/**
 * QC Engine — 4-stage AI pipeline.
 * IDENTICAL port from Python qc-report-generator/src/qc/checker.py + src/engines/claude_engine.py
 */

import type Anthropic from "@anthropic-ai/sdk";
import { getClaudeClient } from "../ai/claude-client";
import type { QCIssue, QCLine, QCModule, QCSeverity, QCSummary, QCResult } from "./types";
import { runRuleChecks } from "./rule-checks";
import { runNetflixChecks } from "./netflix-checks";
import { parseNeonStep3, getSourceTextForTimeRange } from "../neon/parser";
import { type KNPData, formatKNPForPrompt } from "./knp-parser";
import {
  QC_SYSTEM_PROMPT,
  VALIDATION_SYSTEM_PROMPT,
  SECOND_PASS_SYSTEM_PROMPT,
  AMAZON_RULES,
  AMAZON_SYSTEM_EXTRA,
  NETFLIX_RULES,
  buildCheckPrompt,
  buildValidationPrompt,
  buildSecondPassPrompt,
} from "./prompts";

const BATCH_SIZE = 15;
const CONTEXT_LINES = 3;
const MAX_RETRIES = 8;
const BACKOFF_BASE_429 = 5;
const BACKOFF_CAP_429 = 60;
const BACKOFF_BASE_ERR = 2;
const BACKOFF_CAP_ERR = 15;
const JITTER_FACTOR = 0.3;

const log = (msg: string) => console.log(`[QC:engine] ${msg}`);

// Global semaphore — max 50 concurrent Claude API calls across all files
const MAX_CONCURRENCY = 50;
let activeCalls = 0;
const waitQueue: Array<() => void> = [];

async function withSemaphore<T>(fn: () => Promise<T>): Promise<T> {
  if (activeCalls >= MAX_CONCURRENCY) {
    await new Promise<void>((resolve) => waitQueue.push(resolve));
  }
  activeCalls++;
  try {
    return await fn();
  } finally {
    activeCalls--;
    const next = waitQueue.shift();
    if (next) next();
  }
}

/** Parse timecode string to milliseconds (supports HH:MM:SS:FF and HH:MM:SS,mmm) */
function tcToMsHelper(tc: string, fps: number = 23.976): number {
  const frameParts = tc.match(/^(\d{2}):(\d{2}):(\d{2}):(\d{2})$/);
  if (frameParts) {
    const [, h, m, s, f] = frameParts.map(Number);
    return Math.round((h * 3600 + m * 60 + s) * 1000 + (f / fps) * 1000);
  }
  const msParts = tc.match(/^(\d{2}):(\d{2}):(\d{2})[.,](\d{3})$/);
  if (msParts) {
    const [, h, m, s, ms] = msParts.map(Number);
    return h * 3600000 + m * 60000 + s * 1000 + ms;
  }
  return 0;
}

// ─── Retry Logic (identical to Python claude_engine.py) ─────────────────────

function isRetryable(err: Error): { retryable: boolean; isRateLimit: boolean } {
  const s = err.message.toLowerCase();
  if (s.includes("429") || s.includes("rate") || s.includes("overloaded"))
    return { retryable: true, isRateLimit: true };
  for (const code of ["500", "502", "503", "504", "529"])
    if (s.includes(code)) return { retryable: true, isRateLimit: false };
  if (s.includes("timeout")) return { retryable: true, isRateLimit: false };
  if (s.includes("could not resolve authentication"))
    return { retryable: true, isRateLimit: true };
  return { retryable: false, isRateLimit: false };
}

function backoffSeconds(attempt: number, isRateLimit: boolean): number {
  const base = isRateLimit
    ? Math.min(BACKOFF_CAP_429, BACKOFF_BASE_429 * 2 ** Math.min(attempt, 6))
    : Math.min(BACKOFF_CAP_ERR, BACKOFF_BASE_ERR * 2 ** Math.min(attempt, 4));
  const jitter = base * JITTER_FACTOR * (2 * Math.random() - 1);
  return Math.max(1, base + jitter);
}

