fix: 双Claw能力对齐 + MCP/LLM代理绕过（P0止血）

- 新增 app/core/capability_discovery.py：统一能力发现层，Web/IM 双Claw共用单一实现
  （三级降级：用户安装 → 平台published → 相关性裁剪；MCP 自动挂载已启用服务）
- im_llm.py：IM Bot Claw 改用统一发现层，恢复 skills(12) + MCP(26工具) 能力，
  修复 #866 部署回退导致的 agents=[]/mcp=[] 回归；mcp__ 前缀分流到 call_mcp_tool
- mcp_client.py：注入 httpx_client_factory(trust_env=False) 绕过容器 HTTP_PROXY，
  修复外部 MCP(xiaobenyang.com / anysearch.com) tools/list 全部静默失败（mcp_tools 0→26）
- llm_client.py / im_bot_user.py：2026-07-31 代理劫持内部请求 + 媒体断连修复（随镜像已上线）

backend/python/app/api/im_bot_user.py

@@ -143,12 +143,33 @@ def _qihua_reply(text: str, user_id: str = None) -> str:
                 f"{hist}\n\n"
                 f"用户当前问题：{text}"
             )
+    # 磁盘调试日志（不依赖 uvicorn logger）
+    try:
+        with open("/tmp/im_llm_debug.log", "a", encoding="utf-8") as _f:
+            _f.write(_beijing_now().isoformat() + f" _qihua_reply user={user_id} prompt_len={len(prompt)} prompt={repr(prompt)[:500]}\n")
+    except Exception:
+        pass
     db = None
     try:
         from app.api.im_llm import _call_real_llm
         db = SessionLocal()
-        return _call_real_llm(db, None, prompt, "个人微信", user_id=user_id)
+        result = _call_real_llm(db, None, prompt, "个人微信", user_id=user_id)
+        try:
+            with open("/tmp/im_llm_debug.log", "a", encoding="utf-8") as _f:
+                _f.write(_beijing_now().isoformat() + f" _qihua_reply RESULT len={len(result) if result else 0} preview={repr(result)[:200]}\n")
+        except Exception:
+            pass
+        return result
     except Exception as e:
+        import traceback as _tb2
+        import logging
+        _log = logging.getLogger('im_bot_user')
+        _log.warning(f'[_qihua_reply] 异常: {e}\n{_tb2.format_exc()[:1000]}')
+        try:
+            with open("/tmp/im_llm_debug.log", "a", encoding="utf-8") as _f:
+                _f.write(_beijing_now().isoformat() + f" _qihua_reply EXCEPTION: {e}\n{_tb2.format_exc()[:800]}\n")
+        except Exception:
+            pass
         return f"（AI 回复生成失败：{e}）"
     finally:
         if db:
@@ -164,7 +185,7 @@ def _wechat_get_updates(baseurl: str, bot_token: str, cursor: str):
         resp = requests.post(
             baseurl.rstrip("/") + "/ilink/bot/getupdates",
             headers=_wechat_headers(bot_token),
-            json={"get_updates_buf": cursor, "base_info": {"channel_version": "1.0.2"}},
+            json={"get_updates_buf": cursor},
             timeout=40,
         )
         if not resp.ok:
@@ -178,20 +199,26 @@ def _wechat_get_updates(baseurl: str, bot_token: str, cursor: str):
 def _wechat_post_msg(baseurl: str, bot_token: str, to_user_id: str, context_token: str, item_list: list):
     """向 iLink 发送一条消息（item_list 可含文本/图片/视频等）。返回 (ok, ret, body)。
     ret=-14=会话过期需重扫；ret=-2=token 失效。"""
-    import uuid as _uuid
     payload = {
         "msg": {
-            "from_user_id": "",
             "to_user_id": to_user_id,
-            "client_id": "py-" + str(_uuid.uuid4()),
             "message_type": 2,
             "message_state": 2,
             "context_token": context_token,
             "item_list": item_list,
-        },
+        }
-        "base_info": {"channel_version": "1.0.2"},
     }
     last_body = {}
+    # 07-30 diagnostic
+    import json as _json3
+    try:
+        with open("/tmp/wechat_send_debug.log", "a", encoding="utf-8") as _f:
+            _f.write(_beijing_now().isoformat() + " SEND_DEBUG token=" + repr(bot_token[:20] if bot_token else "NULL")
+                     + " to=" + repr(to_user_id[:30]) + " ctx=" + repr(context_token[:30]) + "\n")
+            _f.write("  payload_keys=" + str(list(payload.get("msg", {}).keys())) + "\n")
+    except Exception:
+        pass
+
     for _ in range(2):
         try:
             resp = requests.post(
@@ -239,21 +266,21 @@ def _platform_relay_url(raw_url: str) -> str:
 
 def _wechat_send_media(baseurl: str, bot_token: str, to_user_id: str, context_token: str,
                        media_type: str, url: str):
-    """发送媒体（图片/视频）并在微信内联渲染。
+    """发送媒体（图片/视频）。
 
