astrbot/core/astr_agent_run_util.py

import asyncio
import re
import time
import traceback
from collections.abc import AsyncGenerator

from astrbot.core import logger
from astrbot.core.agent.message import Message
from astrbot.core.agent.runners.tool_loop_agent_runner import ToolLoopAgentRunner
from astrbot.core.astr_agent_context import AstrAgentContext
from astrbot.core.message.components import BaseMessageComponent, Json, Plain
from astrbot.core.message.message_event_result import (
    MessageChain,
    MessageEventResult,
    ResultContentType,
)
from astrbot.core.persona_error_reply import (
    extract_persona_custom_error_message_from_event,
)
from astrbot.core.provider.entities import LLMResponse
from astrbot.core.provider.provider import TTSProvider

AgentRunner = ToolLoopAgentRunner[AstrAgentContext]


def _should_stop_agent(astr_event) -> bool:
    return astr_event.is_stopped() or bool(astr_event.get_extra("agent_stop_requested"))


def _truncate_tool_result(text: str, limit: int = 70) -> str:
    if limit <= 0:
        return ""
    if len(text) <= limit:
        return text
    if limit <= 3:
        return text[:limit]
    return f"{text[: limit - 3]}..."


def _extract_chain_json_data(msg_chain: MessageChain) -> dict | None:
    if not msg_chain.chain:
        return None
    first_comp = msg_chain.chain[0]
    if isinstance(first_comp, Json) and isinstance(first_comp.data, dict):
        return first_comp.data
    return None


def _record_tool_call_name(
    tool_info: dict | None, tool_name_by_call_id: dict[str, str]
) -> None:
    if not isinstance(tool_info, dict):
        return
    tool_call_id = tool_info.get("id")
    tool_name = tool_info.get("name")
    if tool_call_id is None or tool_name is None:
        return
    tool_name_by_call_id[str(tool_call_id)] = str(tool_name)


def _build_tool_call_status_message(tool_info: dict | None) -> str:
    if tool_info:
        return f"🔨 调用工具: {tool_info.get('name', 'unknown')}"
    return "🔨 调用工具..."


def _build_tool_result_status_message(
    msg_chain: MessageChain, tool_name_by_call_id: dict[str, str]
) -> str:
    tool_name = "unknown"
    tool_result = ""

    result_data = _extract_chain_json_data(msg_chain)
    if result_data:
        tool_call_id = result_data.get("id")
        if tool_call_id is not None:
            tool_name = tool_name_by_call_id.pop(str(tool_call_id), "unknown")
        tool_result = str(result_data.get("result", ""))

    if not tool_result:
        tool_result = msg_chain.get_plain_text(with_other_comps_mark=True)
    tool_result = _truncate_tool_result(tool_result, 70)

    status_msg = f"🔨 调用工具: {tool_name}"
    if tool_result:
        status_msg = f"{status_msg}\n📎 返回结果: {tool_result}"
    return status_msg


async def run_agent(
    agent_runner: AgentRunner,
    max_step: int = 30,
    show_tool_use: bool = True,
    show_tool_call_result: bool = False,
    stream_to_general: bool = False,
    show_reasoning: bool = False,
) -> AsyncGenerator[MessageChain | None, None]:
    step_idx = 0
    astr_event = agent_runner.run_context.context.event
    tool_name_by_call_id: dict[str, str] = {}
    while step_idx < max_step + 1:
        step_idx += 1

        if step_idx == max_step + 1:
            logger.warning(
                f"Agent reached max steps ({max_step}), forcing a final response."
            )
            if not agent_runner.done():
                # 拔掉所有工具
                if agent_runner.req:
                    agent_runner.req.func_tool = None
                # 注入提示词
                agent_runner.run_context.messages.append(
                    Message(
                        role="user",
                        content="工具调用次数已达到上限，请停止使用工具，并根据已经收集到的信息，对你的任务和发现进行总结，然后直接回复用户。",
                    )
                )

        stop_watcher = asyncio.create_task(
            _watch_agent_stop_signal(agent_runner, astr_event),
        )
        try:
            async for resp in agent_runner.step():
                if _should_stop_agent(astr_event):
                    agent_runner.request_stop()