async function callWithRetry(
  client: any, // Anthropic | AnthropicVertex — both have .messages.create()
  system: string,
  user: string,
  label: string,
): Promise<string> {
  let lastError: Error | null = null;
  for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
    try {
      const t0 = Date.now();
      const response: any = await withSemaphore(() => client.messages.create({
        model: "claude-opus-4-6",
        max_tokens: 16384,
        system,
        messages: [{ role: "user", content: user }],
      }));
      const text = response.content
        .filter((b: any) => b.type === "text")
        .map((b: any) => b.text)
        .join("");
      const dur = ((Date.now() - t0) / 1000).toFixed(1);
      log(`${label} — ${dur}s, prompt ${user.length} chars → response ${text.length} chars`);
      return text;
    } catch (err) {
      lastError = err as Error;
      const { retryable, isRateLimit } = isRetryable(lastError);
      if (!retryable || attempt >= MAX_RETRIES) {
        log(`${label} FAILED after ${attempt + 1} attempts: ${lastError.message}`);
        throw lastError;
      }
      const wait = backoffSeconds(attempt, isRateLimit);
      const tag = isRateLimit ? "429 rate limit" : "transient error";
      log(`${tag} on ${label} (attempt ${attempt + 1}/${MAX_RETRIES}): ${lastError.message.substring(0, 80)}... retrying in ${wait.toFixed(1)}s`);
      await new Promise((r) => setTimeout(r, wait * 1000));
    }
  }
  throw lastError!;
}

// ─── JSON Parsing (identical to Python base.py parse_issues / parse_validation_response) ─

function parseIssues(responseText: string, lines: QCLine[]): QCIssue[] {
  let text = responseText.trim();
  if (text.startsWith("```")) {
    text = text.split("\n").filter((l) => !l.trim().startsWith("```")).join("\n");
  }

  const jsonMatch = text.match(/\[[\s\S]*\]/);
  if (!jsonMatch) return [];

  try {
    const raw = JSON.parse(jsonMatch[0]);
    if (!Array.isArray(raw)) return [];

    const lineLookup = new Map(lines.map((l) => [l.id, l.text]));
    return raw
      .filter((r: any) => typeof r === "object" && r.line_id != null)
      .map((r: any) => ({
        lineId: r.line_id,
        severity: (r.severity || "warning").replace("error", "critical") as QCSeverity,
        category: r.category || "grammar",
        originalText: lineLookup.get(r.line_id) || "",
        description: r.description || "",
        suggestedFix: r.suggested_fix || "",
        flaggedText: r.flagged_text || "",
      }));
  } catch {
    return [];
  }
}

function parseValidationResponse(responseText: string): Array<{ issue_index: number; verdict: string; reason?: string; severity?: string }> {
  let text = responseText.trim();
  if (text.startsWith("```")) {
    text = text.split("\n").filter((l) => !l.trim().startsWith("```")).join("\n");
  }
  const jsonMatch = text.match(/\[[\s\S]*\]/);
  if (!jsonMatch) return [];
  try {
    const verdicts = JSON.parse(jsonMatch[0]);
    if (!Array.isArray(verdicts)) return [];
    return verdicts.filter((v: any) => typeof v === "object" && "verdict" in v);
  } catch {
    return [];
  }
}

