"""Reusable frame processors shared by cascade and realtime pipelines."""

import asyncio
from collections.abc import Callable
from uuid import uuid4

from loguru import logger
from models import AssistantConfig
from services.brains import Brain
from services.conversation_history import ConversationRecorder
from services.knowledge import search as search_knowledge
from db.session import SessionLocal

from pipecat.frames.frames import (
    InputTransportMessageFrame,
    InterruptionFrame,
    LLMContextFrame,
    LLMFullResponseEndFrame,
    LLMFullResponseStartFrame,
    LLMTextFrame,
    OutputTransportMessageUrgentFrame,
    TextFrame,
    UserImageRawFrame,
    UserImageRequestFrame,
)
from pipecat.processors.aggregators.llm_response_universal import (
    LLMAssistantAggregator,
)
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.turns.user_mute.base_user_mute_strategy import BaseUserMuteStrategy
from pipecat.utils.time import time_now_iso8601


KNOWLEDGE_CONTEXT_MARKER = "<!-- knowledge-context -->"


def _text_input(message) -> tuple[str, bool] | None:
    """解析现有 user-text 与 RTVI send-text 两种前端文字消息。"""
    if not isinstance(message, dict):
        return None
    if message.get("type") == "user-text":
        text = str(message.get("text") or "").strip()
        return (text, True) if text else None
    if message.get("type") == "send-text":
        data = message.get("data")
        if not isinstance(data, dict):
            return None
        text = str(data.get("content") or "").strip()
        options = data.get("options")
        run_immediately = not isinstance(options, dict) or options.get(
            "run_immediately", True
        )
        return (text, bool(run_immediately)) if text else None
    return None


class TextInputProcessor(FrameProcessor):
    """把 transport 文字消息转换成 LLM 可消费的帧。

    run_immediately(默认/打断):先通过 on_text_input 事件把用户文字交给
    run_pipeline 登记,再用 broadcast_interruption() 打断当前播报。新的 LLM
    回复由 assistant aggregator 确认处理完 interruption 后触发。
    run_immediately=False(RTVI send-text 静默追加):仅把文字写进上下文,
    不打断、不触发推理。
    """

    def __init__(self, should_ignore_input: Callable[[], bool] | None = None):
        super().__init__()
        self._should_ignore_input = should_ignore_input or (lambda: False)
        # 立即触发的文字(含打断语义)走 on_text_input;静默追加另走一条事件
        self._register_event_handler("on_text_input")
        self._register_event_handler("on_text_append")
        self._register_event_handler("on_client_ready")

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)

        if not isinstance(frame, InputTransportMessageFrame):
            await self.push_frame(frame, direction)
            return

        if isinstance(frame.message, dict) and frame.message.get("type") == "client-ready":
            await self._call_event_handler("on_client_ready")
            return

        parsed = _text_input(frame.message)
        if not parsed:
            await self.push_frame(frame, direction)
            return

        if self._should_ignore_input():
            logger.debug("通话正在结束,忽略后续文字输入")
            return

        text, run_immediately = parsed
        if run_immediately:
            # 先登记文字再打断。下一轮 LLM 由 assistant aggregator 在真正处理完
            # InterruptionFrame 后触发,避免新回复被这次 interruption 一起取消。
            await self._call_event_handler("on_text_input", text)
            await self.broadcast_interruption()
        else:
            await self._call_event_handler("on_text_append", text)


class CallEndingUserMuteStrategy(BaseUserMuteStrategy):
    """Keep user media muted after an end-call tool starts terminating a call."""

    def __init__(self, is_call_ending: Callable[[], bool]):
        super().__init__()
        self._is_call_ending = is_call_ending

    async def process_frame(self, frame) -> bool:
        await super().process_frame(frame)
        return self._is_call_ending()


class VisionCaptureProcessor(FrameProcessor):
    """Capture one requested video frame for auxiliary vision-model analysis."""

    def __init__(self, timeout_s: float = 3.0):
        super().__init__()
        self._timeout_s = timeout_s
        self._pending: dict[str, asyncio.Future[UserImageRawFrame]] = {}

    async def request_image(
        self,
        requester: FrameProcessor,
        request: UserImageRequestFrame,
    ) -> UserImageRawFrame:
        key = request.tool_call_id or str(uuid4())
        request.tool_call_id = key
        request.append_to_context = False
        request.result_callback = None

        loop = asyncio.get_running_loop()
        future: asyncio.Future[UserImageRawFrame] = loop.create_future()
        self._pending[key] = future
        await requester.push_frame(request, FrameDirection.UPSTREAM)
        try:
            return await asyncio.wait_for(future, timeout=self._timeout_s)
        finally:
            self._pending.pop(key, None)

