From 996d5413da85dade692ddc75675e4a24d0f966e9 Mon Sep 17 00:00:00 2001
From: QihuAI Design <ai@qihuyun.com>
Date: Wed, 15 Jul 2026 12:23:56 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20=E4=BF=AE=E5=A4=8DAI=E8=81=8A=E5=A4=A9/I?=
 =?UTF-8?q?M=E5=A4=9A=E9=A1=B9=E9=97=AE=E9=A2=98(=E5=AA=92=E4=BD=93?=
 =?UTF-8?q?=E7=9B=B4=E8=BF=9E+=E6=8A=80=E8=83=BD=E8=87=AA=E5=8A=A8?=
 =?UTF-8?q?=E5=8F=91=E7=8E=B0+agent/mcp=E4=B8=8A=E4=B8=8B=E6=96=87)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Issue4/7: 图像/视频生成对AGNES国内域名强制直连(trust_env=False)，修复Server disconnected(代理172.18.0.1:7890返回502)

- Issue8: 修复_parse_skill_content缺失def导致的NameError；新增_discover_installed_skills默认技能自动发现

- Issue2/3: system prompt注入agent资料与mcp描述/endpoint

- Issue7: IM Bot制作图片走Auto路由媒体模型
---
 backend/python/app/api/chat.py   | 131 +++++++++++++++++++++++++++----
 backend/python/app/api/im_llm.py | 130 ++++++++++++++++++++++++++----
 2 files changed, 232 insertions(+), 29 deletions(-)

diff --git a/backend/python/app/api/chat.py b/backend/python/app/api/chat.py
index cf29f07..a962cb8 100644
--- a/backend/python/app/api/chat.py
+++ b/backend/python/app/api/chat.py
@@ -125,13 +125,30 @@ def _html_escape(s: str) -> str:
 
 
 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)。"""
+    """出网代理地址(兜底)。默认经 172.18.0.1:7890 智能路由代理；
+    运行时若未注入代理环境变量则回退到该地址。"""
     import os
     return os.environ.get("HTTPS_PROXY") or os.environ.get("HTTP_PROXY") or "http://172.18.0.1:7890"
 
 
+def _proxy_for_url(url: str) -> str | None:
+    """按目标域名选择是否走代理。
+
+    实测(2026-07-15)：apihub.agnes-ai.com / platform-outputs.agnes-ai.space 为国内域名，
+    容器内直连稳定返回 200；而经智能路由代理 172.18.0.1:7890 当前返回 502，导致
+    httpx 报 "Server disconnected without sending a response"。
+    因此对 AGNES 等国内域名返回 None(直连)，其余域名才走 _outbound_proxy()。
+
+    注意：httpx 默认 trust_env=True 会忽略 proxy=None 而改用环境变量 HTTP_PROXY，
+    故调用方必须同时传 trust_env=False 才能真正直连。
+    """
+    from urllib.parse import urlparse
+    host = (urlparse(url).hostname or "").lower()
+    if "agnes-ai.com" in host or "agnes-ai.space" in host:
+        return None
+    return _outbound_proxy()
+
+
 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)。"""
@@ -151,14 +168,16 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
             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。"
+        # 注意：extra_body 是 OpenAI SDK 专属参数，不能直接作为原始 JSON 字段下发，
+        # AGNES 收到未知顶层字段会直接断开连接（表现为 "Server disconnected without
+        # sending a response"）。保持与 OpenAI 兼容的极简结构即可。
         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:
+            async with httpx.AsyncClient(timeout=timeout, proxy=_proxy_for_url(api_url), trust_env=False) 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]}"
@@ -194,7 +213,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
             "frame_rate": 24,
         }
         try:
-            async with httpx.AsyncClient(timeout=timeout, proxy=_outbound_proxy()) as client:
+            async with httpx.AsyncClient(timeout=timeout, proxy=_proxy_for_url(api_url), trust_env=False) 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]}"
@@ -450,6 +469,35 @@ def _resolve_market_mcp(db: Session, mcp_ids: list, user_id: str) -> list:
         })
 
