别再只调 Prompt：用 Harness、测试与上下文打造可靠的编码 Agent

Agent = LLM + Harness + Domain Knowledge

临时 Prompt → 仓库文档 → Lint/类型检查 → 单元测试 → 集成测试 → 上游评测

# Repository Instructions

## Validation
- Run `python -m unittest discover -s tests` after every code change.
- Run `ruff check .` before declaring the task complete.
- Do not suppress failing tests or remove assertions.

## Architecture
- Keep HTTP handlers thin.
- Put business logic under `src/services/`.
- Public API changes require a compatibility test.

## Safety
- Never run recursive deletion commands.
- Never push to a remote repository.
- Do not read `.env`, credential files, or directories outside this repository.

## Completion criteria
- Tests pass.
- The diff contains only task-related changes.
- The final report lists changed files, validation commands, and remaining risks.

from __future__ import annotations

import shutil
import subprocess
import sys
from pathlib import Path

ROOT = Path("agent_demo").resolve()
MAX_ITERATIONS = 5


def create_demo_project() -> None:
    if ROOT.exists():
        shutil.rmtree(ROOT)
    ROOT.mkdir()

    (ROOT / "calc.py").write_text(
        "def add(a, b):\n"
        "    return a - b\n",
        encoding="utf-8",
    )
    (ROOT / "test_calc.py").write_text(
        "import unittest\n"
        "from calc import add\n\n"
        "class CalcTest(unittest.TestCase):\n"
        "    def test_add(self):\n"
        "        self.assertEqual(add(2, 3), 5)\n\n"
        "if __name__ == '__main__':\n"
        "    unittest.main()\n",
        encoding="utf-8",
    )


def run_tests() -> tuple[bool, str]:
    result = subprocess.run(
        [
            sys.executable,
            "-m",
            "unittest",
            "discover",
            "-s",
            str(ROOT),
            "-p",
            "test_*.py",
        ],
        cwd=ROOT,
        text=True,
        capture_output=True,
        timeout=30,
    )
    logs = (result.stdout + "\n" + result.stderr)[-6000:]
    return result.returncode == 0, logs


def propose_patch(test_logs: str) -> tuple[str, str]:
    """模拟模型输出：(相对路径, 新文件内容)。

    实际项目可在这里把任务、相关文件和 test_logs 发给模型，
    但仍应要求模型返回结构化补丁，而不是任意 shell 命令。
    """
    if "AssertionError" in test_logs:
        return "calc.py", "def add(a, b):\n    return a + b\n"
    raise RuntimeError("No safe patch can be proposed from the diagnostics")


def guarded_write(relative_path: str, content: str) -> None:
    target = (ROOT / relative_path).resolve()
    if ROOT not in target.parents:
        raise PermissionError(f"Refusing to write outside workspace: {target}")
    if target.suffix != ".py":
        raise PermissionError(f"Unsupported file type: {target.suffix}")
    target.write_text(content, encoding="utf-8")


def main() -> None:
    create_demo_project()

    for iteration in range(1, MAX_ITERATIONS + 1):
        passed, logs = run_tests()
        print(f"\n--- iteration {iteration} ---")
        print(logs)

        if passed:
            print("Verified: all tests pass.")
            return

        path, content = propose_patch(logs)
        guarded_write(path, content)
        print(f"Applied candidate patch to {path}")

    raise SystemExit("Stopped: iteration limit reached")


if __name__ == "__main__":
    main()