ai-kefu-admin/app/engine/routes/chat_rest.py

"""REST shell for the HESUPIPE app (non-streaming, JSON-in/JSON-out).

These routes wrap `ConversationManager`, shared with the existing `/ws/chat`
widget path. The app uses HTTP because it lacks a WebSocket client and is fine
with a 3-8s "AI is typing" affordance covering the round trip.

Red lines (see spec):
- WS path (`app/api/routes/chat.py`) shares `ConversationManager` with these
  routes; both paths stay drift-free.
- `context` seeding is conditional + whitelist (handled in
  `ConversationManager.create_conversation`; Task 9 起白名单 =
  租户 config.engine.context_seed_keys——jiancai product_*/auto vehicle_*).
- The message route consumes `stream_process_message`; lead email is
  triggered inside `ConversationManager` when contact info is detected
  (contact-driven-lead-email plan).
"""
from fastapi import APIRouter, HTTPException, Request
from pydantic import BaseModel
from sqlalchemy import select

from app.engine.models import ConversationSource, ConversationStatus, Message
from app.engine.routes.runtime import get_runtime
from app.engine.services.conversation import ConversationManager
from app.engine.services.rate_limit import DEFAULT_MAX
from app.engine.services.rate_limit import check as rate_check

router = APIRouter(prefix="/api/chat", tags=["chat-rest"])


class SessionRequest(BaseModel):
    session_id: str | None = None
    context: dict | None = None
    client_info: dict | None = None


def _server_ip(request: Request) -> str:
    """服务端可信 IP:XFF 第一段优先,否则直连 host;全空返回空串。

    同步自上游 ai-kefu(app/api/routes/chat_rest.py):REST 路径原先只在有
    XFF 时覆盖自报 ip,直连(无代理)场景自报假 IP 原样落库。"""
    xff = request.headers.get("x-forwarded-for", "").split(",")[0].strip()
    return xff or (request.client.host if request.client else "")


def _merge_client_info(client_info: dict | None, server_ip: str) -> dict:
    """服务端 IP 覆盖自报 ip 键;服务端为空时保留自报(空串不得冲掉客户端值)。"""
    merged = dict(client_info or {})
    if server_ip:
        merged["ip"] = server_ip
    return merged


@router.post("/session")
async def chat_session(req: SessionRequest, request: Request):
    """Resume an active conversation by session_id, or create a new app one.

    `context` flows into `create_conversation` where it is whitelisted into
    `raw_info` (only when non-None). `client_info` is stored for analytics only
    (ip/ua/device-id/lang) and never fed to the AI prompt.
    """
    # 服务端可信 IP 同步自上游 ai-kefu:XFF 第一段优先,无 XFF 用直连 host
    # 覆盖客户端自报(可伪造面最小);全空保留自报。?site= 处理不变。
    client_info = _merge_client_info(req.client_info, _server_ip(request))

    # Task 10: ?site=<slug> 记入 client_info 作会话站点标(与 WS init 的
    # site 字段同语义)——只加记录,不参与业务逻辑
    site_slug = request.query_params.get("site", "").strip()
    if site_slug:
        client_info["site"] = site_slug

    # Task 5b: 租户运行时经 state/默认租户兜底解析(过渡期无中间件);
    # manager 由装配线从 runtime 造(db 会话 + runtime 引用)
    runtime = get_runtime(request)
    async with runtime.session_factory() as db:
        manager = ConversationManager(db, runtime=runtime)
        if req.session_id:
            conv = await manager.get_by_session(req.session_id)
            if conv and conv.status == ConversationStatus.active:
                return {
                    "session_id": conv.session_id,
                    "is_new": False,
                    "ended": False,
                }
        conv = await manager.create_conversation(
            source=ConversationSource.app,
            client_info=client_info,
            context=req.context,
        )
        return {
            "session_id": conv.session_id,
            "is_new": True,
            "ended": False,
        }


@router.get("/{session_id}/messages")
async def chat_history(session_id: str, request: Request):
    """Public history endpoint for session replay. App-source only, to narrow
    the read surface; an unknown or non-app session yields 404.
    """
    runtime = get_runtime(request)
    async with runtime.session_factory() as db:
        manager = ConversationManager(db, runtime=runtime)
        conv = await manager.get_by_session(session_id)
        if not conv or conv.source != ConversationSource.app:
            raise HTTPException(status_code=404, detail="session not found")
        result = await db.execute(
            select(Message)
            .where(Message.conversation_id == conv.id)
            .order_by(Message.created_at)
        )
        msgs = result.scalars().all()
        return [
            {
                "role": m.role.value,
                "content": m.content,
                "created_at": m.created_at.isoformat(),
            }
            for m in msgs
        ]


class MessageRequest(BaseModel):
    session_id: str
    content: str


def _resolve_rate_limit(config_data: dict | None) -> int:
    """engine.rate_limit_per_hour 取值(Task 10): 0 或负 = 回退出厂默认
    （与姊妹键 ws_conn_limit 一致）——config_validation 已拒 0/负,此处兜
    直写快照/旧库数据。"""
    v = ((config_data or {}).get("engine") or {}).get("rate_limit_per_hour")
    return v if v and v > 0 else DEFAULT_MAX


@router.post("/message")
async def chat_message(req: MessageRequest, request: Request):
    """Process one app turn: stream-collect the AI reply, then evaluate.

    🔴 Consumes `stream_process_message` so the REST path shares one code
    path with the WS widget and never drifts. Lead email is triggered inside
    `ConversationManager` when contact info is detected
    (contact-driven-lead-email plan) — this route carries no email logic.

    A light in-memory rate limit (per tenant + ip + X-Device-Id, rolling 1h)
    guards GLM spend; only this route is throttled. Task 9: 桶键加租户维度
    (两租户同出口 IP 不互挤配额),上限读租户 config.engine.rate_limit_per_hour
    (默认 30,与出厂一致)。
    """
    runtime = get_runtime(request)
    ip = request.client.host if request.client else ""
    dev = request.headers.get("X-Device-Id", "")
    limit = _resolve_rate_limit(runtime.config.data)
    if not rate_check(runtime.config.slug, ip, dev, max_requests=limit):
        raise HTTPException(status_code=429, detail="rate limit exceeded")

    async with runtime.session_factory() as db:
        manager = ConversationManager(db, runtime=runtime)
        conv = await manager.get_by_session(req.session_id)
        if not conv:
            raise HTTPException(status_code=404, detail="session not found")
        if conv.status != ConversationStatus.active:
            raise HTTPException(status_code=409, detail="conversation ended")

        reply_parts: list[str] = []
        ended = False
        intent_level = "pending"
        async for event in manager.stream_process_message(
            req.session_id, req.content
        ):
            t = event.get("type")
            if t == "stream_chunk":
                reply_parts.append(event.get("content", ""))
            elif t == "evaluated":
                ended = bool(event.get("ended"))
                intent_level = event.get("intent_level", "pending")
            elif t == "error":
                msg = event.get("message", "error")
                raise HTTPException(
                    status_code=409 if "ended" in msg else 404, detail=msg
                )

        return {
            "reply": "".join(reply_parts),
            "ended": ended,
            "intent_level": intent_level,
        }