mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-17 04:11:57 +00:00
03c18c4d78
一、文案(前端约 1600 处,后端传到界面的约 880 处) - 术语统一:工作流层面用「运行」(继续运行、重新运行、试运行),节点和工具层面用「执行」,不再用「跑」; 总时长 / 执行时长 / 等待审批;服务端;助手;Agent;token;工作流;固定版本;模型接入;系统提示;批准 / 驳回;受管级别 - 口语改书面:取不到、没取回来、读不懂、连不上、多半、照样、兜底、揉成这类说法改成「无法 / 未能 / 失败 / 请」;去掉拟人写法 - 不再露出实现细节:界面上不写 checkpoint、同一拍、哈希(凭证处除外),不写改版历史(「以前…」「老数据」); 报错、校验、门禁、自动修复里裸写的配置键名和枚举值换成界面上的中文标签,由新的 backend/app/engine/labels.py 统一提供; 发布弹窗里结论句的选项不再是 off / require_citation - 同一件事只留一种说法:生成阶段文案、出具档位说明、画布锁定提示各合成一份 - 去掉界面上的工业场景示例(MES、ERP、「张工」「王工」)和真实产品名,换成通用示例 二、说错的提示(都有测试证明) - 数据库测连接失败时的专用提示以前从不出现(按报错文字全等比较,永远比不上):改为按错误类型判断,不再拼出重复的前缀 - 「设置 → 供应商」改指「设置 → 模型接入」;「在设置里建数据源」改指「数据」页 - 缺少 API Key、模型接入已停用的报错不再夹英文,写明去哪里填 - 两处原因说错(「这种引用在后续版本支持」「这一版还不支持」)改成真实原因 - 节点说明里不准确的说法改正:代码节点在 microVM 或系统沙箱中执行;报告撰写、口径卡的说明不再说过头 - 导出的证据 CSV 里两列同名,第三列改为「片段编号」 三、前后端对齐 - 前端按后端原文识别日志、报错的地方,新旧两种写法都认得(历史运行里存的是旧文字),逐条补了新旧原文的核对 - 后端交回模型的原指令放进单独的 for_model 字段,前端在解析响应时去掉,不会显示到界面上 四、检查 - 检查脚本里按原文匹配的断言改成新文案;改完会变成永远成立的反向断言逐条修正,并做了变异验证 - 新增 check-copy(静态扫描前端界面字符串和后端上界面的文案):再出现「跑」「后端」「多半」「令牌」、 界面上的 Copilot、裸写的配置键名、半角标点等就报错;确需保留的写进 scripts/copy-allow.json 并注明理由
180 lines
7.1 KiB
Python
180 lines
7.1 KiB
Python
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import logging
|
||
import os
|
||
import resource
|
||
import shutil
|
||
import sys
|
||
import time
|
||
from pathlib import Path
|
||
|
||
from app.core.config import settings, sanitize_session
|
||
from app.core.errors import describe_exception, raw
|
||
from app.sandbox import interpreter
|
||
from app.sandbox.base import ENTRY_PREFIX, ExecResult, Sandbox, SandboxLimits, entry_name, truncate
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
# 入口脚本后缀 + 执行命令。文件名运行时生成(entry_name),不能写死:
|
||
# 同会话并发执行都写 main.py 会互相覆盖。
|
||
_LANG = {
|
||
"python": (".py", None), # 运行时用 interpreter.python_argv()
|
||
"bash": (".sh", ["bash"]),
|
||
"sh": (".sh", ["sh"]),
|
||
"node": (".js", ["node"]),
|
||
"javascript": (".js", ["node"]),
|
||
}
|
||
|
||
|
||
class LocalSandbox(Sandbox):
|
||
"""裸子进程执行,所有隔离后端都不可用时的最后兜底。
|
||
|
||
只有 rlimit 和一个独立工作目录,**没有任何访问控制**:代码能读你的整个
|
||
文件系统、能联网。正常情况下不会走到这里 —— macOS 有 Seatbelt、
|
||
Linux 装个 bubblewrap 即可。真要用它跑不信任的代码,别。
|
||
"""
|
||
|
||
name = "local"
|
||
|
||
def __init__(self) -> None:
|
||
self._root = settings.workspace_dir / "local"
|
||
self._root.mkdir(parents=True, exist_ok=True)
|
||
|
||
async def health(self) -> dict[str, object]:
|
||
return {
|
||
"backend": self.name,
|
||
"available": True,
|
||
"isolated": False,
|
||
"warning": "以本地子进程执行,无访问控制,请勿运行不可信的代码",
|
||
"enforced": {"timeout": True, "cpu": True,
|
||
"memory": not sys.platform == "darwin", "network": False, "fsize": True},
|
||
"root": str(self._root),
|
||
"interpreter": interpreter.describe(),
|
||
}
|
||
|
||
def _session_dir(self, session_id: str) -> Path:
|
||
# 净化规则统一在 config.sanitize_session 里,四个后端和文件工具共用一份:
|
||
# 这段逻辑以前各抄各的,文件工具那份漏抄了,就成了越界读取的口子。
|
||
path = self._root / sanitize_session(session_id)
|
||
path.mkdir(parents=True, exist_ok=True)
|
||
return path
|
||
|
||
@staticmethod
|
||
def _preexec(limits: SandboxLimits): # pragma: no cover - 只在子进程里跑
|
||
def _apply() -> None:
|
||
resource.setrlimit(resource.RLIMIT_CPU, (limits.timeout, limits.timeout + 2))
|
||
# 注意 macOS 不强制执行 RLIMIT_AS,这条在 Darwin 上是无效的
|
||
nbytes = limits.memory_mb * 1024 * 1024
|
||
for res_name in ("RLIMIT_AS", "RLIMIT_DATA"):
|
||
res_id = getattr(resource, res_name, None)
|
||
if res_id is not None:
|
||
try:
|
||
resource.setrlimit(res_id, (nbytes, nbytes))
|
