"""接地验证器(D4)—— Worker 自验证的"接地"来源。 验证分层优先级(D4): 可执行验证(跑代码/跑测试) > facts 对照 > 结构检查 > 模型自由判断(最后手段) - 可执行验证:code 域 .py 工件在临时目录子进程沙箱运行(timeout、-I 隔离、捕获输出)。 - facts 对照:把回答/工件与知识库 facts(statement/keywords)比对覆盖度。 - 结构检查:工件非空、长度达标、不含纯占位。 说明:沙箱目前做"临时目录 + 超时 + 解释器隔离",Windows 下真正禁网需系统级工具, 此处以超时与隔离为主要护栏(文档如实记录)。验证器为纯逻辑,可注入 runner 便于单测。 """ from __future__ import annotations import subprocess import sys import tempfile from pathlib import Path from typing import Any, Callable, Dict, List, Optional, Tuple # 判定为"含代码"的启发标记 _CODE_HINTS = ("def ", "class ", "import ", "return ", "if __name__", "print(") def detect_artifact_language(name: str) -> str: """按文件名推断工件语言:python / json / text。""" suffix = Path(name).suffix.lower() if suffix in (".py", ".pyw"): return "python" if suffix in (".json",): return "json" return "text" def extract_code_block(text: str) -> str: """从模型输出提取 python 代码块(剥除 markdown 围栏),无则返回原文本。""" t = text.strip() fence = chr(96) * 3 # 三个反引号 marker = fence + "python" start = t.find(marker) if start == -1: return t body_start = start + len(marker) end = t.find(fence, body_start) if end == -1: return t[body_start:].strip() return t[body_start:end].strip() def run_code_sandbox(code: str, timeout_s: float = 10.0, runner: Optional[Callable[[List[str], Dict[str, str], float], Tuple[int, str, str]]] = None ) -> Tuple[int, str, str]: """在临时目录子进程运行 python 代码。返回 (returncode, stdout, stderr)。 隔离措施:临时工作目录、-I 隔离模式、timeout 超时强杀、捕获输出。 runner 可注入(测试用假执行器,避免真跑任意代码)。 """ if runner is not None: return runner([sys.executable, "-I"], {}, timeout_s) with tempfile.TemporaryDirectory(prefix="v2_sandbox_") as tmp: script = Path(tmp) / "main.py" script.write_text(code, encoding="utf-8") try: proc = subprocess.run( [sys.executable, "-I", str(script)], capture_output=True, text=True, timeout=timeout_s, cwd=tmp, creationflags=subprocess.CREATE_NO_WINDOW, ) return proc.returncode, proc.stdout or "", proc.stderr or "" except subprocess.TimeoutExpired: return -1, "", "timeout exceeded" class Verifier: """按 D4 分层执行接地验证。""" def __init__(self, code_timeout_s: float = 10.0, sandbox_runner: Optional[Callable[..., Tuple[int, str, str]]] = None): self.code_timeout_s = code_timeout_s self._sandbox_runner = sandbox_runner def verify(self, domain: str, artifact_name: str, artifact_text: str, query: str, kb: Any = None) -> Tuple[bool, List[str]]: """返回 (passed, details)。kb 为 KnowledgeBase(可 None,跳过 facts 层)。""" lang = detect_artifact_language(artifact_name) details: List[str] = [] # 1) 可执行验证(D4 最高优先级):code + python 工件且含代码 if lang == "python" and any(h in artifact_text for h in _CODE_HINTS): rc, out, err = run_code_sandbox(artifact_text, self.code_timeout_s, runner=self._sandbox_runner) if rc == 0: details.append("代码沙箱运行通过 (rc=0)") return True, details # 可执行验证失败 -> 直接拒绝(不落回结构检查,避免"假通过") details.append(f"代码沙箱运行失败 rc={rc}: {(err or out)[:120]}") return False, details elif lang == "python": details.append("工件不含可执行代码(跳过沙箱,进入 facts/结构检查)") # 2) facts 对照 if kb is not None: fact_hits = self._check_facts(domain, artifact_text, kb) if fact_hits: details.append(f"facts 对照命中 {fact_hits}") return True, details # 3) 结构检查 text = artifact_text.strip() if len(text) < 20: details.append(f"工件过短({len(text)} 字符)") return False, details if text.lower() in ("pass", "none", "todo", "待实现", "略"): details.append("工件为占位内容") return False, details details.append("结构检查通过(非空、长度达标)") return True, details def _check_facts(self, domain: str, text: str, kb: Any) -> int: """统计工件文本命中知识库事实的数量。""" facts = kb.facts(domain) or [] if not facts: return 0 hits = 0 for f in facts: kw = f.get("keywords") or [] if any(str(k) in text for k in kw): hits += 1 return hits def build_verifier(cfg: Dict[str, Any]) -> Verifier: """cfg 为 config.worker 段。""" return Verifier(code_timeout_s=float(cfg.get("code_timeout_s", 10)))