-    问题：AGNES 媒体 URL 位于境外，iLink(中国大陆服务器)直连常被墙/超时，
+    问题：AGNES 媒体 URL 位于境外，iLink(中国大陆服务器)直连常被墙/超时。
-    微信里只能看到"链接"而无法内联播放。
+    解决：改写为平台中转地址(经香港代理可达 AGNES)。
-    解决：改写为平台中转地址(经香港代理可达 AGNES)，微信即可内联显示图片/视频。
+
+    2026-07-31 兼容修复：iLink 对 type=2/5 的媒体 item 返回 ret=-2 invalid arguments，
+    且旧代码把 ret=-2 误判为 token 失效，导致连接假性断开。先以纯文本链接形式发送，
+    确保用户可点击访问且不会破坏会话；待拿到 iLink 媒体协议文档后再恢复内联 item。
     """
     relay_url = _platform_relay_url(url)
     if media_type == "image":
-        media_item = {"type": 2, "image_item": {"url": relay_url}}
+        caption = "🖼️ 图片已生成，点击查看：" + relay_url
-        caption = "🖼️ 图片已生成（若未直接显示，请点此查看）：" + relay_url
     else:
-        media_item = {"type": 5, "video_item": {"url": relay_url}}
+        caption = "🎬 视频已生成，点击查看：" + relay_url
-        caption = "🎬 视频已生成（若未直接播放，请点此查看）：" + relay_url
+    return _wechat_send_message(baseurl, bot_token, to_user_id, context_token, caption)
-    return _wechat_post_msg(baseurl, bot_token, to_user_id, context_token,
-                             [media_item, {"type": 1, "text_item": {"text": caption}}])
 
 
 def _wechat_extract_text(msg: dict) -> str:
@@ -326,7 +353,6 @@ def _get_or_create_wechat_conv(db: Session, user_id: str) -> str:
     ), {"uid": user_id})
     if rows:
         return _row_to_dict(rows[0])["id"]
-    import uuid as _uuid
     cid = str(_uuid.uuid4())
     now = _beijing_now()
     _execute_raw(db, text(
@@ -338,7 +364,6 @@ def _get_or_create_wechat_conv(db: Session, user_id: str) -> str:
 
 
 def _append_ai_message(db: Session, conv_id: str, role: str, content: str, created=None):
-    import uuid as _uuid
     _execute_raw(db, text(
         "INSERT INTO messages (id, conversation_id, role, content, token_used, created_at) "
         "VALUES (:id, :cid, :role, :content, 0, :c)"
@@ -425,7 +450,14 @@ def _wechat_receiver_loop(account_id: int):
                 except Exception:
                     pass
                 # 先生成回复（此时历史尚未包含本条，避免重复）
+                import logging as _lrx, inspect as _insp
+                _lrx.warning(f'[receiver] calling _qihua_reply({repr(content)}, {repr(owner_uid)})')
+                # 验证使用的 _call_real_llm 版本
+                from app.api.im_llm import _call_real_llm as _crl
+                _src = _insp.getsource(_crl)
+                _lrx.warning(f'[receiver] _call_real_llm source len={len(_src)}, has _try_models={"_try_models" in _src}')
                 reply = _qihua_reply(content, owner_uid)
+                _lrx.warning(f'[receiver] reply ({type(reply).__name__}): {repr(reply)[:200]}')
                 # 媒体结构化返回：_call_real_llm 生成图片/视频时返回
                 # {"__media__": True, "media_type": "image"/"video", "url": "..."}
                 if isinstance(reply, dict) and reply.get("__media__"):
@@ -457,11 +489,23 @@ def _wechat_receiver_loop(account_id: int):
                                      % (ret, repr(sender)))
                     except Exception:
                         pass
-                    if ret in (-2, -14):
+                    if ret == -14:
-                        # 会话/鉴权失效：标记需重扫但继续循环，让下一轮 poll 自然检测过期
+                        # -14=会话过期，需重扫
                         _mark_wechat_expired(d["id"])
+                    elif ret == -2:
+                        # -2=参数错误（如媒体消息格式不被 iLink 接受），并非 token 失效。
+                        # 标记 expired 会导致连接假性断开，后续消息全部丢失，故仅记录。
+                        pass
             time.sleep(1)
         except Exception:
+            # 2026-07-31 修复：此前裸 except 静默吞掉所有异常，导致接收器出错时毫无日志、
+            # 完全无法排查。改为记录异常堆栈到磁盘，便于定位。
+            import traceback as _tb
+            try:
+                with open("/tmp/wechat_receiver_err.log", "a", encoding="utf-8") as _f:
+                    _f.write(_beijing_now().isoformat() + " RECEIVER EXCEPTION:\n" + _tb.format_exc() + "\n")
+            except Exception:
+                pass
             time.sleep(5)
         finally:
             db.close()
@@ -555,6 +599,13 @@ async def wechat_qrcode_status(
         data = j.get("data", j) if isinstance(j, dict) else j
         status = data.get("status")
         if status == "confirmed":
+            # 07-30 diagnostic: log full iLink confirm response
+            import json as _json
+            try:
+                with open("/tmp/wechat_confirm.log", "a", encoding="utf-8") as _f:
+                    _f.write(_beijing_now().isoformat() + " CONFIRM_RAW=" + _json.dumps(j, ensure_ascii=False)[:2000] + "\n")
+            except Exception:
+                pass
             creds = data.get("credentials", {}) or {}
             bot_token = data.get("bot_token") or creds.get("bot_token")
             baseurl = data.get("baseurl") or creds.get("baseurl") or ILINK_BASE
@@ -590,7 +641,15 @@ async def wechat_qrcode_status(
                         "updated_at": _beijing_now(), "id": _row_to_dict(row)["id"],
                         })
                 row = _wechat_row_for_user(db, user.id)
-                aid = _row_to_dict(row)["id"] if row else 1
+                saved = _row_to_dict(row) if row else {}
+                import json as _json2
+                try:
+                    with open("/tmp/wechat_confirm.log", "a", encoding="utf-8") as _f:
+                        _f.write(_beijing_now().isoformat() + " SAVED bot_token=" + repr(saved.get("bot_token", "NULL")[:20]) 
+                                 + " ilink_bot_id=" + repr(saved.get("ilink_bot_id")) + "\n")
+                except Exception:
+                    pass
+                aid = saved.get("id", 1)
                 _start_wechat_receiver(aid)
             finally:
                 db.close()

