261 lines
8.5 KiB
Python
261 lines
8.5 KiB
Python
import os
|
||
import re
|
||
import asyncio
|
||
import yaml
|
||
from functools import partial
|
||
import hashlib
|
||
from pathlib import Path
|
||
from datetime import datetime
|
||
from typing import List, Dict, Any
|
||
from ahserver.serverenv import ServerEnv
|
||
from appPublic.dictObject import DictObject
|
||
from appPublic.log import info, debug, error, exception
|
||
from appPublic.jsonConfig import getConfig
|
||
from appPublic.uniqueID import getID
|
||
|
||
# 配置审计日志
|
||
|
||
class LLMHandler:
|
||
def __init__(self, request, llmid, apikey=None):
|
||
self.llmid = llmid
|
||
self.apikey = apikey
|
||
self.request = request
|
||
|
||
async def __call__(self, prompt):
|
||
env = ServerEnv()
|
||
kw = {
|
||
"stream": False,
|
||
"llmid": self.llmid,
|
||
"prompt": prompt
|
||
}
|
||
kw = DictObject(**kw)
|
||
txt = ''
|
||
async for d in env.inference_generator(self, request, params_kw=kw):
|
||
debug(f'{d=}, {type(d)=}')
|
||
txt += d.content
|
||
return txt
|
||
|
||
async def run_subprocess(command, cwd, env, timeout=30.0):
|
||
try:
|
||
# 启动子进程
|
||
# stdout/stderr 设置为 PIPE 对应原来的 capture_output=True
|
||
process = await asyncio.create_subprocess_shell(
|
||
command,
|
||
stdout=asyncio.subprocess.PIPE,
|
||
stderr=asyncio.subprocess.PIPE,
|
||
cwd=cwd,
|
||
env=env,
|
||
# 注意：asyncio 默认处理字节流，如果需要文本，后面需手动 decode
|
||
)
|
||
|
||
# 使用 wait_for 实现 timeout 功能
|
||
try:
|
||
# wait_for 会在超时后抛出 TimeoutError 并自动 cancel 协程
|
||
stdout, stderr = await asyncio.wait_for(process.communicate(), timeout=timeout)
|
||
except asyncio.TimeoutError:
|
||
process.kill() # 超时强制杀掉进程
|
||
await process.wait() # 确保资源回收
|
||
raise Exception(f"{command} 执行超时")
|
||
|
||
# 对应 text=True，手动解码
|
||
return {
|
||
"stdout": stdout.decode().strip(),
|
||
"stderr": stderr.decode().strip(),
|
||
"return_code": process.returncode
|
||
}
|
||
|
||
except Exception as e:
|
||
print(f"{command} 执行异常: {e}")
|
||
return None
|
||
|
||
class IndustrialSkillEngine:
|
||
def __init__(self, request, llmid: str, apikey: str=None):
|
||
self.request = request
|
||
config = getConfig()
|
||
skills_dir = config.skills_dir
|
||
self.root = Path(skills_dir).resolve()
|
||
self.llm = LLMHandler(request, llmid, apikey=apikey)
|
||
self.registry = {}
|
||
self.task_queue = asyncio.Queue(maxsize=20)
|
||
# 状态机：记录当前任务执行到的步骤
|
||
self.state = {"current_skill": None, "history": [], "pending_params": []}
|
||
|
||
async def write_output(self, data=None):
|
||
await self.task_queue.put(data)
|
||
|
||
# --- 1. 工业级初始化：依赖检查与索引 ---
|
||
async def boot(self, refresh=False):
|
||
env = self.request._run_ns
|
||
userid = await env.get_user()
|
||
key = f'skillregister_{userid}'
|
||
skills = await env.session_getvalue(key)
|
||
if not refresh and skills:
|
||
self.self.registry = skills
|
||
return
|
||
|
||
for skill_md in self.root.glob("**/SKILL.md"):
|
||
with open(skill_md, 'r', encoding='utf-8') as f:
|
||
content = f.read()
|
||
meta = yaml.safe_load(re.search(r'^---(.*?)---', content, re.DOTALL).group(1))
|
||
name = meta.get('name')
|
||
|
||
# 检查依赖 (Pre-flight check)
|
||
req_file = skill_md.parent / "requirements.txt"
|
||
has_deps = req_file.exists()
|
||
|
||
self.registry[name] = DictObject(**{
|
||
"root": skill_md.parent,
|
||
"meta": meta,
|
||
"content": content,
|
||
"has_deps": has_deps
|
||
})
|
||
await env.session_setvalue(key, self.registry)
|
||
|
||
# --- 2. 自动化依赖环境隔离 (venv 思想) ---
|
||
async def _ensure_dependencies(self, skill_name: str):
|
||
skill = self.registry[skill_name]
|
||
if skill["has_deps"]:
|
||
# 工业级引擎通常会检查一个隐藏的 .installed 标识
|
||
if not (skill["root"] / ".deps_installed").exists():
|
||
print(f"📦 正在为技能 {skill_name} 安装必要依赖...")
|
||
await run_subprocess(["pip", "install", "-r", "requirements.txt"], cwd=skill["root"])
|
||
(skill["root"] / ".deps_installed").touch()
|
||
|
||
# --- 3. 增强版安全执行器：带重试逻辑与审计 ---
|
||
async def _execute_with_retry(self, command: str, skill_name: str, retry_count=1) -> str:
|
||
await self._ensure_dependencies(skill_name)
|
||
|
||
# 安全预检：禁止敏感命令
|
||
forbidden = ["rm ", "> /dev/", "chmod", "sudo"]
|
||
if any(f in command for f in forbidden):
|
||
return "🚫 安全风险：检测到非法指令，执行被拦截。"
|
||
|
||
info(f"Executing: {command} in {skill_name}")
|
||
|
||
try:
|
||
env = os.environ.copy()
|
||
env["SKILL_CONTEXT"] = skill_name
|
||
|
||
res = await run_subprocess(command, cwd=self.registry[skill_name]["root"], env=env, timeout=30)
|
||
|
||
if res.return_code != 0:
|
||
# 工业级特性：自动将错误回传给 LLM 进行自愈 (Self-healing)
|
||
error(f"Command failed: {res.stderr}")
|
||
if retry_count > 0:
|
||
print(f"⚠️ 执行失败，尝试让 AI 自愈修复参数...")