                if resp.type == "aborted":
                    if not stop_watcher.done():
                        stop_watcher.cancel()
                        try:
                            await stop_watcher
                        except asyncio.CancelledError:
                            pass
                    astr_event.set_extra("agent_user_aborted", True)
                    astr_event.set_extra("agent_stop_requested", False)
                    return

                if _should_stop_agent(astr_event):
                    continue

                if resp.type == "tool_call_result":
                    msg_chain = resp.data["chain"]

                    astr_event.trace.record(
                        "agent_tool_result",
                        tool_result=msg_chain.get_plain_text(
                            with_other_comps_mark=True
                        ),
                    )

                    if msg_chain.type == "tool_direct_result":
                        # tool_direct_result 用于标记 llm tool 需要直接发送给用户的内容
                        await astr_event.send(msg_chain)
                        continue
                    if astr_event.get_platform_id() == "webchat":
                        await astr_event.send(msg_chain)
                    elif show_tool_use and show_tool_call_result:
                        status_msg = _build_tool_result_status_message(
                            msg_chain, tool_name_by_call_id
                        )
                        await astr_event.send(
                            MessageChain(type="tool_call").message(status_msg)
                        )
                    # 对于其他情况，暂时先不处理
                    continue
                elif resp.type == "tool_call":
                    if agent_runner.streaming:
                        # 用来标记流式响应需要分节
                        yield MessageChain(chain=[], type="break")

                    tool_info = _extract_chain_json_data(resp.data["chain"])
                    astr_event.trace.record(
                        "agent_tool_call",
                        tool_name=tool_info if tool_info else "unknown",
                    )
                    _record_tool_call_name(tool_info, tool_name_by_call_id)

                    if astr_event.get_platform_name() == "webchat":
                        await astr_event.send(resp.data["chain"])
                    elif show_tool_use:
                        if show_tool_call_result and isinstance(tool_info, dict):
                            # Delay tool status notification until tool_call_result.
                            continue
                        chain = MessageChain(type="tool_call").message(
                            _build_tool_call_status_message(tool_info)
                        )
                        await astr_event.send(chain)
                    continue

                if stream_to_general and resp.type == "streaming_delta":
                    continue

                if stream_to_general or not agent_runner.streaming:
                    content_typ = (
                        ResultContentType.LLM_RESULT
                        if resp.type == "llm_result"
                        else ResultContentType.GENERAL_RESULT
                    )
                    astr_event.set_result(
                        MessageEventResult(
                            chain=resp.data["chain"].chain,
                            result_content_type=content_typ,
                        ),
                    )
                    yield
                    astr_event.clear_result()
                elif resp.type == "streaming_delta":
                    chain = resp.data["chain"]
                    if chain.type == "reasoning" and not show_reasoning:
                        # display the reasoning content only when configured
                        continue
                    yield resp.data["chain"]  # MessageChain
            if not stop_watcher.done():
                stop_watcher.cancel()
                try:
                    await stop_watcher
                except asyncio.CancelledError:
                    pass
            if agent_runner.done():
                # send agent stats to webchat
                if astr_event.get_platform_name() == "webchat":
                    await astr_event.send(
                        MessageChain(
                            type="agent_stats",
                            chain=[Json(data=agent_runner.stats.to_dict())],
                        )
                    )

                break

        except Exception as e:
            if "stop_watcher" in locals() and not stop_watcher.done():
                stop_watcher.cancel()
                try:
                    await stop_watcher
                except asyncio.CancelledError:
                    pass
            logger.error(traceback.format_exc())

            custom_error_message = extract_persona_custom_error_message_from_event(
                astr_event
            )
            if custom_error_message:
                err_msg = custom_error_message
            else:
                err_msg = (
                    f"Error occurred during AI execution.\n"
                    f"Error Type: {type(e).__name__}\n"
                    f"Error Message: {str(e)}"
                )