backend/python/app/api/im_llm.py

@@ -59,6 +59,17 @@ def _proxy_for_url(url: str) -> str | None:
     return os.environ.get("HTTPS_PROXY") or os.environ.get("HTTP_PROXY") or "http://172.18.0.1:7890"
 
 
+_DEBUG_LOG_PATH = "/tmp/im_llm_debug.log"
+
+def _debug_log(msg: str):
+    """写磁盘调试日志（不依赖 uvicorn logging 配置）。"""
+    try:
+        with open(_DEBUG_LOG_PATH, "a", encoding="utf-8") as _f:
+            _f.write(datetime.now(timezone(timedelta(hours=8))).isoformat() + " " + msg + "\n")
+    except Exception:
+        pass
+
+
 def _run_async(coro, timeout=None):
     """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。
 
@@ -126,7 +137,15 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
             _discover_installed_skills,
             _build_system_prompt,
             _execute_skill_tool,
+            _list_enabled_mcp_servers,
         )
+        # 2026-08-04 双 Claw 对齐：能力发现收敛为唯一实现，Web / IM 共用
+        from app.core.capability_discovery import (
+            build_mcp_tool_defs,
+            discover_mcp_servers,
+            discover_skills,
+        )
+        from app.mcp_client import call_mcp_tool
 
         # ===== 1) 媒体意图路由（Issue 7）=====
         # 2026-07-30 fix: 只对当前用户消息做媒体意图检测，避免历史中的图像请求污染
@@ -169,23 +188,59 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
         ai_model = resolve_auto_model("text", db)
         _try_models = [ai_model] if (ai_model and ai_model.is_enabled) else []
         # Auto 模式附加备选（IM Bot 始终 Auto）
+        # 2026-07-31 根因修复：仅选用「纯文本」模型作为 failover 备选。
+        # 此前按 is_enabled 顺序取前 4 个，会混入图像/视频/语音(ASR)模型
+        # （这些模型 capability 含 "text" 但本质是媒体模型，收到文本 prompt 返回空内容、
+        #  不抛异常 → 被误判为「模型失败」，最终导致「已尝试N个模型均调用失败」）。
+        _MEDIA_CAPS = {"image", "video", "audio", "asr", "embedding", "rerank", "ocr"}
+        def _is_text_only(m):
+            _caps = m.capabilities or []
+            return "text" in _caps and not any(c in _MEDIA_CAPS for c in _caps)
         _fallbacks = db.query(AIModel).filter(
             AIModel.is_enabled == True,
             AIModel.id != (ai_model.id if ai_model else ""),
-        ).order_by(AIModel.is_default.desc()).limit(4).all()
+        ).order_by(AIModel.is_default.desc()).all()
-        _try_models.extend([m for m in _fallbacks if m not in _try_models])
+        _fallbacks = [m for m in _fallbacks if _is_text_only(m)]
+        _try_models.extend([m for m in _fallbacks[:4] if m not in _try_models])
+        _debug_log(f"TRY_MODELS user={user_id} count={len(_try_models)} names={[m.name for m in _try_models]}")
 
         if not _try_models:
             return "（未配置可用文本模型，无法生成回复）"
 
-        # 技能自动发现（提前执行一次，不必每个候选模型重复做）
+        # 能力自动发现（提前执行一次，不必每个候选模型重复做）
+        # 2026-08-04 根因修复：
+        #   旧实现 `_discover_installed_skills(db, user_id)[:15]` 强依赖
+        #   app_installations 表，线上该表 0 行 → 97 个 published skill
+        #   在 Web / IM 双端同时发现不到；`[:15]` 截断的是一个空列表。
+        #   现改走统一发现层（安装优先 → published 降级 → 相关性裁剪）。
+        _query_hint = _cur_msg or prompt
         skills = []
-        if user_id:
         try:
-                skills = _discover_installed_skills(db, user_id)[:15]
+            skills = discover_skills(db, user_id, query=_query_hint, claw="im")
-            except Exception:
+        except Exception as _e:
+            logger.warning(f"[im_llm] 技能发现失败(降级为空): {_e}")
             skills = []
 
