import { generateText } from "ai";
import { execFile } from "node:child_process";
import { promisify } from "node:util";
import { mkdtemp, readFile, rm } from "node:fs/promises";
import os from "node:os";
import path from "node:path";
import { NextResponse } from "next/server";
import sharp from "sharp";
import { z } from "zod";
import { auth } from "@/app/(auth)/auth";
import { getLanguageModelById } from "@/lib/ai/dynamic-provider";
import { inlineVisionImages } from "@/lib/ai/vision-inline";
import { getAiModelAgentCodingModelId, isAiModelMultimodal } from "@/lib/db/ai-model-flags";
import {
  getAgentAiModelWithProvider,
  getEnabledAiModelsWithProvider,
} from "@/lib/db/queries";

export const runtime = "nodejs";
export const maxDuration = 180;

const execFileAsync = promisify(execFile);
const VIDEO_ANALYSIS_FRAME_COUNT = 4;
const VIDEO_ANALYSIS_MAX_SIDE = 960;
const VIDEO_ANALYSIS_TIMEOUT_MS = 15000;

const requestSchema = z.object({
  mode: z.enum(["video-analysis", "character-scene-extraction", "agent-plan"]),
  prompt: z.string().trim().optional(),
  imageUrls: z.array(z.string().trim()).optional(),
  videoUrls: z.array(z.string().trim()).optional(),
  nodeTitle: z.string().trim().optional(),
  preferredModelRef: z.string().trim().optional(),
  model: z.string().trim().optional(),
});

function sanitizeUrls(value: unknown, limit = 6) {
  if (!Array.isArray(value)) return [];
  return value
    .map((item) => String(item || "").trim())
    .filter((url) => /^https?:\/\//i.test(url))
    .slice(0, limit);
}

function supportsVision(model: {
  supportsVision?: boolean | null;
  capabilities?: unknown;
  defaultParams?: unknown;
}) {
  return (
    Boolean(model.supportsVision) ||
    isAiModelMultimodal({
      capabilities: model.capabilities,
      defaultParams: model.defaultParams,
    })
  );
}

function extractJsonBlock<T = Record<string, unknown>>(text: string): T | null {
  const matched = String(text || "").match(/\{[\s\S]*\}/);
  if (!matched) return null;
  try {
    return JSON.parse(matched[0]) as T;
  } catch {
    return null;
  }
}

function normalizeText(value: unknown) {
  return String(value || "").trim();
}

function normalizeStringList(value: unknown, limit = 8) {
  if (!Array.isArray(value)) return [];
  return Array.from(
    new Set(
      value
        .map((item) => String(item || "").trim())
        .filter(Boolean)
        .slice(0, limit)
    )
  );
}

function normalizeShotIdeas(value: unknown) {
  if (!Array.isArray(value)) return [];
  return value
    .map((item, index) => {
      const raw =
        item && typeof item === "object" && !Array.isArray(item)
          ? (item as Record<string, unknown>)
          : {};
      const title = normalizeText(raw.title || `镜头 ${index + 1}`);
      const description = normalizeText(raw.description || raw.prompt || raw.summary);
      if (!title && !description) return null;
      return {
        title: title || `镜头 ${index + 1}`,
        description,
      };
    })
    .filter(
      (item): item is { title: string; description: string } => Boolean(item)
    )
    .slice(0, 6);
}