            error_llm_response = LLMResponse(
                role="err",
                completion_text=err_msg,
            )
            try:
                await agent_runner.agent_hooks.on_agent_done(
                    agent_runner.run_context, error_llm_response
                )
            except Exception:
                logger.exception("Error in on_agent_done hook")

            if agent_runner.streaming:
                yield MessageChain().message(err_msg)
            else:
                astr_event.set_result(MessageEventResult().message(err_msg))
            return


async def _watch_agent_stop_signal(agent_runner: AgentRunner, astr_event) -> None:
    while not agent_runner.done():
        if _should_stop_agent(astr_event):
            agent_runner.request_stop()
            return
        await asyncio.sleep(0.5)


async def run_live_agent(
    agent_runner: AgentRunner,
    tts_provider: TTSProvider | None = None,
    max_step: int = 30,
    show_tool_use: bool = True,
    show_tool_call_result: bool = False,
    show_reasoning: bool = False,
) -> AsyncGenerator[MessageChain | None, None]:
    """Live Mode 的 Agent 运行器，支持流式 TTS

    Args:
        agent_runner: Agent 运行器
        tts_provider: TTS Provider 实例
        max_step: 最大步数
        show_tool_use: 是否显示工具使用
        show_tool_call_result: 是否显示工具返回结果
        show_reasoning: 是否显示推理过程

    Yields:
        MessageChain: 包含文本或音频数据的消息链
    """
    # 如果没有 TTS Provider，直接发送文本
    if not tts_provider:
        async for chain in run_agent(
            agent_runner,
            max_step=max_step,
            show_tool_use=show_tool_use,
            show_tool_call_result=show_tool_call_result,
            stream_to_general=False,
            show_reasoning=show_reasoning,
        ):
            yield chain
        return

    support_stream = tts_provider.support_stream()
    if support_stream:
        logger.info("[Live Agent] 使用流式 TTS（原生支持 get_audio_stream）")
    else:
        logger.info(
            f"[Live Agent] 使用 TTS（{tts_provider.meta().type} "
            "使用 get_audio，将按句子分块生成音频）"
        )

    # 统计数据初始化
    tts_start_time = time.time()
    tts_first_frame_time = 0.0
    first_chunk_received = False

    # 创建队列
    text_queue: asyncio.Queue[str | None] = asyncio.Queue()
    # audio_queue stored bytes or (text, bytes)
    audio_queue: asyncio.Queue[bytes | tuple[str, bytes] | None] = asyncio.Queue()

    # 1. 启动 Agent Feeder 任务：负责运行 Agent 并将文本分句喂给 text_queue
    feeder_task = asyncio.create_task(
        _run_agent_feeder(
            agent_runner,
            text_queue,
            max_step,
            show_tool_use,
            show_tool_call_result,
            show_reasoning,
        )
    )

    # 2. 启动 TTS 任务：负责从 text_queue 读取文本并生成音频到 audio_queue
    if support_stream:
        tts_task = asyncio.create_task(
            _safe_tts_stream_wrapper(tts_provider, text_queue, audio_queue)
        )
    else:
        tts_task = asyncio.create_task(
            _simulated_stream_tts(tts_provider, text_queue, audio_queue)
        )

    # 3. 主循环：从 audio_queue 读取音频并 yield
    try:
        while True:
            queue_item = await audio_queue.get()

            if queue_item is None:
                break

            text = None
            if isinstance(queue_item, tuple):
                text, audio_data = queue_item
            else:
                audio_data = queue_item

            if not first_chunk_received:
                # 记录首帧延迟（从开始处理到收到第一个音频块）
                tts_first_frame_time = time.time() - tts_start_time
                first_chunk_received = True

            # 将音频数据封装为 MessageChain
            import base64

            audio_b64 = base64.b64encode(audio_data).decode("utf-8")
            comps: list[BaseMessageComponent] = [Plain(audio_b64)]
            if text:
                comps.append(Json(data={"text": text}))
            chain = MessageChain(chain=comps, type="audio_chunk")
            yield chain

    except Exception as e:
        logger.error(f"[Live Agent] 运行时发生错误: {e}", exc_info=True)
    finally:
        # 清理任务
        if not feeder_task.done():
            feeder_task.cancel()
        if not tts_task.done():
            tts_task.cancel()