+        # MCP 自动挂载：IM Bot Claw 没有勾选 UI，自动发现即其等价通道。
+        # （Web Claw 靠用户在界面勾选 → effective_mcp_ids）
+        mcp_servers = []
+        mcp_function_definitions = []
+        mcp_tool_map = {}
+        try:
+            mcp_servers = discover_mcp_servers(db, None, claw="im")
+            if mcp_servers:
+                mcp_function_definitions, mcp_tool_map = _run_async(
+                    build_mcp_tool_defs(mcp_servers), timeout=45
+                )
+        except Exception as _e:
+            logger.warning(f"[im_llm] MCP 发现失败(降级为空): {_e}")
+            mcp_servers, mcp_function_definitions, mcp_tool_map = [], [], {}
+
+        _debug_log(
+            f"CAP skills={len(skills)} mcp_servers={len(mcp_servers)} "
+            f"mcp_tools={len(mcp_function_definitions)}"
+        )
+
         # 平台知识库注入（对齐 Web Chat——微信用户提问"平台有哪些技能"也能得到完整清单）
         kb_context = ""
         try:
@@ -213,13 +268,18 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                 _retry_log.append(f"{ai_model.name}: 缺少密钥或端点")
                 continue
 
+            # agents 暂留空：线上 market_apps 中 type='agent' 记录为 0，
+            # 且 IM 侧尚无 agent 执行分支；待薄壳化接入 EngineCore 后自动获得。
             system_prompt, function_definitions = _build_system_prompt(
                 skills=skills,
                 agents=[],
                 kb_context=kb_context,
                 user_content=prompt,
-                mcp=[],
+                mcp=mcp_servers,
+                mcp_functions=mcp_function_definitions,
                 auto_discovered=bool(skills),
+                all_mcp=_list_enabled_mcp_servers(db),
+                db=db,
             )
             messages = [
                 {"role": "system", "content": system_prompt},
@@ -233,7 +293,7 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                 _http_ok = True
                 for round_num in range(max_rounds):
                     request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048, "enable_thinking": False, "temperature": 0.7}
-                    if function_definitions and skills:
+                    if function_definitions:
                         request_body["tools"] = [
                             {"type": "function", "function": fd} for fd in function_definitions
                         ]
@@ -241,6 +301,7 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                     try:
                         import logging as _l5
                         _l5.warning(f'[im_llm REQ] {ai_model.name}: url={api_url[:60]}, proxy={_proxy_for_url(api_url)}, api_key={api_key[:20]}..., messages_count={len(request_body.get("messages", []))}, tools_count={len(request_body.get("tools", []))}')
+                        _debug_log(f"REQ model={ai_model.name} url={api_url} proxy={_proxy_for_url(api_url)} body={json.dumps(request_body, ensure_ascii=False)[:600]}")
                         data = chat_complete(
                             api_url,
                             api_key,
@@ -249,11 +310,13 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                             timeout=ai_model.timeout or 120,
                         )
                         _l5.warning(f'[im_llm RES] {ai_model.name}: data_keys={list(data.keys()) if data else None}, resp_type={type(data).__name__}')
+                        _debug_log(f"RES model={ai_model.name} data={json.dumps(data, ensure_ascii=False)[:600]}")
                     except Exception as _exc:
                         _retry_log.append(f"{ai_model.name}: {str(_exc)[:60]}")
                         import traceback as _tb
                         logger.warning(f"[im_llm] 模型 {ai_model.name} 异常: {_exc}")
                         logger.warning(f"[im_llm] {ai_model.name} 异常详情: {_tb.format_exc()[:2000]}")
+                        _debug_log(f"EXC model={ai_model.name} err={str(_exc)[:200]} trace={_tb.format_exc()[:500]}")
                         _http_ok = False
                         break
                     message = data.get("choices", [{}])[0].get("message", {})
@@ -265,10 +328,11 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                     _msg_keys = list(message.keys()) if message else []
                     _tool_calls = tool_calls
                     _lcr.warning(f'[CRITICAL] MODEL={ai_model.name} msg_keys={_msg_keys} content_len={len(_raw_content or "")} reasoning_len={len(_raw_reasoning or "")} has_tool_calls={bool(_tool_calls)}')
+                    _debug_log(f"PARSE model={ai_model.name} content_len={len(_raw_content or '')} reasoning_len={len(_raw_reasoning or '')} content_preview={repr(_raw_content)[:120]}")
                     ai_content = message.get("content", "") or message.get("reasoning_content") or ""
                     if not ai_content and (message.get("reasoning_content") or ""):
                         ai_content = f"（AI 思考：{message.get('reasoning_content')[:500]}...）"
-                    if tool_calls and skills:
+                    if tool_calls and (skills or mcp_tool_map):
                         messages.append({
                             "role": "assistant",
                             "content": message.get("content") or message.get("reasoning_content") or "",
@@ -280,6 +344,39 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                                 func_args = json.loads(tc.get("function", {}).get("arguments", "{}"))
                             except Exception:
                                 func_args = {"input": tc.get("function", {}).get("arguments", "")}
+                            if func_name.startswith("mcp__"):
+                                # MCP 工具：命名规则与 Web Claw 完全一致
+                                # mcp__{server}__{tool}__{idx}
+                                _info = mcp_tool_map.get(func_name)
+                                if not _info:
+                                    tool_result = (
+                                        "[MCP 工具未找到] 可用工具: "
+                                        + str(list(mcp_tool_map.keys())[:5])
+                                    )
+                                else:
+                                    _srv = next(
+                                        (s for s in mcp_servers
+                                         if s.get("id") == _info.get("server_id")),
+                                        None,
+                                    )
+                                    if not _srv:
+                                        tool_result = f"[MCP 服务未找到: {_info.get('server_name')}]"
+                                    else:
+                                        try:
+                                            tool_result = _run_async(
+                                                call_mcp_tool(_srv, _info["name"], func_args),
+                                                timeout=90,
+                                            )
+                                        except Exception as _mexc:
+                                            tool_result = (
+                                                f"[MCP 工具 {_info.get('name')} 执行失败："
+                                                f"{str(_mexc)[:160]}]"
+                                            )
+                                _debug_log(
+                                    f"MCP_CALL name={func_name} "
+                                    f"result_len={len(tool_result or '')}"
+                                )
+                            else:
                                 tool_result = _execute_skill_tool(db, func_name, func_args, skills, prompt)
                             messages.append({
                                 "role": "tool",
@@ -294,20 +391,25 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                 # 检查本轮调用是否成功
                 import logging as _log2
                 _log2.warning(f'[im_llm trace] {ai_model.name}: http_ok={_http_ok}, ai_content_len={len(ai_content) if ai_content else 0}, ai_content_preview={repr(ai_content)[:200]}')
+                _debug_log(f"CHECK model={ai_model.name} http_ok={_http_ok} ai_content_len={len(ai_content) if ai_content else 0}")
                 if _http_ok and ai_content:
+                    _debug_log(f"SUCCESS model={ai_model.name} returning len={len(ai_content)}")
                     return ai_content or "（AI 回复为空）"
                 # 本轮失败：记录日志，切换备选
                 logger.warning('[im_llm] 模型 %s 失败 (http_ok=%s, content=%s), 切换备选...' % (
                     ai_model.name, _http_ok, repr(ai_content)[:100]
                 ))
+                _debug_log(f"FAIL model={ai_model.name} http_ok={_http_ok} content_preview={repr(ai_content)[:100]}")
                 continue
             except Exception as exc:
                 _retry_log.append(f"{ai_model.name}: {str(exc)[:60]}")
                 logger.warning(f"[im_llm] 模型 {ai_model.name} 异常: {exc}")
+                _debug_log(f"OUTER_EXC model={ai_model.name} err={str(exc)[:120]}")
                 continue
 