function normalizeVideoAnalysisRows(
  value: unknown,
  framePreviews: string[],
) {
  if (!Array.isArray(value)) return [];
  return value
    .map((item, index) => {
      const raw =
        item && typeof item === "object" && !Array.isArray(item)
          ? (item as Record<string, unknown>)
          : {};
      const sceneDescription = normalizeText(
        raw.sceneDescription || raw.scene || raw.title || raw.description,
      );
      const backgroundMusic = normalizeText(
        raw.backgroundMusic || raw.music || raw.bgm,
      );
      const voiceAndSfx = normalizeText(
        raw.voiceAndSfx || raw.sound || raw.audio || raw.sfx,
      );
      const imagePrompt = normalizeText(
        raw.imagePrompt || raw.image_generation_prompt || raw.image,
      );
      const motionPrompt = normalizeText(
        raw.motionPrompt || raw.motion || raw.videoPrompt || raw.video_motion_prompt,
      );
      const keyframeIndex = Number(
        raw.keyframeIndex || raw.frameIndex || raw.key_frame_index || index + 1,
      );
      if (
        !sceneDescription &&
        !backgroundMusic &&
        !voiceAndSfx &&
        !imagePrompt &&
        !motionPrompt
      ) {
        return null;
      }
      const normalizedFrameIndex =
        Number.isFinite(keyframeIndex) && keyframeIndex > 0
          ? Math.max(0, Math.min(framePreviews.length - 1, Math.round(keyframeIndex) - 1))
          : index % Math.max(framePreviews.length, 1);
      return {
        sceneDescription: sceneDescription || `镜头 ${index + 1}`,
        backgroundMusic: backgroundMusic || "无",
        voiceAndSfx: voiceAndSfx || "无",
        imagePrompt: imagePrompt || sceneDescription || "延续当前视频气质生成关键帧",
        motionPrompt: motionPrompt || "保持镜头运动自然、连贯、符合前后节奏。",
        keyframeUrl: framePreviews[normalizedFrameIndex] || "",
      };
    })
    .filter(
      (
        item,
      ): item is {
        sceneDescription: string;
        backgroundMusic: string;
        voiceAndSfx: string;
        imagePrompt: string;
        motionPrompt: string;
        keyframeUrl: string;
      } => Boolean(item),
    )
    .slice(0, 8);
}

function normalizeCharacters(value: unknown) {
  if (!Array.isArray(value)) return [];
  return value
    .map((item) => {
      const raw =
        item && typeof item === "object" && !Array.isArray(item)
          ? (item as Record<string, unknown>)
          : {};
      const name = normalizeText(raw.name || raw.title);
      const description = normalizeText(raw.description || raw.summary);
      const prompt = normalizeText(raw.prompt);
      if (!name && !description && !prompt) return null;
      return {
        name: name || "未命名角色",
        description,
        prompt,
        look: normalizeText(raw.look) || undefined,
        wardrobe: normalizeText(raw.wardrobe) || undefined,
        signatureAction: normalizeText(raw.signatureAction) || undefined,
        traits: normalizeStringList(raw.traits, 6),
      };
    })
    .filter((item): item is NonNullable<typeof item> => Boolean(item))
    .slice(0, 8);
}

function normalizeScenes(value: unknown) {
  if (!Array.isArray(value)) return [];
  return value
    .map((item) => {
      const raw =
        item && typeof item === "object" && !Array.isArray(item)
          ? (item as Record<string, unknown>)
          : {};
      const name = normalizeText(raw.name || raw.title);
      const description = normalizeText(raw.description || raw.summary);
      const prompt = normalizeText(raw.prompt);
      if (!name && !description && !prompt) return null;
      return {
        name: name || "未命名场景",
        description,
        prompt,
        mood: normalizeText(raw.mood) || undefined,
        lighting: normalizeText(raw.lighting) || undefined,
        props: normalizeStringList(raw.props, 6),
      };
    })
    .filter((item): item is NonNullable<typeof item> => Boolean(item))
    .slice(0, 8);
}

function buildVideoAnalysisPrompt(params: {
  nodeTitle?: string;
  prompt: string;
  frameCount: number;
  imageCount: number;
  videoCount: number;
}) {
  return [
    "你是视频画布工作流里的视频分析节点，需要为后续分镜和生成节点提供结构化输出。",
    params.nodeTitle ? `当前节点：${params.nodeTitle}` : "",
    `上游参考图数量：${params.imageCount}`,
    `上游视频数量：${params.videoCount}`,
    `已提供按时间顺序抽取的视频关键帧数量：${params.frameCount}`,
    params.prompt ? `上游脚本/说明：\n${params.prompt}` : "上游没有额外脚本，请重点依据参考画面分析。",
    "",
    "请只输出 JSON，不要输出 markdown，不要输出解释。",
    "JSON 结构如下：",
    `{
  "summary": "一句话总结当前视频/镜头风格，不超过120字",
  "highlights": ["3-5条镜头分析要点"],
  "analysisRows": [
    {
      "sceneDescription": "镜头内容与景别描述",
      "backgroundMusic": "这一段适合的背景音乐氛围",
      "voiceAndSfx": "人声/环境声/音效建议，没有就写无",
      "imagePrompt": "后续生成关键帧时可直接使用的图像提示词",
      "motionPrompt": "后续生成视频时可直接使用的运动提示词",
      "keyframeIndex": 1
    }
  ],
  "shotIdeas": [
    { "title": "镜头 1", "description": "一句话描述镜头目标、景别或动作" }
  ],
  "visualMotifs": ["2-4个可复用的视觉母题"],
  "recommendedNextSteps": ["2-4条建议的下游动作"]
}`,
    "analysisRows 请尽量输出 2-6 行；keyframeIndex 使用 1 开始的序号，对应你看到的关键帧顺序。",
    "如果没有真实视频帧，只能依据参考图和脚本推断，也要正常输出，但措辞要稳健。",
  ]
    .filter(Boolean)
    .join("\n");
}

