QihuAI Design
29a1ce2e54
- 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）
278 lines
9.6 KiB
Python
278 lines
9.6 KiB
Python
"""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
|