nuxt-app/server/api/agent/chat/index.post.ts

import { defineEventHandler, getQuery, getRouterParam, readBody, setResponseHeaders } from "h3";
import { R } from "#server/utils/response";
import { getCurrentUser, getConfigGlobal, getConfigUser } from "#server/utils/context";
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { type LanguageModel, streamText, stepCountIs } from "ai";
import { getSessionByIdAndUser, getMessagesBySession, getMaxSortOrder, saveMessage, truncateMessagesAfter, deleteMessage, getMessageById, updateMessageContent, touchSession, countAssistantMessages } from "#server/service/agent/session";
import { getAgentToolsForChat } from "#server/service/agent-tool";
import { getTempTokenFromCookie } from "#server/service/agent/temp-token";
import { checkRateLimit, incrementRateLimit } from "#server/service/agent/rate-limit";
import { getModelWithProviderById, getSystemModelWithProviderById } from "#server/service/llm";
import { generateSessionTitle } from "#server/service/agent/title";
import log4js from "logger";

const logger = log4js.getLogger("APP");

function resolveModel(
  provider: {
    name: string;
    apiKey: string | null;
    baseUrl: string | null;
    parseMode: string;
  },
  modelId: string,
): LanguageModel {
  const baseUrl = provider.baseUrl?.replace(/\/+$/, "") || undefined;
  if (provider.parseMode === "anthropic") {
    throw createError({
      statusCode: 400,
      statusMessage: "Anthropic 解析模式暂不支持流式对话，请使用 OpenAI 兼容模式",
    });
  }
  const openaiCompatible = createOpenAICompatible({
    name: provider.name,
    apiKey: provider.apiKey || undefined,
    baseURL: baseUrl || "https://api.openai.com/v1",
  });
  return openaiCompatible(modelId) as LanguageModel;
}

export default defineEventHandler(async (event) => {
  const user = await getCurrentUser(event);
  const tempToken = getTempTokenFromCookie(event);

  if (!user && !tempToken) {
    throw createError({ statusCode: 401, statusMessage: "请先登录或创建临时会话" });
  }

  const body = await readBody(event);
  const { sessionId, content, editMessageId, regenerate } = body as {
    sessionId: string;
    content: string;
    editMessageId?: string;
    regenerate?: boolean;
  };

  if (!sessionId || !content || typeof content !== "string" || content.trim().length === 0) {
    throw createError({ statusCode: 400, statusMessage: "参数无效" });
  }

  const session = await getSessionByIdAndUser(sessionId, user?.id ?? null, tempToken);
  if (!session) {
    throw createError({ statusCode: 404, statusMessage: "会话不存在" });
  }

  const ip = getRequestIP(event, { xForwardedFor: true }) || "unknown";

  if (!user) {
    const rateLimit = checkRateLimit(sessionId, ip);
    if (rateLimit.blocked) {
      setResponseHeaders(event, {
        "X-RateLimit-Session-Remaining": String(rateLimit.sessionRemaining),
        "X-RateLimit-Ip-Remaining": String(rateLimit.ipRemaining),
      });
      throw createError({
        statusCode: 429,
        statusMessage: rateLimit.sessionRemaining === 0 ? "当前会话已达回复上限，请登录后继续" : "今日回复次数已达上限，请登录或明日再试",
      });
    }
    setResponseHeaders(event, {
      "X-RateLimit-Session-Remaining": String(rateLimit.sessionRemaining),
      "X-RateLimit-Ip-Remaining": String(rateLimit.ipRemaining),
    });
  }

  let modelRow;
  if (user) {
    const preferredModelId = await getConfigUser(event, "preferredLlmModelId");
    const targetModelId = session.modelId ?? preferredModelId;
    if (!targetModelId) {
      throw createError({ statusCode: 400, statusMessage: "未配置模型，请先在设置中选择模型" });
    }
    modelRow = await getModelWithProviderById(targetModelId, user.id);
  } else {
    if (!session.modelId) {
      throw createError({ statusCode: 400, statusMessage: "系统未配置默认模型" });
    }
    modelRow = await getSystemModelWithProviderById(session.modelId);
  }

  if (!modelRow) {
    throw createError({ statusCode: 404, statusMessage: "模型不存在或无权访问" });
  }

  const { model, provider } = modelRow;
  if (!provider.apiKey) {
    throw createError({ statusCode: 400, statusMessage: "供应商未配置 API Key" });
  }

  const systemPrompt = await getConfigGlobal("agentSystemPrompt");
  const titleModelId = await getConfigGlobal("agentTitleModelId");

  let publicToolSlugs: string[] = [];
  if (!user) {
    const raw = await getConfigGlobal("agentPublicToolSlugs");
    if (raw && Array.isArray(raw)) {
      publicToolSlugs = raw as string[];
    }
  }

  const { tools, approvalConfig } = await getAgentToolsForChat({
    userId: user?.id ?? null,
    userRole: user?.role ?? null,
    enableTools: session.enableTools === 1,
    publicToolSlugs,
  });

  let historyMessages = await getMessagesBySession(sessionId, { limit: 100 });

  let userMessage;
  let shouldGenerateTitle = false;