async function getVideoDurationSeconds(videoUrl: string) {
  try {
    const { stdout } = await execFileAsync(
      "ffprobe",
      [
        "-v",
        "error",
        "-show_entries",
        "format=duration",
        "-of",
        "default=noprint_wrappers=1:nokey=1",
        videoUrl,
      ],
      {
        timeout: VIDEO_ANALYSIS_TIMEOUT_MS,
        maxBuffer: 1024 * 1024,
      },
    );
    const duration = Number.parseFloat(String(stdout || "").trim());
    return Number.isFinite(duration) && duration > 0 ? duration : null;
  } catch {
    return null;
  }
}

function buildVideoFrameTimestamps(durationSeconds: number | null, frameCount: number) {
  const safeCount = Math.max(1, frameCount);
  if (!durationSeconds || !Number.isFinite(durationSeconds) || durationSeconds <= 0.5) {
    return Array.from({ length: safeCount }).map((_, index) =>
      Number((0.3 + index * 0.45).toFixed(2)),
    );
  }
  const start = Math.min(0.35, durationSeconds * 0.12);
  const end = Math.max(start + 0.6, durationSeconds * 0.88);
  if (safeCount === 1) return [Number(((start + end) / 2).toFixed(2))];
  const step = (end - start) / (safeCount - 1);
  return Array.from({ length: safeCount }).map((_, index) =>
    Number((start + step * index).toFixed(2)),
  );
}

async function downscaleVideoFrameForVision(inputBuffer: Buffer) {
  return sharp(inputBuffer, { failOn: "none" })
    .rotate()
    .resize({
      width: VIDEO_ANALYSIS_MAX_SIDE,
      height: VIDEO_ANALYSIS_MAX_SIDE,
      fit: "inside",
      withoutEnlargement: true,
    })
    .jpeg({ quality: 82, mozjpeg: true })
    .toBuffer();
}

async function extractVideoFrameDataUrls(videoUrl: string, frameCount = VIDEO_ANALYSIS_FRAME_COUNT) {
  const tempDir = await mkdtemp(path.join(os.tmpdir(), "canvas-video-analysis-"));
  try {
    const durationSeconds = await getVideoDurationSeconds(videoUrl);
    const timestamps = buildVideoFrameTimestamps(durationSeconds, frameCount);
    const frames = await Promise.all(
      timestamps.map(async (timestamp, index) => {
        const outputPath = path.join(tempDir, `frame-${index + 1}.jpg`);
        await execFileAsync(
          "ffmpeg",
          [
            "-hide_banner",
            "-loglevel",
            "error",
            "-ss",
            String(timestamp),
            "-i",
            videoUrl,
            "-frames:v",
            "1",
            "-q:v",
            "4",
            "-y",
            outputPath,
          ],
          {
            timeout: VIDEO_ANALYSIS_TIMEOUT_MS,
            maxBuffer: 2 * 1024 * 1024,
          },
        );
        const buffer = await readFile(outputPath);
        const processed = await downscaleVideoFrameForVision(buffer);
        return `data:image/jpeg;base64,${processed.toString("base64")}`;
      }),
    );
    return frames.filter(Boolean);
  } catch (error) {
    console.warn("[Canvas Workflow Analyze] extract video frames failed:", {
      message: error instanceof Error ? error.message : String(error),
      videoUrl,
    });
    return [];
  } finally {
    await rm(tempDir, { force: true, recursive: true }).catch(() => undefined);
  }
}

