Skip to content

JiuwenSwarm is an intelligent AI Agent built on openJiuwen. It extends the powerful capabilities of large language models directly to your fingertips through various communication apps you use daily.

Python
5.9 KB
Raw Blame
# Copyright (c) Huawei Technologies Co., Ltd. 2026. All rights reserved.

from __future__ import annotations

import warnings

warnings.filterwarnings(
    "ignore",
    message=r"Protobuf gencode version.*",
)

import os
import tempfile
import time

from jiuwenswarm.agents.harness.common.rails.task_execution_rail import (
    _ARTIFACT_SCAN_MAX_TEXT_BYTES,
    READONLY_INNER_TOOLS,
    _extract_artifact_paths_from_result,
    _extract_file_paths_from_write_tool,
    _extract_image_paths_from_tool_result,
    detect_artifact_paths,
)


def test_write_tool_prefers_path_from_tool_args() -> None:
    paths = _extract_file_paths_from_write_tool(
        "write_file",
        {"path": "/tmp/demo.py"},
        "Wrote /tmp/other.py",
    )
    assert paths == ["/tmp/demo.py"]


def test_write_tool_fallback_extracts_py_from_result_text() -> None:
    paths = _extract_file_paths_from_write_tool(
        "write_file",
        {},
        "Successfully wrote /workspace/out/demo.py",
    )
    assert paths == ["/workspace/out/demo.py"]


def test_write_tool_fallback_ignores_non_python_paths() -> None:
    paths = _extract_file_paths_from_write_tool(
        "write_file",
        {},
        "Saved /workspace/out/chart.png and /workspace/out/notes.txt",
    )
    assert paths == []


def test_image_path_extraction_keeps_image_whitelist() -> None:
    # generate_image 真实输出格式:每行一个路径,可能含空格(effective_project_dir)
    paths = _extract_image_paths_from_tool_result(
        "Generated 1 image(s) successfully!\n"
        "Local file paths (use for attachments or send_file):\n"
        "- E:\\01 code\\proj\\generated_images\\chart.png\n"
        "- /workspace/out/demo.py\n"
        "Prompt: a chart"
    )
    assert paths == ["E:\\01 code\\proj\\generated_images\\chart.png"]


def test_weak_key_candidate_falls_back_to_body_scan(tmp_path) -> None:
    # P1-1 回归:结构化弱键(result: "ok")校验失败后,不应屏蔽
    # stdout 正文中的真实产物路径
    base = tmp_path.resolve()
    report = base / "report.xlsx"
    report.write_bytes(b"x")
    paths = _extract_artifact_paths_from_result(
        {"stdout": f"Saved to {report}\nDone.", "result": "ok"},
        workspace_base=base,
    )
    assert paths == [str(report)]


def test_structured_hit_short_circuits_body_scan(tmp_path) -> None:
    # 结构化候选校验通过即返回,不再扫描正文(对齐 clowder-ai 语义)
    base = tmp_path.resolve()
    report = base / "report.xlsx"
    report.write_bytes(b"x")
    junk = base / "junk.txt"
    junk.write_bytes(b"x")
    paths = _extract_artifact_paths_from_result(
        {"output_path": str(report), "stdout": f"see {junk}"},
        workspace_base=base,
    )
    assert paths == [str(report)]


# evaluate_script 大文本短路:避免 800K HTML 触发 findall + 逐条 stat 阻塞(633s 根因)。


def _big_html(size: int) -> str:
    r"""构造含海量路径样子串的 HTML(href/src/D:\..\.ext 满地都是)。"""
    chunk = (
        '<a href="/api/v1/users/list.html">'
        '<img src="/static/img/foo.png" /> '
        '<script src="https://x.com/a/b/c.js"></script> '
        r'<link rel="stylesheet" href="D:\proj\dist\main.css"> '
        r'C:\Users\me\app\node_modules\pkg\index.js '
    )
    out = []
    while len("".join(out)) < size:
        out.append(chunk)
    return "".join(out)[:size]


def test_invoke_tool_evaluate_script_short_circuits_no_stat_storm() -> None:
    """evaluate_script 经 invoke_tool 间接调用且结果为大文本 HTML:
    READONLY_INNER_TOOLS 命中 → 直接返回空检测,不进正则扫描/逐条 stat。

    回归此路径会阻塞事件循环 633s(见 officeclaw 历史 633075ms AFTER_TOOL_CALL)。
    """
    assert "evaluate_script" in READONLY_INNER_TOOLS

    big_html = _big_html(800_000)
    tool_args = {"tool_name": "evaluate_script", "arguments": {"script": "x"}}
    tool_result = {
        "success": True,
        "tool_name": "evaluate_script",
        "result": big_html,
    }

    t0 = time.monotonic()
    det = detect_artifact_paths("invoke_tool", tool_args, tool_result, workspace_base=None)
    elapsed_ms = (time.monotonic() - t0) * 1000

    # 解包后内部工具名为 evaluate_script,且不产生任何产物路径
    assert det.tool_name == "evaluate_script"
    assert det.paths == []
    # 短路必须极快:旧路径 633s,修复后应在毫秒级
    assert elapsed_ms < 100, f"short-circuit too slow: {elapsed_ms:.1f}ms"


def test_oversize_result_text_skips_regex_scan() -> None:
    """即便绕过只读工具白名单(如某未知只读 MCP 工具返回超大 stdout),
    _ARTIFACT_SCAN_MAX_TEXT_BYTES 兜底也应跳过 findall,避免 stat 风暴。

    正则对 80 万字符 findall 会匹配出海量候选路径,逐条 stat() 串行阻塞。
    """
    big_text = _big_html(_ARTIFACT_SCAN_MAX_TEXT_BYTES * 8)
    assert len(big_text) > _ARTIFACT_SCAN_MAX_TEXT_BYTES

    t0 = time.monotonic()
    paths = _extract_artifact_paths_from_result(big_text, workspace_base=None)
    elapsed_ms = (time.monotonic() - t0) * 1000

    assert paths == []
    assert elapsed_ms < 100, f"oversize guard too slow: {elapsed_ms:.1f}ms"


def test_normal_stdout_still_detects_real_artifact() -> None:
    """兜底防御不能误伤正常 code/bash stdout 产物检测:
    小文本中含真实存在的文件路径时,仍应检出。
    """
    with tempfile.TemporaryDirectory() as tmpdir:
        real_file = os.path.join(tmpdir, "report.csv")
        with open(real_file, "w", encoding="utf-8") as f:
            f.write("a,b\n1,2\n")

        stdout = f"done, output written to {real_file}\nmore lines\n"
        paths = _extract_artifact_paths_from_result(
            {"stdout": stdout}, workspace_base=None
        )

        assert any(
            os.path.normcase(real_file) == os.path.normcase(p) for p in paths
        ), f"regression: real artifact not detected: {paths}"