     return results
+
+
+def _discover_installed_skills(db: Session, user_id: str, limit: int = 15) -> list:
+    """
+    Issue 8: 默认技能自动发现。
+    检索「本账号已安装(app_installations) 且 平台已部署发布(MarketApp.status='published')」的
+    skill 型应用，返回可供 function calling 使用的技能定义列表（最多 limit 个）。
+    仅当用户未显式选择技能时才启用；多个技能适用时由模型自行判断选 1 个最优。
+    """
+    if not user_id:
+        return []
+    installed = db.query(AppInstallation).filter(
+        AppInstallation.user_id == user_id
+    ).all()
+    app_ids = [i.app_id for i in installed]
+    if not app_ids:
+        return []
+    apps = db.query(MarketApp).filter(
+        MarketApp.id.in_(app_ids),
+        MarketApp.type == "skill",
+        MarketApp.status == "published",
+    ).all()
+    ids = [a.id for a in apps]
+    if not ids:
+        return []
+    return _resolve_market_skills(db, ids, user_id)[:limit]
+
+
+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:
@@ -727,6 +775,7 @@ def _build_system_prompt(
     kb_context: str,
     user_content: str,
     mcp: list = None,
+    auto_discovered: bool = False,
 ) -> tuple:
     """
     动态组装system prompt + function definitions。
@@ -747,21 +796,41 @@ def _build_system_prompt(
 
     function_definitions = []
 
-    # Stage B.1: Agent角色注入
+    # Stage B.1: Agent角色注入（Issue 2：让用户询问 agent 能力时，AI 能“看到”该 agent 资料）
     if agents:
         for agent in agents:
+            name = agent.get("name", "")
             role = agent.get("role_prompt", "")
             instructions = agent.get("instructions", "")
+            desc = agent.get("description", "")
+            block = []
+            if name:
+                block.append(
+                    f"用户已选择智能体「{name}」。当用户询问该智能体的功能、能力、用法或相关问题时，"
+                    f"请基于以下资料准确作答，不要说“不知道这个智能体”。"
+                )
             if role:
-                system_parts.append(f"当前角色设定: {role}")
+                block.append(f"该智能体的角色设定: {role}")
+            if desc:
+                block.append(f"该智能体的简介: {desc}")
             if instructions:
-                system_parts.append(f"角色行为指引: {instructions}")
+                block.append(f"该智能体的行为指引: {instructions}")
+            if block:
+                system_parts.append("\n".join(block))
 
     # 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。")
+        if auto_discovered:
+            # Issue 8：每轮默认先寻找本账号已安装且平台已部署的技能，多个适用时 AI 自行判断选 1 个最优
+            system_parts.append(
+                "重要规则: 以上技能为本账号已安装且平台已部署的技能。每轮对话请先判断用户请求是否可由其中某个技能处理；"
+                "若多个技能都适用，请自行判断并选择最贴切的 1 个优先调用，不要一次性全部调用。"
+                "调用对应工具后将获得执行结果，基于结果生成最终回复。若这些技能均不适用，则直接基于你的知识回答。"
+            )
+        else:
+            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:
@@ -812,11 +881,26 @@ def _build_system_prompt(
     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}")
 
-    # MCP 应用注入
+    # MCP 应用注入（Issue 3：让用户询问 MCP 能力时，AI 能“看到”该 MCP 的资料与地址）
     if mcp:
-        mcp_descs = ", ".join([f"{s['name']}({s.get('transport', 'stdio')})" for s in mcp])
-        system_parts.append(f"已连接的 MCP 应用(外部工具服务): {mcp_descs}")
-        system_parts.append("当用户请求需要调用外部系统/工具（如查询数据库、调用API、操作第三方服务）时，可参考上述 MCP 应用提供的能力；若平台已接入对应 MCP 工具，优先通过工具调用完成。")
+        mcp_lines = []
+        for m in mcp:
+            name = m.get("name", "")
+            desc = m.get("description", "")
+            transport = m.get("transport", "stdio")
+            endpoint = m.get("endpoint", "")
+            line = f"用户已选择 MCP 应用「{name}」（传输方式: {transport}）"
+            if desc:
+                line += f"，功能简介: {desc}"
+            if endpoint:
+                line += f"，服务地址: {endpoint}"
+            mcp_lines.append(line)
+        system_parts.append("\n".join(mcp_lines))
+        system_parts.append(
+            "用户已选择上述 MCP 应用。当请求需要借助外部工具/系统（如查询数据、调用 API、操作第三方服务）时，"
+            "请优先参考这些 MCP 应用的能力来回答；若平台已接入对应工具，优先通过工具调用完成；"
+            "不要说“不了解该 MCP”。"
+        )
 
     system_prompt = "\n".join(system_parts)
     return system_prompt, function_definitions
@@ -1281,6 +1365,20 @@ async def send_message(
                     agents = _resolve_market_agents(db, effective_agent_ids, user_id)
                     mcps = _resolve_market_mcp(db, effective_mcp_ids, user_id)
 
