Cursor Composer 2.5：文本反馈强化学习如何让 AI 编程智能体更稳更准

import json
from pathlib import Path

# ---- 1. 定义文本反馈生成器 ----
def generate_text_feedback(step_record: dict, instruction: str) -> str:
    """根据单步执行记录和原始指令，生成定向文本反馈"""
    violations = []

    # 检查指令遵循：只改 controller？
    if "只改 controller" in instruction:
        changed_files = step_record.get("changed_files", [])
        for f in changed_files:
            if "model" in f or "schema" in f:
                violations.append(
                    f"你修改了 {f}，但指令要求只改 controller 层文件。"
                )

    # 检查是否跳步：要求改 3 个函数，实际只改了 1 个？
    required_changes = step_record.get("required_changes", [])
    actual_changes = step_record.get("actual_changes", [])
    if len(actual_changes) < len(required_changes):
        violations.append(
            f"指令要求修改 {len(required_changes)} 处，你只改了 {len(actual_changes)} 处，遗漏了 "
            + ", ".join(set(required_changes) - set(actual_changes))
        )

    if violations:
        return "指令遵循失败：" + "; ".join(violations)
    return "这步执行正确，遵循了所有约束。"


# ---- 2. 模拟一个训练数据收集循环 ----
def collect_feedback_loop(
    trajectory: list[dict], instruction: str
) -> list[dict]:
    """对一条完整轨迹的每一步生成文本反馈"""
    feedback_trajectory = []
    for i, step in enumerate(trajectory):
        feedback = generate_text_feedback(step, instruction)
        feedback_trajectory.append({
            "step_index": i,
            "action": step["action"],
            "text_feedback": feedback,
            "is_positive": "失败" not in feedback,
        })
    return feedback_trajectory


# ---- 3. 演示数据构造 ----
instruction = "只改 controller 层，保持 API 签名不变，修改 3 个函数的参数校验逻辑"

trajectory = [
    {"action": "edit user_controller.py", "changed_files": ["controllers/user_controller.py"],
     "required_changes": ["validate_email", "validate_age", "validate_name"],
     "actual_changes": ["validate_email"]},
    {"action": "edit user_model.py", "changed_files": ["models/user_model.py", "controllers/user_controller.py"],
     "required_changes": ["validate_email", "validate_age", "validate_name"],
     "actual_changes": ["validate_email", "validate_age"]},
]

feedback_data = collect_feedback_loop(trajectory, instruction)

for item in feedback_data:
    print(json.dumps(item, ensure_ascii=False, indent=2))

请重构 payment 模块的错误处理逻辑。
约束：
1. 只改 payment/controller.py 和 payment/service.py，不动 model
2. 保持现有 HTTP 状态码映射不变
3. 新增的异常类放在 payment/exceptions.py
4. 每个函数改完后立即跑 pytest payment/tests/

你在第 4 步修改了 payment/model.py 的 Transaction 类，
这违反了"不动 model"的约束。
请撤销这个修改，只调整 controller 层的异常捕获逻辑。

# 在 .cursorrules 中加入自动检查点规则
cat >> .cursorrules << 'EOF'
## 执行规则
- 每完成一个文件的修改，输出变更摘要并等待确认
- 修改超过 3 个文件时，每 3 个文件暂停，列出已改文件清单
- 遇到测试失败时，先分析失败原因再尝试修复，不要直接改测试
EOF