"""OpenAI-compatible HTTP client, persistent per-endpoint."""

from __future__ import annotations

from typing import Any

from openai import OpenAI

# Bound every LLM call. Without this the SDK default (600s) lets a stalled
# endpoint hang the worker thread that ran the tool; enough such threads exhaust
# FastMCP's pool and wedge the whole stdio server. 90s is well above a healthy
# deepseek-v4-flash response yet short enough to fail fast and free the thread.
DEFAULT_REQUEST_TIMEOUT = 90.0


def _resolve_timeout(endpoint_cfg: dict[str, Any]) -> float:
    """Per-endpoint request timeout in seconds; falls back to the bounded default."""
    return float(endpoint_cfg.get("request_timeout", DEFAULT_REQUEST_TIMEOUT))


def make_client(endpoint_cfg: dict[str, Any], api_key: str) -> OpenAI:
    """Create a persistent OpenAI client for an endpoint."""
    kwargs: dict[str, Any] = {
        "base_url": endpoint_cfg["base_url"],
        "api_key": api_key,
        "timeout": _resolve_timeout(endpoint_cfg),
    }
    defaults = endpoint_cfg.get("request_defaults", {})
    if "extra_body" in defaults:
        kwargs["default_query"] = {}
    return OpenAI(**kwargs)


class ClientPool:
    """Manages persistent HTTP clients per endpoint."""

    def __init__(self) -> None:
        self._clients: dict[str, OpenAI] = {}

    def get(self, name: str, endpoint_cfg: dict[str, Any], api_key: str) -> OpenAI:
        if name not in self._clients:
            self._clients[name] = make_client(endpoint_cfg, api_key)
        return self._clients[name]