+                    # Issue 8: 默认技能自动发现（仅当用户未显式选择技能时启用）
+                    # 每轮先寻找本账号已安装且平台已部署的技能，多个适用时由模型自行判断选 1 个最优。
+                    used_auto_skills = False
+                    if not effective_skill_ids:
+                        try:
+                            discovered = _discover_installed_skills(db, user_id)
+                        except Exception as e:
+                            logger.warning(f"自动发现已安装技能失败: {e}")
+                            discovered = []
+                        if discovered:
+                            skills = discovered
+                            used_auto_skills = True
+                            effective_use_tools = True
+
                     # Stage B.1: Agent可能自带skill_ids和kb_ids
                     agent_extra_skill_ids = []
                     agent_extra_kb_ids = []
@@ -1312,6 +1410,7 @@ async def send_message(
                         kb_context=kb_context,
                         user_content=req.content,
                         mcp=mcps,
+                        auto_discovered=used_auto_skills,
                     )
 
                     # 构建消息列表
@@ -1325,7 +1424,9 @@ async def send_message(
                         if h.role == "user" or h.role == "assistant":
                             messages.append({"role": h.role, "content": h.content})
 
-                    messages.append({"role": "user", "content": req.content})
+                    # Issue 1: 必须把附件内容(attachment_block)一并发给模型，
+                    # 否则用户在对话窗上传文件后模型“看不到”文件内容。
+                    messages.append({"role": "user", "content": req.content + attachment_block})
 
                     # Stage B.3: Function Calling循环(最多5轮)
                     max_tool_rounds = 5
diff --git a/backend/python/app/api/im_llm.py b/backend/python/app/api/im_llm.py
index 7cd0f9e..ad87c2b 100644
--- a/backend/python/app/api/im_llm.py
+++ b/backend/python/app/api/im_llm.py
@@ -3,7 +3,15 @@ IM Bot 真实 LLM 调用。
 
 原实现位于 channels.py（通用频道管理模块），该模块已移除。
 为保持 IM Bot（微信机器人等）正常生成回复，将 _call_real_llm 迁移至此独立模块。
+
+行为（对齐 AI 聊天）：
+- Issue 7：图像/视频/语音意图优先走平台媒体模型（与 AI 聊天同一套 Auto 路由），
+  不再对“制作图片”这类请求只回文本。
+- Issue 8：默认 Auto 模型；每轮先自动发现本账号已安装且平台已部署的技能，
+  多个适用时由模型自行判断选 1 个最优，通过 function calling 执行。
 """
+import asyncio
+import json
 from typing import Optional
 
 import httpx
@@ -13,9 +21,54 @@ from ..capability_registry import resolve_auto_model
 from ..models import AIModel, ModelProvider
 