         # 全部失败
         _summary = "; ".join(_retry_log[-3:])
+        _debug_log(f"ALL_FAILED try_models={len(_try_models)} retry_log={_retry_log}")
         return f"（已尝试 {len(_try_models)} 个模型均调用失败。最后错误: {_summary}。请稍后重试。）"
     except Exception as e:
         return f"（AI 回复生成失败：{e}）"

backend/python/app/core/capability_discovery.py

@@ -0,0 +1,277 @@
+"""统一能力发现层（双 Claw 共用 · 单一实现）。
+
+背景
+----
+七虎平台对外有两个 Claw 端口：
+  * Web Claw    —— SSE 流式，用户可在 UI 勾选 skills / mcp / agents
+  * IM Bot Claw —— 微信等 IM，server push 整条回复，**没有任何勾选 UI**
+
+历史问题（2026-08-04 线上实证）：
+  1. `im_llm.py` 写死 `agents=[]` / `mcp=[]`，IM 端永远拿不到 MCP 工具；
+  2. `_discover_installed_skills()` 强依赖 `app_installations` 表，而线上该表
+     **0 行** → 97 个 published skill 在 Web / IM **双端同时发现不到**；
+  3. IM 端 skills 还被 `[:15]` 截断（截断一个空列表，属噪声）。
+
+设计原则
+--------
+能力发现只允许有 **一份实现**（本模块）。任何 Claw 想要"技能/工具"，
+都必须调用这里的函数；禁止在端侧另写一套发现逻辑——历史上 #866 部署
+就是因为存在第二份实现而把 #764/#765 的注入改动整体覆盖回退。
+
+发现策略（三级降级，双端一致）
+--------------------------
+  1) 用户显式安装（app_installations）—— 最高优先，尊重用户意愿
+  2) 无安装记录 → 降级为平台已发布（status='published'）全集
+  3) 池子超过预算 → 按当前问题做轻量相关性排序，取 top-K 控制 token
+
+运维旋钮（环境变量，改完重启生效）
+--------------------------------
+  CAP_SKILL_LIMIT               单轮注入技能数上限，默认 12
+  CAP_MCP_LIMIT                 单轮注入 MCP 服务数上限，默认 8
+  CAP_SKILL_FALLBACK_PUBLISHED  安装表为空时是否降级为 published，默认 开
+  CAP_MCP_AUTO                  未显式指定时是否自动挂载已启用 MCP，默认 开
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import re
+from typing import Any, Iterable
+
+logger = logging.getLogger(__name__)
+
+__all__ = [
+    "discover_skills",
+    "discover_mcp_servers",
+    "build_mcp_tool_defs",
+    "rank_by_query",
+]
+
+
+# --------------------------------------------------------------------------
+# 环境旋钮
+# --------------------------------------------------------------------------
+def _env_int(name: str, default: int) -> int:
+    try:
+        raw = (os.environ.get(name) or "").strip()
+        return int(raw) if raw else default
+    except Exception:
+        return default
+
+
+def _env_flag(name: str, default: bool = True) -> bool:
+    raw = (os.environ.get(name) or "").strip().lower()
+    if not raw:
+        return default
+    return raw in ("1", "true", "yes", "on")
+
+
+# --------------------------------------------------------------------------
+# 轻量相关性排序（中文友好：整词 + 2-gram）
+# --------------------------------------------------------------------------
+def _tokens(text: str) -> set:
+    text = re.sub(r"[^\w\u4e00-\u9fff]+", " ", (text or "").lower())
+    out: set = set()
+    for w in text.split():
+        if len(w) >= 2:
+            out.add(w)
+        for i in range(len(w) - 1):
+            out.add(w[i : i + 2])
+    return out
+
+
+def rank_by_query(items: list, query: str, top_k: int) -> list:
+    """按与 query 的词元重叠度排序取前 top_k。
+
+    池子本来就不超预算时原样返回，避免无谓抖动（同一问题两次调用结果稳定）。
+    """
+    if not items:
+        return []
+    if top_k <= 0 or len(items) <= top_k:
+        return items[:top_k] if top_k > 0 else []
+    q = _tokens(query)
+    if not q:
+        return items[:top_k]
+
+    scored = []
+    for idx, it in enumerate(items):
+        blob = " ".join(
+            str(it.get(k) or "")
+            for k in ("name", "description", "category", "func_name")
+        )
+        t = _tokens(blob)
+        if t:
+            # 交集数按文档长度做平方根归一，抑制"长描述天然占优"
+            score = len(q & t) / (len(t) ** 0.5)
+        else:
+            score = 0.0
+        # idx 作为稳定次序的 tie-breaker，保证同分时结果可复现
+        scored.append((-score, idx, it))
+    scored.sort(key=lambda x: (x[0], x[1]))
+    return [it for _, _, it in scored[:top_k]]
+
+
+# --------------------------------------------------------------------------
+# 技能发现
+# --------------------------------------------------------------------------
+def discover_skills(
+    db,
+    user_id: str | None,
+    query: str = "",
+    limit: int | None = None,
+    claw: str = "web",
+) -> list:
+    """发现当前请求可用的技能列表（双 Claw 共用）。
+
+    返回 `_resolve_market_skills` 的结构：
+        [{id, name, description, skill_content, category, func_name, parameters}]
+    """
+    limit = limit if limit is not None else _env_int("CAP_SKILL_LIMIT", 12)
+    if limit <= 0:
+        return []
+
+    # 延迟导入：避免 app.core ←→ app.api 循环依赖
+    from app.api.chat_skill import _resolve_market_skills
+    from app.models import AppInstallation, MarketApp
+
+    pool: list = []
+    source = "empty"
+
+    # 1) 用户显式安装
+    if user_id:
+        try:
+            app_ids = [
+                i.app_id
+                for i in db.query(AppInstallation)