// ─── Module Config ──────────────────────────────────────────────────────────

function getModuleConfig(module: QCModule) {
  if (module === "amazon") return { rules: AMAZON_RULES, systemExtra: AMAZON_SYSTEM_EXTRA };
  return { rules: NETFLIX_RULES, systemExtra: "" };
}

function sevRank(severity: string): number {
  return { critical: 0, warning: 1, info: 2 }[severity] ?? 1;
}

function buildSummary(issues: QCIssue[], totalLines: number): QCSummary {
  const byCategory: Record<string, number> = {};
  let critical = 0, warnings = 0, info = 0;
  for (const issue of issues) {
    byCategory[issue.category] = (byCategory[issue.category] || 0) + 1;
    if (issue.severity === "critical") critical++;
    else if (issue.severity === "warning") warnings++;
    else info++;
  }
  return { totalLines, totalIssues: issues.length, critical, warnings, info, byCategory };
}

// ─── Pipeline (identical to Python checker.py) ──────────────────────────────

export type LogCallback = (message: string, data?: any) => void;

export async function runQCPipelineWithLogs(
  lines: QCLine[],
  language: string,
  module: QCModule,
  filename: string,
  deliverableType: string = "sub",
  neonSource?: any,
  onLog?: LogCallback,
  knpData?: KNPData,
): Promise<QCResult> {
  return runQCPipeline(lines, language, module, filename, deliverableType, neonSource, onLog, knpData);
}

export async function runQCPipeline(
  lines: QCLine[],
  language: string,
  module: QCModule,
  filename: string,
  deliverableType: string = "sub",
  neonSource?: any,
  onLog?: LogCallback,
  knpData?: KNPData,
): Promise<QCResult> {
  const start = Date.now();
  const emit = (msg: string, data?: any) => {
    log(msg);
    onLog?.(msg, data);
  };

  if (!lines.length) {
    return { filename, issues: [], summary: buildSummary([], 0), module, language, engine: "claude", durationMs: 0 };
  }

  emit(`Pipeline start: ${filename}, ${lines.length} lines, module=${module}, lang=${language}, type=${deliverableType}`);
  const { client, authMethod } = getClaudeClient();
  emit(`Claude auth: ${authMethod}`);
  const { rules, systemExtra } = getModuleConfig(module);

  // ─── Stage 1: AI Batch QC ───
  const batches: QCLine[][] = [];
  for (let i = 0; i < lines.length; i += BATCH_SIZE) {
    batches.push(lines.slice(i, i + BATCH_SIZE));
  }
  emit(`Stage 1: AI Batch QC — ${batches.length} batches of ${BATCH_SIZE}`);
  const s1 = Date.now();

  const batchContexts = batches.map((batch, i) => {
    const startIdx = i * BATCH_SIZE;
    return {
      batch,
      before: lines.slice(Math.max(0, startIdx - CONTEXT_LINES), startIdx),
      after: lines.slice(startIdx + batch.length, Math.min(lines.length, startIdx + batch.length + CONTEXT_LINES)),
    };
  });

  // Parse Neon source for source context (if provided)
  const parsedNeon = neonSource ? parseNeonStep3(neonSource) : null;

  const aiIssues: QCIssue[] = [];
  // All batches in parallel (like Python asyncio.gather)
  const batchResults = await Promise.all(
    batchContexts.map(async ({ batch, before, after }, idx) => {
      const lineRange = `L${batch[0].id}-L${batch[batch.length - 1].id}`;
      const system = systemExtra ? `${QC_SYSTEM_PROMPT}\n\n${systemExtra}` : QC_SYSTEM_PROMPT;

      // Build source context from Neon transcription (time-aligned per line)
      let sourceContext: Array<{ lineId: number; sourceText: string }> | undefined;
      if (parsedNeon && batch[0]?.tcIn) {
        sourceContext = batch
          .filter((l) => l.tcIn && l.tcOut)
          .map((l) => {
            // Parse timecodes — lines may have HH:MM:SS:FF or HH:MM:SS,mmm format
            const tcIn = l.tcIn!;
            const tcOut = l.tcOut!;
            const startMs = tcToMsHelper(tcIn);
            const endMs = tcToMsHelper(tcOut);
            const sourceText = getSourceTextForTimeRange(parsedNeon.transcription, startMs, endMs);
            return sourceText ? { lineId: l.id, sourceText } : null;
          })
          .filter((s): s is { lineId: number; sourceText: string } => s !== null);
      }

