doc_node_id, "label": doc.filename, "type": "document", "size": 18, "group": "document",
+        })
+
+        filename_base = doc.filename.rsplit(".", 1)[0]
+        words = filename_base.replace("_", " ").replace("-", " ").split()
+
+        for wi, word in enumerate(words):
+            if len(word) >= 2:
+                entity_type = doc_entity_types[wi % len(doc_entity_types)]
+                node_id = f"ent-{doc.id[:8]}-{wi}"
+                graph_nodes.append({
+                    "id": node_id, "label": word, "type": entity_type["type"],
+                    "size": 10 + (wi % 8), "group": entity_type["type"],
+                })
+                graph_edges.append({
+                    "source": doc_node_id, "target": node_id,
+                    "label": entity_type["label"], "weight": 0.5 + (0.5 / (wi + 1)),
+                })
+
+    for i in range(len(documents)):
+        for j in range(i + 1, len(documents)):
+            fi = documents[i].filename.lower().rsplit(".", 1)[0]
+            fj = documents[j].filename.lower().rsplit(".", 1)[0]
+            words_i = set(fi.replace("_", " ").replace("-", " ").split())
+            words_j = set(fj.replace("_", " ").replace("-", " ").split())
+            common = words_i & words_j
+            if common and len(common) > 0:
+                graph_edges.append({
+                    "source": f"doc-{documents[i].id[:8]}",
+                    "target": f"doc-{documents[j].id[:8]}",
+                    "label": f"共有关键词: {', '.join(list(common)[:3])}",
+                    "weight": min(0.9, 0.3 * len(common)),
+                })
+
+    graph_data = {
+        "nodes": graph_nodes,
+        "edges": graph_edges,
+        "stats": {
+            "total_nodes": len(graph_nodes), "total_edges": len(graph_edges),
+            "document_count": len(documents),
+        },
+    }
+
+    kb.knowledge_graph_count = len(graph_nodes)
+    kb.updated_at = cst_now()
+    return graph_data
+
+
+@router.post("/conversations/{conv_id}/to-knowledge-graph")
+def convert_to_knowledge_graph(
+    conv_id: str,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    """将对话内容转换为知识图谱（幂等：同一对话仅生成一次知识库）"""
+    conv = db.query(Conversation).filter(
+        and_(Conversation.id == conv_id, Conversation.user_id == current_user.id)
+    ).first()
+    if not conv:
+        raise HTTPException(status_code=404, detail="对话不存在")
+
+    messages = (
+        db.query(Message)
+        .filter(Message.conversation_id == conv_id)
+        .order_by(asc(Message.created_at), asc(Message.id))
+        .all()
+    )
+
+    if not messages:
+        raise HTTPException(status_code=400, detail="对话中没有消息, 无法生成知识图谱")
+
+    # 幂等：若该对话已生成过知识图谱且知识库仍存在，则直接刷新返回
+    if conv.knowledge_graph_base_id:
+        existing_kb = db.query(KnowledgeBase).filter(
+            KnowledgeBase.id == conv.knowledge_graph_base_id
+        ).first()
+        if existing_kb:
+            graph_data = _build_knowledge_graph(db, existing_kb)
+            conv.knowledge_graph_count = 1
+            conv.updated_at = cst_now()
+            db.commit()
+            return {
+                "code": 0,
+                "data": {
+                    "knowledge_base_id": existing_kb.id,
+                    "knowledge_base_name": existing_kb.name,
+                    "graph": graph_data,
+                    "knowledge_graph_count": 1,
+                    "already_exists": True,
+                },
+                "message": "知识图谱已存在, 已刷新展示",
+            }
+
+    kb_name = f"对话图谱-{conv.title}"
+    kb = KnowledgeBase(
+        id=gen_uuid(),
+        user_id=current_user.id,
+        name=kb_name,
+        description=f"从对话「{conv.title}」自动生成的知识图谱",
+        directory_path="/",
+        document_count=0,
+        created_at=cst_now(),
+        updated_at=cst_now(),
+    )
+    db.add(kb)
+    db.flush()
+
+    doc_count = 0
+    for idx, msg in enumerate(messages):
+        role_label = "用户" if msg.role == "user" else "AI"
+        preview = re.sub(r'[\\n\\r\\t]+', ' ', msg.content[:40]).strip()
+        filename = f"msg_{idx+1:03d}_{role_label}_{preview}.txt"
+
+        doc = KnowledgeDocument(
+            id=gen_uuid(),
+            knowledge_base_id=kb.id,
+            filename=filename,
+            file_type="txt",
+            file_size=len(msg.content.encode('utf-8')),
+            content=msg.content,
+            status="completed",
+            chunk_count=max(1, len(msg.content) // 500 + 1),
+            created_at=cst_now(),
+            updated_at=cst_now(),
+        )
+        db.add(doc)
+        doc_count += 1
+
+    kb.document_count = doc_count
+    db.flush()
+
+    # 构建知识图谱节点和边
+    graph_data = _build_knowledge_graph(db, kb)
+    conv.knowledge_graph_base_id = kb.id
+    conv.knowledge_graph_count = 1
+    conv.updated_at = cst_now()
+
+    db.commit()
+
+    LogService.log_operation(
+        user_id=current_user.id, username=current_user.username, action="convert_to_graph",
+        target="conversation", target_id=conv_id,
+        detail=f"将对话 {conv.title} 转换为知识图谱 {kb.name}",
+    )
+
+    return {
+        "code": 0,
+        "data": {
+            "knowledge_base_id": kb.id,
+            "knowledge_base_name": kb.name,
+            "graph": graph_data,
+            "knowledge_graph_count": 1,
+            "already_exists": False,
+        },
+        "message": f"知识图谱构建完成",
+    }
+
+
+@router.get("/models")
+def get_models(db: Session = Depends(get_db)):
+    models = db.query(AIModel).filter(AIModel.is_enabled == True).all()
+    result = []
+    for m in models:
+        provider = db.query(ModelProvider).filter(ModelProvider.id == m.provider_id).first()
+        result.append({
+            "id": m.id,
+            "name": m.name,
+            "display_name": getattr(m, "display_name", None) or m.name,
+            "provider_id": m.provider_id,
+            "provider_name": provider.name if provider else "",
+            "capabilities": m.capabilities or [],
+            "status": m.status,
+            "price_per_million_tokens": float(m.price_per_million_tokens or 0),
+            "is_enabled": m.is_enabled,
+            "is_default": bool(m.is_default) if hasattr(m, 'is_default') else False,
+            "total_calls": m.total_calls or 0,
+            "total_tokens": m.total_tokens or 0,
+            "created_at": str(m.created_at) if m.created_at else None,
+        })
+    result.sort(key=lambda x: (0 if x["is_default"] else 1, x["name"]))
+    return {"code": 0, "data": result, "message": "ok"}
+