audio_codebuddy/mcp-for-xiaozhi/mcp_pipe.py

"""
mcp_pipe.py — 小智AI MCP 接入点桥接（v8 in-process）

不用 subprocess。直接 import audio_server 的工具，
在主进程里用 websockets.sync 跟 xiaozhi.me 通信。
完全绕开 Windows + Python 3.14 的 stdio pipe 怪行为。
"""
from __future__ import annotations

import json
import logging
import os
import sys
import time

import websockets.sync.client as ws_sync

import audio_server

logger = logging.getLogger("MCP_PIPE")

HERE = os.path.dirname(os.path.abspath(__file__))

RECONNECT_DELAY = 5  # 断开后等 N 秒重连


def run_session(endpoint: str, raw_log) -> bool:
    """单次 WS 会话，返回 True=正常退出(KeyboardInterrupt)，False=断线需要重连"""
    try:
        ws = ws_sync.connect(endpoint, max_size=4 * 1024 * 1024)
    except Exception as e:
        logger.warning("WS connect failed: %s, will retry in %ds", e, RECONNECT_DELAY)
        return False

    logger.info("WS connected")
    logger.info("READY: ready to receive xiaozhi MCP requests")

    try:
        while True:
            try:
                raw = ws.recv(timeout=300)
            except TimeoutError:
                logger.info("WS recv timeout 5min, reconnecting...")
                return False
            except Exception as e:
                logger.warning("WS recv failed: %s", e)
                return False
            if not raw:
                logger.warning("WS recv empty (EOF)")
                return False
            if isinstance(raw, bytes):
                raw = raw.decode("utf-8", errors="replace")
            logger.info("RAW: %s", raw[:300])
            raw_log.write(f"[<-] {raw[:300]}\n")
            raw_log.flush()
            try:
                req = json.loads(raw)
            except Exception as e:
                logger.warning("parse err: %s", e)
                continue

            method = req.get("method")
            rid = req.get("id")
            logger.info("<- method=%s id=%s", method, rid)

            if rid is None and method and method.startswith("notifications/"):
                logger.info("   (notification, no response)")
                continue

            resp = audio_server.handle_request(req)
            if resp is None:
                continue
            line = (json.dumps(resp, ensure_ascii=False, separators=(",", ":")) + "\n").encode("utf-8")
            logger.info("-> response for %s id=%s (%d bytes)", method, rid, len(line))
            raw_log.write(f"[->] {line.decode('utf-8', errors='replace')[:300]}\n")
            raw_log.flush()
            try:
                ws.send(line)
            except Exception as e:
                logger.warning("WS send failed (connection likely closed): %s", e)
                return False
    except KeyboardInterrupt:
        return True
    except Exception as e:
        logger.warning("Unexpected error in session: %s", e)
        return False
    finally:
        try:
            ws.close()
        except Exception:
            pass

    return False


def main() -> int:
    for k in ("HTTP_PROXY", "HTTPS_PROXY", "http_proxy", "https_proxy", "ALL_PROXY"):
        os.environ.pop(k, None)

    endpoint = os.environ.get("MCP_ENDPOINT")
    if not endpoint:
        print("ERROR: MCP_ENDPOINT 必填")
        return 2

    logging.basicConfig(
        level=os.environ.get("MCP_LOG_LEVEL", "INFO"),
        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
    )

    raw_log_path = os.path.join(HERE, "mcp-pipe-utf8.log")
    raw_log = open(raw_log_path, "a", encoding="utf-8")
    raw_log.write(f"=== mcp_pipe started (auto-reconnect) ===\n")
    raw_log.flush()

    logger.info("AudioBook MCP server connecting to %s", endpoint)
    logger.info("API base = %s", audio_server.API_BASE)

    while True:
        should_exit = run_session(endpoint, raw_log)
        if should_exit:
            logger.info("KeyboardInterrupt, exiting")
            break
        logger.info("Disconnected, reconnecting in %ds...", RECONNECT_DELAY)
        time.sleep(RECONNECT_DELAY)

    return 0


if __name__ == "__main__":
    sys.exit(main())