function buildCharacterScenePrompt(params: {
  nodeTitle?: string;
  prompt: string;
  imageCount: number;
}) {
  return [
    "你是视频画布工作流里的角色场景提取节点，需要从脚本、设定和参考图里提取角色与场景资产。",
    params.nodeTitle ? `当前节点：${params.nodeTitle}` : "",
    `参考图数量：${params.imageCount}`,
    params.prompt ? `输入内容：\n${params.prompt}` : "没有额外脚本，请根据参考图提取角色和场景线索。",
    "",
    "请只输出 JSON，不要输出 markdown，不要输出解释。",
    "JSON 结构如下：",
    `{
  "summary": "一句话总结当前故事世界与资产线索",
  "characters": [
    {
      "name": "角色名",
      "description": "角色外观与身份简介",
      "look": "外貌特征",
      "wardrobe": "服装/道具",
      "traits": ["性格关键词"],
      "signatureAction": "代表动作",
      "prompt": "可直接用于生成角色图/角色视频的提示词"
    }
  ],
  "scenes": [
    {
      "name": "场景名",
      "description": "空间、布景和氛围简介",
      "mood": "情绪与气氛",
      "lighting": "布光特征",
      "props": ["核心道具"],
      "prompt": "可直接用于生成场景图/场景视频的提示词"
    }
  ],
  "relationships": ["角色与场景或角色之间的关系"],
  "recommendedNextSteps": ["建议连接的后续节点动作"]
}`,
    "角色和场景都要尽量具体，避免空泛形容词。",
  ]
    .filter(Boolean)
    .join("\n");
}

function buildAgentPlanPrompt(params: {
  nodeTitle?: string;
  prompt: string;
  imageCount: number;
  videoCount: number;
}) {
  return [
    "你是视频画布里的自定义代理节点，需要把当前创作目标拆成可执行的工作流计划。",
    params.nodeTitle ? `当前节点：${params.nodeTitle}` : "",
    `参考图数量：${params.imageCount}`,
    `视频数量：${params.videoCount}`,
    params.prompt ? `上游需求：\n${params.prompt}` : "没有额外脚本，请基于当前上下文给出稳妥计划。",
    "",
    "请只输出 JSON，不要输出 markdown，不要输出解释。",
    "JSON 结构如下：",
    `{
  "summary": "一句话说明这个代理链的目标",
  "agentPlan": ["3-6个执行步骤"],
  "deliverables": ["2-4个预期交付物"],
  "recommendedNodes": ["建议继续连接的节点类型或能力名"]
}`,
    "步骤要偏工作流视角，而不是泛泛而谈。",
  ]
    .filter(Boolean)
    .join("\n");
}

export async function POST(request: Request) {
  try {
    const session = await auth();
    if (!session?.user?.id) {
      return NextResponse.json({ success: false, error: "未登录" }, { status: 401 });
    }

    const body = await request.json().catch(() => ({}));
    const parsed = requestSchema.safeParse(body);
    if (!parsed.success) {
      return NextResponse.json(
        { success: false, error: "请求参数不合法" },
        { status: 400 }
      );
    }

    const prompt = normalizeText(parsed.data.prompt);
    const imageUrls = sanitizeUrls(parsed.data.imageUrls, 4);
    const videoUrls = sanitizeUrls(parsed.data.videoUrls, 4);
    const extractedVideoFrames =
      parsed.data.mode === "video-analysis" && videoUrls.length > 0
        ? await extractVideoFrameDataUrls(videoUrls[0], VIDEO_ANALYSIS_FRAME_COUNT)
        : [];
    const preferredModelRef = normalizeText(parsed.data.preferredModelRef || parsed.data.model);
    if (!prompt && imageUrls.length === 0 && videoUrls.length === 0) {
      return NextResponse.json(
        { success: false, error: "请至少提供脚本、图片或视频上下文" },
        { status: 400 }
      );
    }

    const [allModels, agentModel] = await Promise.all([
      getEnabledAiModelsWithProvider(false),
      getAgentAiModelWithProvider(),
    ]);
    const compatibleModels = allModels.filter(
      (item) => !item.model.isSubModel && item.model.isEnabled
    );

    const preferredModel =
      preferredModelRef
        ? compatibleModels.find(
            (item) =>
              item.model.id === preferredModelRef ||
              item.model.modelId === preferredModelRef,
          )
        : null;

    const hasVisionInputs = imageUrls.length > 0 || extractedVideoFrames.length > 0;

    let selectedModel =
      preferredModel && (!hasVisionInputs || supportsVision(preferredModel.model))
        ? preferredModel
        : hasVisionInputs &&
            agentModel &&
            supportsVision(agentModel.model)
          ? agentModel
          : hasVisionInputs
            ? compatibleModels.find((item) => supportsVision(item.model))
            : agentModel || compatibleModels[0];