|
||
new_prompt = f"命令 '{command}' 失败，错误信息: {res.stderr}。请根据错误重新生成正确的命令，或提示用户补全参数。"
|
||
# 这里会递归调用逻辑进行修复
|
||
return await self._run(new_prompt, is_retry=True)
|
||
|
||
raise Exception(f"Error: {res.stderr}")
|
||
|
||
return res.stdout
|
||
except Exception as e:
|
||
return str(e)
|
||
|
||
# --- 4. 递归文档注入与意图路由 ---
|
||
async def _get_expanded_context(self, skill_name: str, user_prompt: str, context=None):
|
||
skill = self.registry[skill_name]
|
||
base_content = skill["content"]
|
||
|
||
# 扫描目录下的辅助文件夹
|
||
sub_dirs = ["reference", "examples"]
|
||
found_docs = []
|
||
for d in sub_dirs:
|
||
path = skill["root"] / d
|
||
if path.exists():
|
||
found_docs.extend([f"{d}/{f.name}" for f in path.glob("*.md")])
|
||
|
||
if found_docs:
|
||
# 仅在模型认为有必要时，“点餐式”加载
|
||
choice = await self.llm(f"上下文：{context or ''}\n用户问题: {user_prompt}\n可选深入文档: {found_docs}\n需要读取哪个？(仅返回路径，不需则返回 None)")
|
||
if choice != "None" and any(choice in doc for doc in found_docs):
|
||
with open(skill["root"] / choice, 'r') as f:
|
||
output = {
|
||
"status": "PROCESSING",
|
||
"hint": f"📂 深度加载: {choice}"
|
||
}
|
||
await self.write_output(output)
|
||
base_content += f"\n\n--- 深度参考 ({choice}) ---\n{f.read()}"
|
||
|
||
return base_content
|
||
|
||
async def reference(self, user_prompt:str, context: str=None, is_retry: bool=False):
|
||
f = partial(self.run, user_prompt, context=context, is_retry=is_retry)
|
||
asyncio.create_task(f())
|
||
while True:
|
||
data = await self.task_queue.get()
|
||
if not data:
|
||
break;
|
||
debug(f'{data=}, {type(data)=}')
|
||
yield data
|
||
await asyncio.sleep(0.1)
|
||
|
||
|
||
# --- 5. 主运行接口 ---
|
||
async def run(self, parmas_kw):
|
||
try:
|
||
user_input = json.dumps(params_kw, ensure_ascii=False)
|
||
await self._run(user_input)
|
||
except Exception as e:
|
||
await self.write_output({
|
||
'status': 'FAILED',
|
||
'error': f"{e}"
|
||
})
|
||
self.write_output(None)
|
||
|
||
async def _run(self, user_prompt: str, context=None, is_retry=False):
|
||
# 如果是重试，跳过技能选择
|
||
await self.boot()
|
||
debug(f'{self.registry=}')
|
||
if not is_retry:
|
||
await self.write_output({
|
||
"status": "PROCESSING",
|
||
"hint": "寻找合适的skill"
|
||
})
|
||
skill_map = {n: v.meta.description for n, v in self.registry.items()}
|
||
target = await self.llm(f"用户意图: {user_prompt}\n可选技能清单: {skill_map}\n请返回匹配的技能名:")
|
||
self.state["current_skill"] = target
|
||
await self.write_output({
|
||
"status":"PROCESSING",
|
||
"hint": f"找到skill={target}"
|
||
})
|
||
|
||
skill_name = self.state["current_skill"]
|
||
if skill_name not in self.registry:
|
||
raise Exception(f"技能名{skill_name}未注册")
|
||
|
||
# 获取递归上下文
|
||
context = await self._get_expanded_context(skill_name, user_prompt, context=context)
|
||
|
||
# 决策：是直接回答还是执行脚本
|
||
decision = await self.llm(f"上下文: {context}\n问题: {user_prompt}\n决定动作：EXEC: <command> 或 ANSWER: <text> 或 REPLY: <question>")
|
||
|
||
if "REPLY" in decision:
|
||
sessionkey = getID()
|
||
await self.write_output({
|
||
"status": "REPLY",
|
||
"reply": {
|
||
"questionkey": sessionkey(),
|
||
"question": decision.split("REPLY:")[1].strip()
|
||
}
|
||
})
|
||
env = self.request._run_ns
|
||
user_reply = await env.session_getvalue(sessionkey)
|
||
while not user_reply:
|
||
asyncio.sleep(0.5)
|
||
user_reply = await env.session_getvalue(sessionkey)
|
||
prompt = f"{user_prompt}\n补充输入：{user_reply}"
|
||
await self._run(prompt, context=context, is_retry=True)
|
||
return
|
||
if "EXEC:" in decision:
|
||
cmd = decision.split("EXEC:")[1].strip()
|
||
output = await self._execute_with_retry(cmd, skill_name)
|
||
await self.write_output(output)
|
||
return
|
||
if "ANSWER:" in decision:
|
||
output = decision.replace("ANSWER:", "").strip()
|
||
await self.write_output(output)
|
||
return output
|
||
debug(f' undefined decision:{decision}')
|