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)

        if (
            isinstance(frame, UserImageRawFrame)
            and frame.request
            and frame.request.tool_call_id
            and frame.request.tool_call_id in self._pending
        ):
            future = self._pending[frame.request.tool_call_id]
            if not future.done():
                future.set_result(frame)
            return

        await self.push_frame(frame, direction)


class RealtimeDynamicVariableProcessor(FrameProcessor):
    """Keep realtime system turn/history variables current between responses."""

    def __init__(self, brain: Brain, cfg: AssistantConfig, realtime):
        super().__init__()
        self._brain = brain
        self._cfg = cfg
        self._realtime = realtime

    async def _refresh_instructions(self) -> None:
        update = getattr(self._realtime, "update_instructions", None)
        if callable(update):
            await update(self._brain.system_prompt(self._cfg))

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)
        if isinstance(frame, OutputTransportMessageUrgentFrame):
            message = frame.message
            if isinstance(message, dict):
                event_type = message.get("type")
                if event_type == "transcript" and message.get("role") == "user":
                    content = str(message.get("content") or "").strip()
                    if content:
                        self._brain.record_user_message(content)
                        await self._refresh_instructions()
                elif event_type == "assistant-text-end":
                    await self._brain.on_assistant_text_end(
                        str(message.get("turn_id") or ""),
                        str(message.get("content") or ""),
                        bool(message.get("interrupted", False)),
                    )
                    await self._refresh_instructions()
        await self.push_frame(frame, direction)


class RealtimeTextInputProcessor(FrameProcessor):
    """Route text input directly to a realtime service without cascade semantics."""

    def __init__(self):
        super().__init__()
        self._register_event_handler("on_text_input")
        self._register_event_handler("on_text_append")

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)

        if not isinstance(frame, InputTransportMessageFrame):
            await self.push_frame(frame, direction)
            return

        parsed = _text_input(frame.message)
        if not parsed:
            await self.push_frame(frame, direction)
            return

        text, run_immediately = parsed
        await self._call_event_handler(
            "on_text_input" if run_immediately else "on_text_append",
            text,
        )


class ConversationHistoryProcessor(FrameProcessor):
    """从最终客户端事件旁路保存历史，不改变 Pipecat 的上下文与帧语义。"""

    def __init__(self, recorder: ConversationRecorder | None):
        super().__init__()
        self._recorder = recorder

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)
        await self.push_frame(frame, direction)
        if self._recorder and isinstance(frame, OutputTransportMessageUrgentFrame):
            await self._recorder.record_transport_message(frame.message)


class KnowledgeRetrievalProcessor(FrameProcessor):
    """Retrieve before local LLM inference without changing Pipecat internals."""

    def __init__(
        self,
        knowledge_base_id: str | None,
        top_n: int = 5,
        score_threshold: float = 0.0,
    ):
        super().__init__()
        self._knowledge_base_id = knowledge_base_id
        self._top_n = top_n
        self._score_threshold = score_threshold
        self._mode = "automatic" if knowledge_base_id else "disabled"
        self._last_signature = ""

    def set_scope(self, scope: dict) -> None:
        self._knowledge_base_id = scope.get("knowledge_base_id") or None
        self._mode = str(scope.get("mode") or "disabled")
        self._top_n = int(scope.get("top_n") or 5)
        self._score_threshold = float(scope.get("score_threshold") or 0.0)
        self._last_signature = ""

    def _clear_context(self, messages: list[dict]) -> None:
        # Remove the legacy Workflow knowledge message so an in-flight context
        # created before this compatibility fix cannot keep sending that role.
        messages[:] = [
            message
            for message in messages
            if not (
                message.get("role") == "developer"
                and KNOWLEDGE_CONTEXT_MARKER in str(message.get("content") or "")
            )
        ]
        system_message = next(
            (message for message in messages if message.get("role") == "system"),
            None,
        )
        if system_message is not None:
            content = str(system_message.get("content") or "")
            system_message["content"] = content.split(KNOWLEDGE_CONTEXT_MARKER, 1)[0].rstrip()

    def _set_context(self, messages: list[dict], block: str) -> None:
        """Store retrieved knowledge in a provider-compatible system message."""
        self._clear_context(messages)
        system_message = next(
            (message for message in messages if message.get("role") == "system"),
            None,
        )
        if system_message is None:
            messages.insert(0, {"role": "system", "content": block})
            return
        content = str(system_message.get("content") or "").rstrip()
        system_message["content"] = f"{content}\n\n{block}" if content else block