+                .filter(AppInstallation.user_id == user_id)
+                .all()
+            ]
+            if app_ids:
+                ids = [
+                    a.id
+                    for a in db.query(MarketApp)
+                    .filter(
+                        MarketApp.id.in_(app_ids),
+                        MarketApp.type == "skill",
+                        MarketApp.status == "published",
+                    )
+                    .all()
+                ]
+                if ids:
+                    pool = _resolve_market_skills(db, ids, user_id)
+                    source = "installed"
+        except Exception as e:
+            logger.warning("[cap] 安装技能发现失败(降级继续): %s", e)
+
+    # 2) 降级：平台已发布全集
+    if not pool and _env_flag("CAP_SKILL_FALLBACK_PUBLISHED", True):
+        try:
+            ids = [
+                a.id
+                for a in db.query(MarketApp)
+                .filter(MarketApp.type == "skill", MarketApp.status == "published")
+                .all()
+            ]
+            if ids:
+                pool = _resolve_market_skills(db, ids, user_id or "")
+                source = "published"
+        except Exception as e:
+            logger.warning("[cap] published 技能发现失败: %s", e)
+
+    # 3) 预算裁剪
+    picked = rank_by_query(pool, query, limit)
+    logger.info(
+        "[cap] skills claw=%s source=%s pool=%d picked=%d",
+        claw,
+        source,
+        len(pool),
+        len(picked),
+    )
+    return picked
+
+
+# --------------------------------------------------------------------------
+# MCP 发现
+# --------------------------------------------------------------------------
+def discover_mcp_servers(
+    db,
+    explicit_ids: Iterable[str] | None = None,
+    limit: int | None = None,
+    claw: str = "web",
+) -> list:
+    """发现当前请求可用的 MCP 服务连接（双 Claw 共用）。
+
+    Web Claw 传 explicit_ids（用户在 UI 勾选的）；
+    IM Bot Claw 无勾选通道 → 传 None，自动挂载平台已启用的 MCP。
+    返回 [{id, name, description, transport, endpoint, url, headers}]
+    """
+    limit = limit if limit is not None else _env_int("CAP_MCP_LIMIT", 8)
+    if limit <= 0:
+        return []
+
+    from app.api.chat_mcp import _list_enabled_mcp_servers, _resolve_active_mcp_servers
+
+    ids = [i for i in (explicit_ids or []) if i]
+    if ids:
+        out = _resolve_active_mcp_servers(db, ids)[:limit]
+        logger.info("[cap] mcp claw=%s source=explicit picked=%d", claw, len(out))
+        return out
+
+    if not _env_flag("CAP_MCP_AUTO", True):
+        return []
+
+    out = _list_enabled_mcp_servers(db)[:limit]
+    logger.info("[cap] mcp claw=%s source=auto picked=%d", claw, len(out))
+    return out
+
+
+# --------------------------------------------------------------------------
+# MCP → function definitions
+# --------------------------------------------------------------------------
+def _safe_name(raw: str, maxlen: int) -> str:
+    return re.sub(r"[^A-Za-z0-9_-]", "_", (raw or "")[:maxlen])
+
+
+async def build_mcp_tool_defs(servers: list, idx_start: int = 0) -> tuple:
+    """把 MCP 服务列表展开为 (function_definitions, tool_map)。
+
+    命名规则与 Web Claw 现有实现严格一致：`mcp__{server}__{tool}__{idx}`，
+    保证两端产生的工具名可互相对照、可被同一套审批/统计逻辑识别。
+
+    返回:
+        function_definitions: [{name, description, parameters}]
+        tool_map:             {func_name: {server_id, name, server_name}}
+    """
+    from app.mcp_client import list_mcp_tools
+
+    defs: list = []
+    tool_map: dict = {}
+
+    for offset, srv in enumerate(servers or []):
+        idx = idx_start + offset
+        try:
+            tools = await list_mcp_tools(srv)
+        except Exception as e:
+            logger.warning("[cap] MCP 拉取工具失败 [%s]: %s", srv.get("name"), e)
+            continue
+        for t in tools or []:
+            sname = _safe_name(srv.get("name") or "unknown", 30)
+            tname = _safe_name(t.get("name") or "tool", 50)
+            fname = f"mcp__{sname}__{tname}__{idx}"
+            tool_map[fname] = {
+                "server_id": srv.get("id"),
+                "name": t.get("name"),
+                "server_name": srv.get("name"),
+            }
+            defs.append(
+                {
+                    "name": fname,
+                    "description": (
+                        f"[MCP:{srv.get('name')}] {t.get('description') or t.get('name') or ''}"
+                    )[:900],
+                    "parameters": t.get("inputSchema")
+                    or {"type": "object", "properties": {}},
+                }
+            )
+    return defs, tool_map