        # 确保队列被消费
        pass

    tts_end_time = time.time()

    # 发送 TTS 统计信息
    try:
        astr_event = agent_runner.run_context.context.event
        if astr_event.get_platform_name() == "webchat":
            tts_duration = tts_end_time - tts_start_time
            await astr_event.send(
                MessageChain(
                    type="tts_stats",
                    chain=[
                        Json(
                            data={
                                "tts_total_time": tts_duration,
                                "tts_first_frame_time": tts_first_frame_time,
                                "tts": tts_provider.meta().type,
                                "chat_model": agent_runner.provider.get_model(),
                            }
                        )
                    ],
                )
            )
    except Exception as e:
        logger.error(f"发送 TTS 统计信息失败: {e}")


async def _run_agent_feeder(
    agent_runner: AgentRunner,
    text_queue: asyncio.Queue,
    max_step: int,
    show_tool_use: bool,
    show_tool_call_result: bool,
    show_reasoning: bool,
) -> None:
    """运行 Agent 并将文本输出分句放入队列"""
    buffer = ""
    try:
        async for chain in run_agent(
            agent_runner,
            max_step=max_step,
            show_tool_use=show_tool_use,
            show_tool_call_result=show_tool_call_result,
            stream_to_general=False,
            show_reasoning=show_reasoning,
        ):
            if chain is None:
                continue

            # 提取文本
            text = chain.get_plain_text()
            if text:
                buffer += text

                # 分句逻辑：匹配标点符号
                # r"([.。!！?？\n]+)" 会保留分隔符
                parts = re.split(r"([.。!！?？\n]+)", buffer)

                if len(parts) > 1:
                    # 处理完整的句子
                    # range step 2 因为 split 后是 [text, delim, text, delim, ...]
                    temp_buffer = ""
                    for i in range(0, len(parts) - 1, 2):
                        sentence = parts[i]
                        delim = parts[i + 1]
                        full_sentence = sentence + delim
                        temp_buffer += full_sentence

                        if len(temp_buffer) >= 10:
                            if temp_buffer.strip():
                                logger.info(f"[Live Agent Feeder] 分句: {temp_buffer}")
                                await text_queue.put(temp_buffer)
                            temp_buffer = ""

                    # 更新 buffer 为剩余部分
                    buffer = temp_buffer + parts[-1]

        # 处理剩余 buffer
        if buffer.strip():
            await text_queue.put(buffer)

    except Exception as e:
        logger.error(f"[Live Agent Feeder] Error: {e}", exc_info=True)
    finally:
        # 发送结束信号
        await text_queue.put(None)


async def _safe_tts_stream_wrapper(
    tts_provider: TTSProvider,
    text_queue: asyncio.Queue[str | None],
    audio_queue: "asyncio.Queue[bytes | tuple[str, bytes] | None]",
) -> None:
    """包装原生流式 TTS 确保异常处理和队列关闭"""
    try:
        await tts_provider.get_audio_stream(text_queue, audio_queue)
    except Exception as e:
        logger.error(f"[Live TTS Stream] Error: {e}", exc_info=True)
    finally:
        await audio_queue.put(None)


async def _simulated_stream_tts(
    tts_provider: TTSProvider,
    text_queue: asyncio.Queue[str | None],
    audio_queue: "asyncio.Queue[bytes | tuple[str, bytes] | None]",
) -> None:
    """模拟流式 TTS 分句生成音频"""
    try:
        while True:
            text = await text_queue.get()
            if text is None:
                break

            try:
                audio_path = await tts_provider.get_audio(text)

                if audio_path:
                    with open(audio_path, "rb") as f:
                        audio_data = f.read()
                    await audio_queue.put((text, audio_data))
            except Exception as e:
                logger.error(
                    f"[Live TTS Simulated] Error processing text '{text[:20]}...': {e}"
                )
                # 继续处理下一句

    except Exception as e:
        logger.error(f"[Live TTS Simulated] Critical Error: {e}", exc_info=True)
    finally:
        await audio_queue.put(None)