||
except (ValueError, OSError):
|
||
pass
|
||
resource.setrlimit(resource.RLIMIT_FSIZE, (64 * 1024 * 1024,) * 2)
|
||
resource.setrlimit(resource.RLIMIT_CORE, (0, 0))
|
||
# 不设 RLIMIT_NPROC:macOS 上它按 uid 全局计数,设小了会让沙箱里
|
||
# 任何 fork 都失败(bash 直接起不来),而不是只限制沙箱自己。
|
||
os.setsid() # 独立进程组,超时时能整组 kill
|
||
|
||
return _apply
|
||
|
||
async def run(
|
||
self,
|
||
code: str,
|
||
*,
|
||
language: str = "python",
|
||
limits: SandboxLimits | None = None,
|
||
session_id: str | None = None,
|
||
env: dict[str, str] | None = None,
|
||
files: dict[str, str] | None = None,
|
||
) -> ExecResult:
|
||
limits = limits or SandboxLimits()
|
||
language = (language or "python").lower()
|
||
if language not in _LANG:
|
||
return ExecResult(ok=False, backend=self.name, error=f"不支持的语言:{language}")
|
||
|
||
ext, cmd = _LANG[language]
|
||
# python 的命令前缀运行时才定:要挑一个看不见后端依赖的解释器
|
||
cmd = cmd if cmd is not None else interpreter.python_argv()
|
||
entry = entry_name(ext)
|
||
argv = [*cmd, entry]
|
||
if shutil.which(argv[0]) is None and argv[0] != sys.executable:
|
||
return ExecResult(
|
||
ok=False, backend=self.name, error=f"本机找不到 {argv[0]}"
|
||
)
|
||
|
||
workdir = self._session_dir(session_id or f"tmp-{int(time.time()*1000)}")
|
||
(workdir / entry).write_text(code)
|
||
for rel, content in (files or {}).items():
|
||
target = (workdir / rel).resolve()
|
||
if not str(target).startswith(str(workdir.resolve())):
|
||
return ExecResult(ok=False, backend=self.name, error=f"非法文件路径:{rel}")
|
||
target.parent.mkdir(parents=True, exist_ok=True)
|
||
target.write_text(content)
|
||
|
||
# 只透传最小环境变量,避免把宿主机的 API key 泄漏给被执行的代码
|
||
child_env = {
|
||
"PATH": os.environ.get("PATH", "/usr/bin:/bin"),
|
||
"HOME": str(workdir),
|
||
"LANG": "en_US.UTF-8",
|
||
"PYTHONDONTWRITEBYTECODE": "1",
|
||
**(env or {}),
|
||
}
|
||
|
||
started = time.perf_counter()
|
||
try:
|
||
proc = await asyncio.create_subprocess_exec(
|
||
*argv,
|
||
cwd=str(workdir),
|
||
env=child_env,
|
||
stdout=asyncio.subprocess.PIPE,
|
||
stderr=asyncio.subprocess.PIPE,
|
||
preexec_fn=self._preexec(limits),
|
||
)
|
||
except Exception as e: # noqa: BLE001
|
||
# 首行只说原因,类名和 errno 这些留给日志
|
||
logger.warning("%s 沙箱起进程失败:%s", self.name, raw(e))
|
||
return ExecResult(ok=False, backend=self.name, error=f"无法启动进程:{describe_exception(e)}")
|
||
|
||
try:
|
||
out, err = await asyncio.wait_for(proc.communicate(), timeout=limits.timeout)
|
||
except asyncio.TimeoutError:
|
||
try:
|
||
os.killpg(os.getpgid(proc.pid), 9)
|
||
except (ProcessLookupError, PermissionError):
|
||
proc.kill()
|
||
await proc.wait()
|
||
return ExecResult(
|
||
ok=False,
|
||
backend=self.name,
|
||
timed_out=True,
|
||
exit_code=124,
|
||
duration_ms=int((time.perf_counter() - started) * 1000),
|
||
error=f"执行超过 {limits.timeout} 秒,已终止",
|
||
)
|
||
|
||
stdout, t1 = truncate(out.decode(errors="replace"), limits.max_output)
|
||
stderr, t2 = truncate(err.decode(errors="replace"), limits.max_output)
|
||
written = [
|
||
p.name
|
||
for p in workdir.iterdir()
|
||
if p.is_file() and not p.name.startswith(ENTRY_PREFIX)
|
||
][:50]
|
||
return ExecResult(
|
||
ok=proc.returncode == 0,
|
||
exit_code=proc.returncode or 0,
|
||
stdout=stdout,
|
||
stderr=stderr,
|
||
truncated=t1 or t2,
|
||
backend=self.name,
|
||
duration_ms=int((time.perf_counter() - started) * 1000),
|
||
files_written=written,
|
||
)
|
||
|
||
async def cleanup(self, session_id: str) -> None:
|
||
path = self._session_dir(session_id)
|
||
await asyncio.to_thread(shutil.rmtree, path, True)
|