MAI-Code-1-Flash：微软的"反榜单"编程模型，少花 token 多干活

# step1_locate.py — 先用 grep 精确定位，缩小模型需要处理的范围
import subprocess
import json

def locate_references(project_dir: str, old_name: str) -> dict:
    """用 grep 找到所有引用行，只把命中行及上下文交给模型"""
    result = subprocess.run(
        ["grep", "-rn", old_name, project_dir, "--include=*.py"],
        capture_output=True, text=True
    )
    hits = []
    for line in result.stdout.splitlines():
        file_path, lineno, content = line.split(":", 2)
        hits.append({
            "file": file_path,
            "line": int(lineno),
            "content": content.strip()
        })
    return {"old_name": old_name, "references": hits}

# 运行定位
report = locate_references("./my_project", "process_request")
print(json.dumps(report, indent=2))

{
  "old_name": "process_request",
  "references": [
    {"file": "my_project/old_handler.py", "line": 12, "content": "def process_request(req):"},
    {"file": "my_project/api/routes.py", "line": 45, "content": "from old_handler import process_request"},
    {"file": "my_project/api/routes.py", "line": 67, "content": "result = process_request(payload)"},
    {"file": "my_project/tests/test_handler.py", "line": 23, "content": "resp = process_request(mock_req)"}
  ]
}

# step2_refactor.py — 构造精简 prompt，让模型用最少 token 完成重构
import openai

def build_compact_prompt(report: dict, new_name: str) -> str:
    """只包含命中行，不塞整个文件——这就是 '更少 token' 的关键"""
    refs_text = "\n".join(
        f"  {r['file']}:{r['line']} — {r['content']}"
        for r in report["references"]
    )
    return (
        f"将函数 `{report['old_name']}` 重命名为 `{new_name}`。\n"
        f"以下是需要修改的引用点：\n{refs_text}\n\n"
        f"给出每个文件的修改 diff，格式为：\n"
        f"文件路径 | 旧行 | 新行"
    )

prompt = build_compact_prompt(report, "handle_request")

# 调用模型（替换为实际部署的 endpoint）
client = openai.OpenAI(base_url="YOUR_DEPLOYMENT_URL", api_key="YOUR_KEY")
response = client.chat.completions.create(
    model="mai-code-1-flash",  # 或你的模型名
    messages=[{"role": "user", "content": prompt}],
    max_tokens=512,  # 任务明确，不需要长输出
    temperature=0
)

print(response.choices[0].message.content)

my_project/old_handler.py:12 | def process_request(req): | def handle_request(req):
my_project/api/routes.py:45 | from old_handler import process_request | from old_handler import handle_request
my_project/api/routes.py:67 | result = process_request(payload) | result = handle_request(payload)
my_project/tests/test_handler.py:23 | resp = process_request(mock_req) | resp = handle_request(mock_req)