  if (editMessageId) {
    const editMsg = await getMessageById(editMessageId);
    if (!editMsg || editMsg.sessionId !== sessionId) {
      throw createError({ statusCode: 400, statusMessage: "编辑的消息不存在" });
    }
    await truncateMessagesAfter(sessionId, editMsg.sortOrder);
    await updateMessageContent(editMessageId, content.trim());
    historyMessages = await getMessagesBySession(sessionId, { limit: 100 });
    userMessage = editMsg;
  } else if (regenerate) {
    const lastAssistant = historyMessages.filter((m) => m.role === "assistant").pop();
    if (lastAssistant) {
      await deleteMessage(lastAssistant.id);
      historyMessages = await getMessagesBySession(sessionId, { limit: 100 });
    }
    const lastUser = historyMessages.filter((m) => m.role === "user").pop();
    if (lastUser) {
      await updateMessageContent(lastUser.id, content.trim());
      userMessage = lastUser;
    } else {
      const sortOrder = (await getMaxSortOrder(sessionId)) + 1;
      userMessage = await saveMessage({
        sessionId,
        role: "user",
        content: content.trim(),
        sortOrder,
      });
      shouldGenerateTitle = true;
    }
  } else {
    const sortOrder = (await getMaxSortOrder(sessionId)) + 1;
    userMessage = await saveMessage({
      sessionId,
      role: "user",
      content: content.trim(),
      sortOrder,
    });
    const assistantCount = await countAssistantMessages(sessionId);
    if (assistantCount === 0) {
      shouldGenerateTitle = true;
    }
  }

  const modelMessages = [
    ...historyMessages
      .filter((m) => m.id !== userMessage.id || editMessageId || regenerate)
      .map((m) => ({
        role: m.role as "user" | "assistant",
        content: m.content,
      })),
    { role: "user" as const, content: content.trim() },
  ];

  const languageModel = resolveModel(provider, model.modelId);

  logger.info(
    "[%s] [AGENT-CHAT] userId=%s sessionId=%s modelId=%d thinking=%s tools=%s edit=%s regen=%s",
    event.context.requestId ?? "-",
    user?.id ?? "temp",
    sessionId,
    model.id,
    session.enableThinking === 1 ? "on" : "off",
    Object.keys(tools).length > 0 ? "on" : "off",
    editMessageId ? "yes" : "no",
    regenerate ? "yes" : "no",
  );

  const result = streamText({
    model: languageModel,
    system: systemPrompt || undefined,
    messages: modelMessages,
    maxOutputTokens: model.maxTokens || undefined,
    ...(Object.keys(tools).length > 0
      ? { tools, stopWhen: stepCountIs(8), toolApproval: approvalConfig }
      : {}),
    ...(session.enableThinking === 1
      ? {
          providerOptions: {
            openaiCompatible: { reasoningEffort: "high" },
          },
        }
      : {}),
    onError: (errorData) => {
      const errMsg = errorData?.error instanceof Error
        ? errorData.error.message
        : String(errorData?.error ?? "未知错误");
      logger.error("[%s] [AGENT-CHAT] streamText error: %s", event.context.requestId ?? "-", errMsg);
    },
    onFinish: async ({ finishReason, usage, steps, text: assistantText }) => {
      logger.info(
        "[%s] [AGENT-CHAT] finished: reason=%s steps=%d inputTokens=%d outputTokens=%d",
        event.context.requestId ?? "-",
        finishReason,
        steps.length,
        usage?.inputTokens ?? 0,
        usage?.outputTokens ?? 0,
      );

      const parts: Array<Record<string, unknown>> = [];
      for (const step of steps) {
        for (const r of step.reasoning) {
          if (r.type === "reasoning" && r.text) {
            parts.push({ id: `p_${parts.length}`, type: "reasoning", text: r.text });
          }
        }
        const toolResultMap = new Map(step.toolResults.map((tr) => [tr.toolCallId, tr.output]));
        for (const tc of step.toolCalls) {
          parts.push({
            id: `p_${parts.length}`,
            type: "tool-call",
            toolCallId: tc.toolCallId,
            toolName: tc.toolName,
            args: tc.input,
            result: toolResultMap.get(tc.toolCallId),
            state: "result",
          });
        }
        if (step.text) {
          parts.push({ id: `p_${parts.length}`, type: "text", text: step.text });
        }
      }

      const assistantSortOrder = (await getMaxSortOrder(sessionId)) + 1;
      await saveMessage({
        sessionId,
        role: "assistant",
        content: assistantText,
        parts: parts.length > 0 ? JSON.stringify(parts) : null,
        modelId: model.id,
        inputTokens: usage?.inputTokens ?? null,
        outputTokens: usage?.outputTokens ?? null,
        sortOrder: assistantSortOrder,
      });

      await touchSession(sessionId);

      if (!user) {
        incrementRateLimit(sessionId, ip);
        const rateLimit = checkRateLimit(sessionId, ip);
        setResponseHeaders(event, {
          "X-RateLimit-Session-Remaining": String(rateLimit.sessionRemaining),
          "X-RateLimit-Ip-Remaining": String(rateLimit.ipRemaining),
        });
      }

      if (shouldGenerateTitle && titleModelId && assistantText.trim().length > 0) {
        generateSessionTitle(sessionId, content.trim(), assistantText, titleModelId, user?.id ?? null).catch((e) => { logger.error("[AGENT-CHAT] title gen failed: %s", e?.message ?? e); });
      }
    },
  });

  const response = result.toUIMessageStreamResponse({
    sendReasoning: true,
    headers: {
      "X-User-Message-Id": userMessage.id,
    },
  });

  return response;
});