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)
        if not isinstance(frame, LLMContextFrame):
            await self.push_frame(frame, direction)
            return

        messages = frame.context.get_messages()
        if self._mode != "automatic" or not self._knowledge_base_id:
            self._clear_context(messages)
            await self.push_frame(frame, direction)
            return
        user_messages = [message for message in messages if message.get("role") == "user"]
        if not user_messages:
            await self.push_frame(frame, direction)
            return
        query = str(user_messages[-1].get("content") or "").strip()
        signature = f"{len(user_messages)}:{query}"
        if not query or signature == self._last_signature:
            await self.push_frame(frame, direction)
            return
        self._last_signature = signature

        try:
            async with SessionLocal() as session:
                results = await search_knowledge(
                    session,
                    self._knowledge_base_id,
                    query,
                    top_k=self._top_n,
                    score_threshold=self._score_threshold,
                )
        except Exception as exc:
            logger.warning(f"自动知识库检索失败: {exc}")
            results = []

        sources = "\n\n".join(
            f"[{index + 1}] 来源：{item['document']}（相关度 {item['score']}）\n{item['content']}"
            for index, item in enumerate(results)
        ) or "未检索到相关资料。"
        block = f"{KNOWLEDGE_CONTEXT_MARKER}\n当前问题的知识库检索结果：\n{sources}"
        self._set_context(messages, block)
        await self.push_frame(frame, direction)


class UserTurnRoutingProcessor(FrameProcessor):
    """Give a brain first right of refusal before a new user turn reaches the LLM."""

    def __init__(self, brain: Brain):
        super().__init__()
        self._brain = brain
        self._last_user_message: dict | None = None

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)
        if direction != FrameDirection.DOWNSTREAM or not isinstance(
            frame, LLMContextFrame
        ):
            await self.push_frame(frame, direction)
            return

        user_message = next(
            (
                message
                for message in reversed(frame.context.get_messages())
                if message.get("role") == "user"
                and isinstance(message.get("content"), str)
                and str(message.get("content") or "").strip()
            ),
            None,
        )
        if user_message is None:
            await self.push_frame(frame, direction)
            return

        if user_message is self._last_user_message:
            # Programmatic LLMRunFrame after a node transition reuses the same
            # user message. It is a response run, not another routing event.
            await self.push_frame(frame, direction)
            return
        self._last_user_message = user_message

        content = str(user_message.get("content") or "").strip()
        handled = await self._brain.on_user_turn_end(content)
        if not handled:
            await self.push_frame(frame, direction)


class PassthroughLLMAssistantAggregator(LLMAssistantAggregator):
    """聚合 LLM 回复进上下文,同时继续把回复帧交给下游 TTS。"""

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._register_event_handler("on_interruption_processed")
        self._register_event_handler("on_assistant_text_start")
        self._register_event_handler("on_assistant_text_delta")
        self._register_event_handler("on_assistant_text_end")
        self._stream_turn_id: str | None = None
        self._stream_timestamp = ""
        self._stream_text = ""
    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)

        if isinstance(frame, LLMFullResponseStartFrame):
            self._stream_turn_id = uuid4().hex
            self._stream_timestamp = time_now_iso8601()
            self._stream_text = ""
            await self._call_event_handler(
                "on_assistant_text_start",
                self._stream_turn_id,
                self._stream_timestamp,
            )
        elif isinstance(frame, LLMTextFrame) and self._stream_turn_id:
            self._stream_text += frame.text
            await self._call_event_handler(
                "on_assistant_text_delta",
                self._stream_turn_id,
                frame.text,
            )
        elif isinstance(frame, LLMFullResponseEndFrame):
            await self._finish_text_stream(interrupted=False)

        # LLMAssistantAggregator 默认会消费这些帧。放在 TTS 前用于中断时保存
        # 已生成前缀时,必须显式透传,否则 TTS 收不到任何 LLM 回复。
        if isinstance(
            frame,
            (LLMFullResponseStartFrame, LLMFullResponseEndFrame, TextFrame),
        ):
            await self.push_frame(frame, direction)
        elif isinstance(frame, InterruptionFrame):
            await self._finish_text_stream(interrupted=True)
            await self._call_event_handler("on_interruption_processed")

    async def _finish_text_stream(self, *, interrupted: bool):
        if not self._stream_turn_id:
            return
        await self._call_event_handler(
            "on_assistant_text_end",
            self._stream_turn_id,
            self._stream_text,
            interrupted,
        )
        self._stream_turn_id = None
        self._stream_timestamp = ""
        self._stream_text = ""


class WorkflowAggregatorPair:
    """Small public-shape adapter required by Pipecat FlowManager."""

    def __init__(self, user_aggregator, assistant_aggregator):
        self._user = user_aggregator
        self._assistant = assistant_aggregator

    def user(self):
        return self._user

    def assistant(self):
        return self._assistant