 
-def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im") -> str:
-    """调用七虎AI 默认文本模型生成回复，供 IM Bot 复用。"""
+def _run_async(coro):
+    """在同步上下文中运行协程，兼容「无事件循环」与「已有事件循环」两种场景。
+
+    IM 个人微信在独立线程中调用（无事件循环）→ asyncio.run；
+    IM 多平台 Webhook 可能在 FastAPI 协程内调用（已有事件循环）→ 丢到线程池执行。
+    """
     try:
+        loop = asyncio.get_running_loop()
+    except RuntimeError:
+        loop = None
+    if loop is not None and loop.is_running():
+        import concurrent.futures
+        with concurrent.futures.ThreadPoolExecutor(1) as ex:
+            return ex.submit(lambda: asyncio.run(coro)).result()
+    return asyncio.run(coro)
+
+
+def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im", user_id: str = None) -> str:
+    """调用七虎AI 生成回复，供 IM Bot 复用。支持媒体意图路由与已安装技能自动发现。"""
+    try:
+        from app.api.chat import (
+            _generate_media,
+            detect_media_intent,
+            _discover_installed_skills,
+            _build_system_prompt,
+            _execute_skill_tool,
+        )
+
+        # ===== 1) 媒体意图路由（Issue 7）=====
+        intent = detect_media_intent(prompt)
+        if intent:
+            ai_model = resolve_auto_model(intent, db)
+            if ai_model:
+                try:
+                    result = _run_async(_generate_media(db, ai_model, prompt, user_id))
+                except Exception as e:
+                    result = f"（媒体生成失败：{e}）"
+                if result:
+                    # IM 不渲染 HTML，把 <img>/<video> 还原为纯文本链接返回
+                    if result.startswith("<img") or result.startswith("<video"):
+                        import re as _re
+                        m = _re.search(r'src="([^"]+)"', result)
+                        url = m.group(1) if m else ""
+                        label = "图片" if result.startswith("<img") else "视频"
+                        return f"已为你生成{label}：{url}" if url else result
+                    return result
+
+        # ===== 2) 文本回复 + 技能自动发现（Issue 8）=====
         ai_model = resolve_auto_model("text", db)
         if not ai_model:
             return "（未配置默认文本模型，无法生成回复）"
@@ -26,20 +79,69 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
             api_url = f"{provider.api_base_url.rstrip('/')}/v1/chat/completions"
         if not (api_key and api_url):
             return "（模型未配置 API Key 或地址，无法生成回复）"
+
+        # 自动发现本账号已安装且平台已部署的技能
+        skills = []
+        if user_id:
+            try:
+                skills = _discover_installed_skills(db, user_id)[:15]
+            except Exception:
+                skills = []
+
+        system_prompt, function_definitions = _build_system_prompt(
+            skills=skills,
+            agents=[],
+            kb_context="",
+            user_content=prompt,
+            mcp=[],
+            auto_discovered=bool(skills),
+        )
         messages = [
-            {"role": "system", "content": "你是七虎AI 助手，负责在即时通讯中回复用户消息。"},
+            {"role": "system", "content": system_prompt},
             {"role": "user", "content": prompt},
         ]
-        request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
-        with httpx.Client(timeout=ai_model.timeout or 120) as client:
-            resp = client.post(
-                api_url,
-                headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
-                json=request_body,
-            )
-            if resp.status_code != 200:
-                return f"（AI 模型调用失败 HTTP {resp.status_code}）"
-            data = resp.json()
-            return data.get("choices", [{}])[0].get("message", {}).get("content", "")
+
+        max_rounds = 3
+        ai_content = ""
+        for _ in range(max_rounds):
+            request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
+            if function_definitions and skills:
+                request_body["tools"] = [
+                    {"type": "function", "function": fd} for fd in function_definitions
+                ]
+                request_body["tool_choice"] = "auto"
+            with httpx.Client(timeout=ai_model.timeout or 120) as client:
+                resp = client.post(
+                    api_url,
+                    headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+                    json=request_body,
+                )
+                if resp.status_code != 200:
+                    return f"（AI 模型调用失败 HTTP {resp.status_code}）"
+                data = resp.json()
+                message = data.get("choices", [{}])[0].get("message", {})
+                tool_calls = message.get("tool_calls", [])
+                if tool_calls and skills:
+                    messages.append({
+                        "role": "assistant",
+                        "content": message.get("content") or "",
+                        "tool_calls": tool_calls,
+                    })
+                    for tc in tool_calls:
+                        func_name = tc.get("function", {}).get("name", "")
+                        try:
+                            func_args = json.loads(tc.get("function", {}).get("arguments", "{}"))
+                        except Exception:
+                            func_args = {"input": tc.get("function", {}).get("arguments", "")}
+                        tool_result = _execute_skill_tool(db, func_name, func_args, skills, prompt)
+                        messages.append({
+                            "role": "tool",
+                            "tool_call_id": tc.get("id", "call_1"),
+                            "content": tool_result,
+                        })
+                    continue
+                ai_content = message.get("content", "")
+                break
+        return ai_content or "（AI 回复为空）"
     except Exception as e:
         return f"（AI 回复生成失败：{e}）"