把 GPT-6 家族用进生产：模型选择、推理预算与工具编排指南

# model-policy.yaml
routes:
  extraction:
    model_role: fast
    reasoning_effort: low
    max_retries: 1
  customer_reply:
    model_role: balanced
    reasoning_effort: medium
    max_retries: 1
  production_change:
    model_role: deep
    reasoning_effort: high
    human_approval: true
    max_retries: 0

技能：refund-review/v3

目标：判断退款请求是否满足政策，并生成给客服人员的建议。

流程：
1. 读取订单状态，不得猜测。
2. 检查购买时间、商品类型和退款窗口。
3. 缺少字段时返回 needs_more_information。
4. 只生成建议，不直接执行退款。
5. 输出 JSON：decision、reason、missing_fields、recommended_action。

#!/usr/bin/env python3
import json
import os
import sys
import urllib.request

MODELS = {
    "fast": os.getenv("GPT6_FAST_MODEL", "replace-with-fast-model"),
    "balanced": os.getenv("GPT6_BALANCED_MODEL", "replace-with-balanced-model"),
    "deep": os.getenv("GPT6_DEEP_MODEL", "replace-with-deep-model"),
}


def choose_route(task: str, text: str) -> tuple[str, str]:
    high_risk_words = ("production", "payment", "permission", "delete", "生产", "付款", "权限", "删除")
    complex_words = ("compare", "debug", "plan", "比较", "调试", "规划")

    if task == "production_change" or any(word in text.lower() for word in high_risk_words):
        return "deep", "high"
    if len(text) > 4000 or any(word in text.lower() for word in complex_words):
        return "balanced", "medium"
    return "fast", "low"


def build_request(task: str, text: str) -> dict:
    role, effort = choose_route(task, text)
    return {
        "model": MODELS[role],
        "reasoning_effort": effort,
        "input": [
            {
                "role": "system",
                "content": (
                    "Return JSON with keys answer, confidence, and needs_review. "
                    "Do not claim that an external action succeeded unless a tool result confirms it."
                ),
            },
            {"role": "user", "content": text},
        ],
        "metadata": {"task": task, "model_role": role},
    }


def send(payload: dict) -> dict:
    if os.getenv("LIVE") != "1":
        return {"dry_run": True, "request": payload}

    url = os.environ["GPT6_API_URL"]
    api_key = os.environ["GPT6_API_KEY"]
    request = urllib.request.Request(
        url,
        data=json.dumps(payload).encode("utf-8"),
        headers={
            "Authorization": f"Bearer {api_key}",
            "Content-Type": "application/json",
        },
        method="POST",
    )
    with urllib.request.urlopen(request, timeout=60) as response:
        return json.loads(response.read().decode("utf-8"))


if __name__ == "__main__":
    task = sys.argv[1] if len(sys.argv) > 1 else "customer_reply"
    text = sys.argv[2] if len(sys.argv) > 2 else "Summarize this customer request."
    print(json.dumps(send(build_request(task, text)), ensure_ascii=False, indent=2))

python3 gpt6_router.py customer_reply "总结用户的问题并列出缺失信息"
python3 gpt6_router.py production_change "删除生产环境中的旧索引"

export GPT6_FAST_MODEL="实际的快速模型名"
export GPT6_BALANCED_MODEL="实际的均衡模型名"
export GPT6_DEEP_MODEL="实际的深度推理模型名"
export GPT6_API_URL="https://your-provider.example/v1/responses"
export GPT6_API_KEY="your-api-key"
export LIVE=1
python3 gpt6_router.py customer_reply "为延期订单生成回复草稿"