"""供 CLI 和外部嵌入使用的 LLM client 工厂。"""

from __future__ import annotations

import os
import logging

from pam_deploy_graph.logging_utils import json_for_log
from .base import LlmClient
from .openai_compatible import OpenAICompatibleLlmClient, load_prompt_text
from .rule_based import RuleBasedLlmClient

logger = logging.getLogger(__name__)


def build_llm_client(
    *,
    base_url: str | None = None,
    api_key: str | None = None,
    model: str | None = None,
    action_analysis_prompt_path: str | None = None,
) -> LlmClient:
    """根据显式参数或环境变量构造 LLM client。"""
    actual_base_url = base_url if base_url is not None else os.getenv("PAM_LLM_BASE_URL", "")
    actual_api_key = api_key if api_key is not None else os.getenv("PAM_LLM_API_KEY", "")
    actual_model = model if model is not None else os.getenv("PAM_LLM_MODEL", "")
    actual_action_prompt_path = (
        action_analysis_prompt_path
        if action_analysis_prompt_path is not None
        else os.getenv("PAM_LLM_ACTION_ANALYSIS_PROMPT_FILE", "")
    )
    logger.info(
        "构建 LLM client base_url=%s model=%s has_api_key=%s action_prompt_path=%s explicit=%s",
        actual_base_url,
        actual_model,
        bool(actual_api_key),
        actual_action_prompt_path,
        json_for_log(
            {
                "base_url": base_url,
                "api_key": api_key,
                "model": model,
                "action_analysis_prompt_path": action_analysis_prompt_path,
            }
        ),
    )

    if not actual_base_url and not actual_api_key and not actual_model:
        logger.info("未配置真实 LLM，使用 RuleBasedLlmClient fallback")
        return RuleBasedLlmClient()

    missing = []
    if not actual_base_url:
        missing.append("base_url")
    if not actual_model:
        missing.append("model")
    if missing:
        logger.info("LLM 配置不完整 missing=%s", missing)
        raise ValueError(f"LLM 配置不完整，缺少: {', '.join(missing)}")

    client = OpenAICompatibleLlmClient(
        base_url=actual_base_url,
        api_key=actual_api_key,
        model=actual_model,
        action_analysis_prompt=load_prompt_text(actual_action_prompt_path),
    )
    logger.info("真实 LLM client 构建完成 client=%s model=%s has_api_key=%s", type(client).__name__, actual_model, bool(actual_api_key))
    return client