    if (!selectedModel) {
      selectedModel = compatibleModels[0];
    }

    if (!selectedModel) {
      return NextResponse.json(
        { success: false, error: "当前站点没有可用的分析模型" },
        { status: 500 }
      );
    }

    const useVision = hasVisionInputs && supportsVision(selectedModel.model);
    const codingModelId = getAiModelAgentCodingModelId(selectedModel.model);
    const model = await getLanguageModelById(selectedModel.model.id, codingModelId);
    if (!model) {
      return NextResponse.json(
        { success: false, error: "分析模型初始化失败" },
        { status: 500 }
      );
    }

    const promptText =
      parsed.data.mode === "video-analysis"
        ? buildVideoAnalysisPrompt({
            frameCount: extractedVideoFrames.length,
            nodeTitle: normalizeText(parsed.data.nodeTitle),
            prompt,
            imageCount: imageUrls.length,
            videoCount: videoUrls.length,
          })
        : parsed.data.mode === "character-scene-extraction"
          ? buildCharacterScenePrompt({
              nodeTitle: normalizeText(parsed.data.nodeTitle),
              prompt,
              imageCount: imageUrls.length,
            })
          : buildAgentPlanPrompt({
              nodeTitle: normalizeText(parsed.data.nodeTitle),
              prompt,
              imageCount: imageUrls.length,
              videoCount: videoUrls.length,
            });

    const inlinedImageReferences = useVision
      ? (
          await inlineVisionImages(imageUrls, {
            maxImages: 4,
            maxSide: 960,
            timeoutMs: 12000,
            retries: 1,
            loggerPrefix: "[Canvas Workflow Analyze]",
          })
        ).inlineDataUrls
      : [];
    const referenceImages = useVision
      ? [...inlinedImageReferences, ...extractedVideoFrames].slice(0, 6)
      : [];

    const { text } = await generateText({
      model,
      messages: [
        {
          role: "user",
          content: [
            { type: "text", text: promptText },
            ...referenceImages.map((image) => ({ type: "image" as const, image })),
          ],
        },
      ],
      temperature: 0.2,
      maxTokens: 1600,
    } as any);

    const parsedJson = extractJsonBlock<Record<string, unknown>>(text) || {};
    const summary =
      normalizeText(parsedJson.summary) ||
      normalizeText(text).slice(0, 160);

    if (parsed.data.mode === "video-analysis") {
      const analysisRows = normalizeVideoAnalysisRows(
        parsedJson.analysisRows,
        extractedVideoFrames,
      );
      const shotIdeas =
        normalizeShotIdeas(parsedJson.shotIdeas).length > 0
          ? normalizeShotIdeas(parsedJson.shotIdeas)
          : analysisRows.map((row, index) => ({
              title: `镜头 ${index + 1}`,
              description: row.sceneDescription,
            }));
      return NextResponse.json({
        success: true,
        data: {
          summary,
          analysisRows,
          framePreviews: extractedVideoFrames,
          highlights: normalizeStringList(parsedJson.highlights, 6),
          shotIdeas,
          visualMotifs: normalizeStringList(parsedJson.visualMotifs, 6),
          recommendedNextSteps: normalizeStringList(parsedJson.recommendedNextSteps, 6),
          model: selectedModel.model.displayName,
          raw: text,
        },
      });
    }

    if (parsed.data.mode === "character-scene-extraction") {
      return NextResponse.json({
        success: true,
        data: {
          summary,
          characters: normalizeCharacters(parsedJson.characters),
          scenes: normalizeScenes(parsedJson.scenes),
          relationships: normalizeStringList(parsedJson.relationships, 8),
          recommendedNextSteps: normalizeStringList(parsedJson.recommendedNextSteps, 6),
          model: selectedModel.model.displayName,
          raw: text,
        },
      });
    }

    return NextResponse.json({
      success: true,
      data: {
        summary,
        agentPlan: normalizeStringList(parsedJson.agentPlan, 8),
        deliverables: normalizeStringList(parsedJson.deliverables, 6),
        recommendedNodes: normalizeStringList(parsedJson.recommendedNodes, 6),
        model: selectedModel.model.displayName,
        raw: text,
      },
    });
  } catch (error) {
    console.error("[Canvas Workflow Analyze] Error:", error);
    return NextResponse.json(
      {
        success: false,
        error: error instanceof Error ? error.message : "工作流分析失败",
      },
      { status: 500 }
    );
  }
}