/**
 * Heuristic diff classifier for feedback analysis.
 * Analyzes editDiffs from a completed job and produces feedback classifications.
 * Layer 1: rule-based. Layer 2 (Claude skill) will enrich with source code analysis.
 */

type EditDiff = {
  id: string;
  subtitleIndex: number;
  field: string;
  oldValue: string | null;
  newValue: string | null;
  rationale: string;
  changeCategory: string;
};

type FeedbackClassification = {
  errorType: "input_error" | "cicd_needed" | "style_preference" | "false_positive";
  confidenceScore: number;
  description: string;
  affectedInput: "neon" | "alto_curated" | "lsx_prompt" | "lsx_logic" | "arc_config" | null;
};

// Keyword patterns for rationale analysis
const INPUT_ERROR_KEYWORDS = ["wrong", "incorrect", "mistranslat", "error in source", "bad input", "original is wrong", "source error"];
const STYLE_KEYWORDS = ["style", "preference", "natural", "sounds better", "flow", "readability", "tone", "nuance"];
const NEON_KEYWORDS = ["neon", "source", "template", "original", "input text", "transcription"];
const ALTO_KEYWORDS = ["alto", "curated", "reference", "base translation"];
const LSX_LOGIC_KEYWORDS = ["timing", "cpl", "cps", "overlap", "gap", "duration", "frame", "sync"];
const LSX_PROMPT_KEYWORDS = ["format", "style guide", "capitalization", "punctuation", "tag", "italic"];

function matchesKeywords(text: string, keywords: string[]): boolean {
  const lower = text.toLowerCase();
  return keywords.some((k) => lower.includes(k));
}

function classifySingleDiff(diff: EditDiff): FeedbackClassification {
  const { changeCategory, rationale, field } = diff;
  const rat = rationale.toLowerCase();

  // timing_fix → usually cicd_needed (LSX should handle timing)
  if (changeCategory === "timing_fix") {
    return {
      errorType: "cicd_needed",
      confidenceScore: matchesKeywords(rat, LSX_LOGIC_KEYWORDS) ? 0.85 : 0.65,
      description: `Timing correction on subtitle ${diff.subtitleIndex + 1}: ${rationale}`,
      affectedInput: "lsx_logic",
    };
  }

  // cpl_fix → cicd_needed (LSX should auto-fix CPL violations)
  if (changeCategory === "cpl_fix") {
    return {
      errorType: "cicd_needed",
      confidenceScore: 0.8,
      description: `CPL violation fix on subtitle ${diff.subtitleIndex + 1}: ${rationale}`,
      affectedInput: "lsx_logic",
    };
  }

  // split_merge → cicd_needed (LSX cueing should handle this)
  if (changeCategory === "split_merge") {
    return {
      errorType: "cicd_needed",
      confidenceScore: 0.75,
      description: `Split/merge on subtitle ${diff.subtitleIndex + 1}: ${rationale}`,
      affectedInput: "lsx_logic",
    };
  }

  // style_correction → style_preference
  if (changeCategory === "style_correction") {
    return {
      errorType: "style_preference",
      confidenceScore: 0.7,
      description: `Style correction on subtitle ${diff.subtitleIndex + 1}: ${rationale}`,
      affectedInput: matchesKeywords(rat, LSX_PROMPT_KEYWORDS) ? "lsx_prompt" : null,
    };
  }

  // translation_fix → depends on rationale
  if (changeCategory === "translation_fix") {
    if (matchesKeywords(rat, INPUT_ERROR_KEYWORDS)) {
      return {
        errorType: "input_error",
        confidenceScore: 0.8,
        description: `Translation error on subtitle ${diff.subtitleIndex + 1}: ${rationale}`,
        affectedInput: matchesKeywords(rat, ALTO_KEYWORDS)
          ? "alto_curated"
          : matchesKeywords(rat, NEON_KEYWORDS)
            ? "neon"
            : "neon",
      };
    }
    if (matchesKeywords(rat, STYLE_KEYWORDS)) {
      return {
        errorType: "style_preference",
        confidenceScore: 0.65,
        description: `Style preference on subtitle ${diff.subtitleIndex + 1}: ${rationale}`,
        affectedInput: "lsx_prompt",
      };
    }
    // Default translation fix → could be input error but low confidence
    return {
      errorType: "input_error",
      confidenceScore: 0.5,
      description: `Translation change on subtitle ${diff.subtitleIndex + 1}: ${rationale}`,
      affectedInput: "neon",
    };
  }

  // other / fallback → false_positive with low confidence
  return {
    errorType: "false_positive",
    confidenceScore: 0.4,
    description: `Edit on subtitle ${diff.subtitleIndex + 1} (${field}): ${rationale}`,
    affectedInput: null,
  };
}

/**
 * Classify all diffs from a completed job.
 * Groups similar patterns to avoid 1:1 diff→analysis noise.
 * Returns deduplicated feedback classifications.
 */
export function classifyDiffs(diffs: EditDiff[]): FeedbackClassification[] {
  if (diffs.length === 0) return [];

  // Classify each diff
  const classified = diffs.map(classifySingleDiff);

  // Group by errorType + affectedInput to reduce noise
  const groups = new Map<string, FeedbackClassification[]>();
  for (const c of classified) {
    const key = `${c.errorType}:${c.affectedInput || "none"}`;
    if (!groups.has(key)) groups.set(key, []);
    groups.get(key)!.push(c);
  }

  // Produce one analysis per group with aggregated description
  const results: FeedbackClassification[] = [];
  for (const [, items] of groups) {
    if (items.length === 0) continue;
    const first = items[0];
    const avgConfidence =
      items.reduce((sum, i) => sum + i.confidenceScore, 0) / items.length;

    results.push({
      errorType: first.errorType,
      confidenceScore: Math.round(avgConfidence * 100) / 100,
      description:
        items.length === 1
          ? first.description
          : `${items.length} ${first.errorType.replace("_", " ")} issues detected. Examples: ${items
              .slice(0, 3)
              .map((i) => i.description)
              .join("; ")}`,
      affectedInput: first.affectedInput,
    });
  }

  // Sort by confidence descending
  results.sort((a, b) => b.confidenceScore - a.confidenceScore);
  return results;
}