      const prompt = buildCheckPrompt(
        batch, language, deliverableType, rules, systemExtra,
        before.length > 0 ? before : undefined,
        after.length > 0 ? after : undefined,
        sourceContext,
      );
      try {
        const text = await callWithRetry(client, system, prompt, `batch ${idx + 1}/${batches.length} ${lineRange}`);
        return parseIssues(text, lines);
      } catch (err) {
        emit(`Batch ${idx + 1} error (continuing): ${(err as Error).message.substring(0, 80)}`);
        return [];
      }
    }),
  );
  for (const r of batchResults) aiIssues.push(...r);
  emit(`Stage 1 done: ${aiIssues.length} AI issues in ${((Date.now() - s1) / 1000).toFixed(1)}s`);

  // ─── Stage 2: Rule-based checks ───
  emit(`Stage 2: Rule-based checks`);
  const s2 = Date.now();
  let ruleIssues = runRuleChecks(lines, language);
  if (module === "netflix") {
    const nfx = runNetflixChecks(lines, language);
    emit(`Stage 2: Netflix checks found ${nfx.length} issues`);
    ruleIssues.push(...nfx);
  }
  emit(`Stage 2 done: ${ruleIssues.length} rule issues in ${Date.now() - s2}ms`);

  // Split rule issues: OFFENSIVE terms (critical sensitive) go direct to final, rest go through AI
  const directIssues = ruleIssues.filter(
    (i) => i.category === "sensitive" && i.severity === "critical" && i.description.startsWith("Netflix offensive"),
  );
  ruleIssues = ruleIssues.filter(
    (i) => !(i.category === "sensitive" && i.severity === "critical" && i.description.startsWith("Netflix offensive")),
  );
  if (directIssues.length > 0) {
    emit(`${directIssues.length} offensive terms added directly (skip AI)`);
  }

  // ─── Stage 2b: AI validation of rule issues (POTENTIALLY_OFFENSIVE + other rule issues) ───
  if (ruleIssues.length > 0) {
    emit(`Stage 2b: AI validation of ${ruleIssues.length} rule issues`);
    const s2b = Date.now();
    const beforeCount = ruleIssues.length;
    try {
      const prompt = buildValidationPrompt(ruleIssues, language, deliverableType, lines);
      const text = await callWithRetry(client, VALIDATION_SYSTEM_PROMPT, prompt, "rule validation");
      const verdicts = parseValidationResponse(text);

      const verdictMap = new Map<number, typeof verdicts[0]>();
      for (const v of verdicts) {
        if (typeof v.issue_index === "number" && v.issue_index >= 0 && v.issue_index < ruleIssues.length) {
          verdictMap.set(v.issue_index, v);
        }
      }

      const validated: QCIssue[] = [];
      for (let i = 0; i < ruleIssues.length; i++) {
        const v = verdictMap.get(i);
        if (!v) { validated.push(ruleIssues[i]); continue; }
        if (v.verdict === "reject") continue;
        if (v.verdict === "downgrade") {
          const downgradeMap: Record<string, QCSeverity> = { critical: "warning", warning: "info", info: "info" };
          ruleIssues[i].severity = downgradeMap[ruleIssues[i].severity] || "info";
          if (v.reason) ruleIssues[i].description += ` [Review note: ${v.reason}]`;
        }
        validated.push(ruleIssues[i]);
      }
      ruleIssues = validated;
      emit(`Stage 2b done: ${ruleIssues.length} confirmed, ${beforeCount - ruleIssues.length} rejected in ${((Date.now() - s2b) / 1000).toFixed(1)}s`);
    } catch (err) {
      emit(`Stage 2b validation error (keeping unvalidated): ${(err as Error).message.substring(0, 80)}`);
    }
  }