backend/python/app/core/llm_client.py

@@ -83,6 +83,11 @@ def _create_kwargs(request_body: dict, stream: bool = False) -> dict:
     if tools:
         kw["tools"] = tools
         kw["tool_choice"] = request_body.get("tool_choice", "auto")
+    # 2026-07-30: 将未在白名单的额外字段（如 enable_thinking）作为 extra_body 传递
+    _whitelist = {"model", "messages", "stream", "temperature", "top_p", "max_tokens", "stop", "tools", "tool_choice"}
+    _extras = {k: v for k, v in request_body.items() if k not in _whitelist}
+    if _extras:
+        kw["extra_body"] = _extras
     return kw
 
 
@@ -223,29 +228,85 @@ def build_openai_client_sync(api_url, api_key, *, proxy=None, timeout=120, auth_
     return OpenAI(base_url=base, api_key=api_key or "EMPTY", timeout=timeout, http_client=http_client, default_headers=_hdrs if _hdrs else None)
 
 
-def chat_complete(api_url, api_key, request_body, *, proxy=None, timeout=120):
+def chat_complete(api_url, api_key, request_body, *, proxy=None, timeout=120, auth_header=None):
-    """同步版（与 achat_complete 行为/返回形状完全一致），供非 async 上下文使用。"""
+    """同步版（与 achat_complete 行为/返回形状完全一致），供非 async 上下文使用。
-    client = build_openai_client_sync(api_url, api_key, proxy=proxy, timeout=timeout, auth_header=auth_header)
+
-    with client:
+    2026-07-30 重写：直接使用 requests 替代 OpenAI SDK/httpx。
-        resp = client.chat.completions.create(**_create_kwargs(request_body))
+    OpenAI SDK 的 httpx 连接池在慢长轮询（如 iLink QR 状态 12-19s）后会出现
-        msg = resp.choices[0].message
+    Connection error，已实测复现。requests 库没有这个问题。
-        return {
+    """
-            "choices": [
+    import requests
-                {
+    import base64
+    import random
+    body = dict(request_body)
+    body["stream"] = False
+    # enable_thinking 等额外字段已在 request_body 中
+
+    # Headers: Bearer token + X-WECHAT-UIN (兼容 iLink)
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": f"Bearer {api_key}" if api_key else "Bearer empty",
+    }
+    if auth_header and ":" in auth_header:
+        _k, _v = auth_header.split(":", 1)
+        headers[_k.strip()] = _v.strip()
+    else:
+        # X-WECHAT-UIN 防重放
+        headers["X-WECHAT-UIN"] = base64.b64encode(str(random.randint(0, 2**32-1)).encode()).decode()
+
+    # 2026-07-31 修复根因: proxy 为 None 时传 proxies={} 而非 None。
+    # requests 在 proxies=None 时会继续读取环境 HTTP_PROXY(trust_env 默认 True),
+    # 而 uvicorn 进程(run.sh)带有 HTTP_PROXY=172.18.0.1:7890, 该代理无法解析
+    # Docker 内部域名(如 qihu-aiflow-litellm), 导致所有内部模型调用静默失败、返回空。
+    # 显式 proxies={} + trust_env=False 可完全绕过环境代理, 使 _proxy_for_url 的
+    # 决策(返回 None=不代理)真正生效。
+    _proxies = {"http": proxy, "https": proxy} if proxy else {}
+    _session = requests.Session()
+    _session.trust_env = False
+    last_body = {}
+    for attempt in range(2):
+        try:
+            resp = _session.post(
+                api_url.rstrip("/"),
+                headers=headers,
+                json=body,
+                proxies=_proxies,
+                timeout=timeout,
+            )
+            try:
+                last_body = resp.json() if resp.content else {}
+            except Exception:
+                last_body = {}
+            if resp.ok:
+                import logging as _l6
+                _l6.warning(f'[chat_complete] body keys={list(last_body.keys()) if isinstance(last_body, dict) else type(last_body)}, content_len={len(resp.content)}, content_preview={str(resp.content)[:200]}')
+                return _shape_response(last_body)
+        except Exception:
+            pass