  // ─── Stage 3: Merge & deduplicate (identical to Python) ───
  emit(`Stage 3: Merge & deduplicate`);
  const issueMap = new Map<string, QCIssue>();
  // AI issues first
  for (const issue of aiIssues) {
    const key = `${issue.lineId}:${issue.category}`;
    const existing = issueMap.get(key);
    if (!existing || sevRank(issue.severity) < sevRank(existing.severity)) {
      issueMap.set(key, issue);
    }
  }
  // Rule issues — only add if not already covered by AI or higher severity
  for (const issue of ruleIssues) {
    const key = `${issue.lineId}:${issue.category}`;
    const existing = issueMap.get(key);
    if (!existing || sevRank(issue.severity) <= sevRank(existing.severity)) {
      issueMap.set(key, issue);
    }
  }
  // Add direct offensive terms (they bypassed AI)
  for (const issue of directIssues) {
    const key = `${issue.lineId}:${issue.category}:offensive`;
    issueMap.set(key, issue);
  }
  let merged = [...issueMap.values()];
  merged.sort((a, b) => a.lineId !== b.lineId ? a.lineId - b.lineId : sevRank(a.severity) - sevRank(b.severity));
  emit(`Stage 3 done: ${aiIssues.length} AI + ${ruleIssues.length} rule + ${directIssues.length} offensive → ${merged.length} merged`);

  // ─── Stage 4: Second validation pass (identical to Python) ───
  emit(`Stage 4: Second validation pass on ${merged.length} issues`);
  const s4 = Date.now();
  try {
    const prompt = buildSecondPassPrompt(merged, lines, language, deliverableType);
    const text = await callWithRetry(client, SECOND_PASS_SYSTEM_PROMPT, prompt, "second pass");
    const verdicts = parseValidationResponse(text);

    const verdictMap = new Map<number, typeof verdicts[0]>();
    for (const v of verdicts) {
      if (typeof v.issue_index === "number" && v.issue_index >= 0 && v.issue_index < merged.length) {
        verdictMap.set(v.issue_index, v);
      }
    }

    const validated: QCIssue[] = [];
    let confirmed = 0, rejected = 0, downgraded = 0;
    for (let i = 0; i < merged.length; i++) {
      const v = verdictMap.get(i);
      if (!v) { validated.push(merged[i]); confirmed++; continue; }
      if (v.verdict === "reject") { rejected++; continue; }
      if (v.verdict === "downgrade") {
        downgraded++;
        // Downgrade by one level: critical→warning, warning→info
        const currentSev = merged[i].severity;
        const downgradeMap: Record<string, QCSeverity> = { critical: "warning", warning: "info", info: "info" };
        merged[i].severity = downgradeMap[currentSev] || "info";
        if (v.reason) merged[i].description += ` [Note: ${v.reason}]`;
      } else {
        confirmed++;
      }
      validated.push(merged[i]);
    }
    merged = validated;
    emit(`Stage 4 done: ${confirmed} confirmed, ${rejected} rejected, ${downgraded} downgraded in ${((Date.now() - s4) / 1000).toFixed(1)}s`);
  } catch (err) {
    emit(`Stage 4 second pass error (keeping unvalidated): ${(err as Error).message.substring(0, 80)}`);
  }

  // ─── Stage 5 (optional): KNP Terminology Consistency Check ───
  if (knpData && knpData.entries.length > 0) {
    emit(`Stage 5: KNP consistency check (${knpData.entries.length} terms, ${knpData.sourceLanguage}→${knpData.targetLanguage})`);
    const s5 = Date.now();
    try {
      const knpPrompt = formatKNPForPrompt(knpData);
      const knpBatchSize = 15;
      const knpBatches: QCLine[][] = [];
      for (let i = 0; i < lines.length; i += knpBatchSize) {
        knpBatches.push(lines.slice(i, i + knpBatchSize));
      }