+    return {"choices": [{"message": {"content": "", "tool_calls": None, "reasoning_content": None}}]}
+
+
+def _shape_response(body: dict) -> dict:
+    """Normalize OpenAI-compatible response to standard shape."""
+    choices = body.get("choices", [])
+    import logging as _l7
+    if choices:
+        msg = choices[0].get("message", {})
+        result = {
+            "choices": [{
                 "message": {
-                        "content": msg.content or "",
+                    "content": msg.get("content") or "",
-                        "tool_calls": _normalize_tool_calls(getattr(msg, "tool_calls", None)),
+                    "tool_calls": _normalize_tool_calls(msg.get("tool_calls")),
-                        "reasoning_content": getattr(msg, "reasoning_content", None),
+                    "reasoning_content": msg.get("reasoning_content") or "",
                 }
+            }]
         }
-            ],
+        _l7.warning(f'[_shape] content_len={len(result["choices"][0]["message"]["content"])}, keys={list(result["choices"][0]["message"].keys())}')
-            "usage": (resp.usage.model_dump() if resp.usage else {}),
+        return result
-        }
+    _l7.warning(f'[_shape] NO choices in body, keys={list(body.keys())[:5]}')
+    return body
 
 
-# ---------------------------------------------------------------------------
-# 嵌入（Embeddings）：OpenAI 兼容 /v1/embeddings，收敛散落的裸 httpx 嵌入调用
-# ---------------------------------------------------------------------------
 def _embed_kwargs(input, model):
     return {"model": model, "input": input if isinstance(input, list) else [input]}
 

backend/python/app/mcp_client.py

@@ -20,6 +20,7 @@ from typing import Optional
 from mcp import ClientSession
 from mcp.client.sse import sse_client
 from mcp.client.streamable_http import streamablehttp_client
+import httpx
 
 logger = logging.getLogger(__name__)
 
@@ -28,18 +29,38 @@ _TOOLS_CACHE: dict = {}
 _TOOLS_CACHE_TTL = 120  # 秒
 
 
+def _no_proxy_httpx_client(headers=None, timeout=None, auth=None):
+    """MCP 传输用 httpx 客户端工厂：强制直连、不走环境代理。
+
+    容器 run.sh 注入了 HTTP_PROXY=http://172.18.0.1:7890，该代理无法解析
+    外部 MCP 域名（xiaobenyang.com / anysearch.com 等），且默认 trust_env
+    会读取 HTTP_PROXY 导致 tools/list 全部静默失败。这里 trust_env=False
+    直连，与 app/core/llm_client.py:chat_complete 修复同源。
+    """
+    kwargs: dict = {"follow_redirects": True, "trust_env": False}
+    kwargs["timeout"] = timeout if timeout is not None else httpx.Timeout(30.0, read=300.0)
+    if headers is not None:
+        kwargs["headers"] = headers
+    if auth is not None:
+        kwargs["auth"] = auth
+    return httpx.AsyncClient(**kwargs)
+
+
 def _transport_cm(server: dict, timeout: float):
     """按 server['transport_type'] 选择官方 SDK 传输层。
 
     返回异步上下文管理器，进入后产出 (read_stream, write_stream, ...)。
+    关键：注入 _no_proxy_httpx_client，使 MCP 流量直连（见其 docstring）。
     """
     url = (server.get("url") or "").strip()
     ttype = (server.get("transport_type") or "streamable_http").lower()
     headers = server.get("headers") or {}
     if ttype == "sse":
-        return sse_client(url, headers=headers, timeout=timeout)
+        return sse_client(url, headers=headers, timeout=timeout,
+                          httpx_client_factory=_no_proxy_httpx_client)
     # 默认 streamable_http（官方 SDK 标准传输）
-    return streamablehttp_client(url, headers=headers, timeout=timeout)
+    return streamablehttp_client(url, headers=headers, timeout=timeout,
+                                 httpx_client_factory=_no_proxy_httpx_client)
 
 
 async def _session_request(