      const knpResults = await Promise.all(
        knpBatches.map(async (batch, idx) => {
          const lineText = batch.map((l) => `Line ${l.id}: ${l.text}`).join("\n");
          const prompt = `${knpPrompt}\n\nCheck these subtitles for KNP terminology consistency:\n\n${lineText}\n\nFor each line where a KNP source term should appear but the approved target translation is NOT used (or is inconsistent), return a JSON array of issues:\n\`\`\`json\n[\n  { "line_id": <int>, "severity": "warning", "category": "consistency", "flagged_text": "<the incorrect term used>", "description": "<explain which KNP term should be used>", "suggested_fix": "<the complete corrected line>" }\n]\n\`\`\`\nIf no KNP issues found, return [].`;

          try {
            const text = await callWithRetry(client, "You are a terminology consistency checker for subtitle QC.", prompt, `knp batch ${idx + 1}/${knpBatches.length}`);
            return parseIssues(text, lines);
          } catch {
            return [];
          }
        }),
      );

      const knpIssues = knpResults.flat();
      emit(`Stage 5 done: ${knpIssues.length} KNP issues in ${((Date.now() - s5) / 1000).toFixed(1)}s`);
      merged.push(...knpIssues);
    } catch (err) {
      emit(`Stage 5 KNP error (non-fatal): ${(err as Error).message.substring(0, 80)}`);
    }
  }

  // ─── Stage 6: AI Translator Credit Check (last 3 subtitles) ───
  if (module === "netflix" && lines.length > 0) {
    emit(`Stage 6: AI translator credit check`);
    const s6 = Date.now();
    try {
      const lastLines = lines.slice(-3);
      const lastText = lastLines.map((l) => `Line ${l.id}: ${l.text}`).join("\n");
      const creditPrompt = `Check if any of these last subtitles contain a translator/subtitler credit. Credits typically say "Translated by", "Subtitled by", "Subtitle translation by" or equivalent in any language.\n\nSubtitles:\n${lastText}\n\nRespond with JSON:\n\`\`\`json\n[\n  { "line_id": <int>, "severity": "warning", "category": "formatting", "flagged_text": "<the credit text>", "description": "Translator credit detected — verify it follows Netflix credit guidelines", "suggested_fix": "Verify against Netflix Originals Credit Translation document" }\n]\n\`\`\`\nIf no credits found, return [].`;

      const creditText = await callWithRetry(client, "You detect translator credits in subtitle files. Be precise — only flag actual credits, not normal dialogue.", creditPrompt, "credit check");
      const creditIssues = parseIssues(creditText, lines);
      if (creditIssues.length > 0) {
        emit(`Stage 6 done: ${creditIssues.length} credit issue(s) in ${((Date.now() - s6) / 1000).toFixed(1)}s`);
        merged.push(...creditIssues);
      } else {
        emit(`Stage 6 done: no credits detected in ${((Date.now() - s6) / 1000).toFixed(1)}s`);
      }
    } catch (err) {
      emit(`Stage 6 credit check error (non-fatal): ${(err as Error).message.substring(0, 80)}`);
    }
  }

  // Final sort
  merged.sort((a, b) => a.lineId !== b.lineId ? a.lineId - b.lineId : sevRank(a.severity) - sevRank(b.severity));

  const durationMs = Date.now() - start;
  emit(`Pipeline complete: ${merged.length} final issues in ${(durationMs / 1000).toFixed(1)}s`);

  // Build timecode lookup for report
  const lineTimecodes: Record<number, { tcIn: string; tcOut: string }> = {};
  for (const line of lines) {
    if (line.tcIn && line.tcOut) {
      lineTimecodes[line.id] = { tcIn: line.tcIn, tcOut: line.tcOut };
    }
  }

  return {
    filename,
    issues: merged,
    summary: buildSummary(merged, lines.length),
    module,
    language,
    engine: "claude",
    durationMs,
    lineTimecodes,
  };
}