fix(automod): flow real LLM analysis + descriptive fallback

Root cause: ai-analysis-worker read llmResult.explanation and
llmResult.toxicityScore — fields the LLM pipeline never produces
(canonical AnalysisResult uses analysis/score). Every message fell back
to the bare template "Tidak ada indikasi pelanggaran." and the stored
score was always 0.

- Map analysis/score correctly; fallback now quotes the message content
- Prompt: ban generic analysis phrasing, require reply context
- LLM context: include replied-to message content (metadata.reference)
  so the model can explain what the user is replying to
- Frontend: show thread/channel names from metadata instead of raw IDs
  (message card, detail views, search overlay); detail panel now
  displays the ai_analysis text
- Auto-delete log/DM include the descriptive analysis as the reason

services/discord-gateway/src/modules/ai-moderation/ai-analysis-worker.ts

@@ -58,11 +58,8 @@ export interface AnalysisResult {
     | "review"
     | "delete"
     | "escalate";
-  toxicityScore: number;
-  harmScore: number;
-  jailbreakScore: number;
-  safetyScore: number;
-  explanation: string;
+  score: number;
+  analysis: string;
   correctedFlags?: string[];
 }
 
@@ -182,7 +179,7 @@ export default async function workerRouter(
  * Runs the LLM moderation analysis on a single message.
  *
  * The LLM verdict IS the result — confidence, severity, flags and
- * explanation all come from the model. On failure the message is marked
+ * analysis all come from the model. On failure the message is marked
  * "error" (explicit, retryable) instead of receiving a heuristic verdict.
  */
 async function runLLMAnalysis(
@@ -214,15 +211,10 @@ async function runLLMAnalysis(
       severity: llmResult.severity ?? "none",
       confidence: normalizeConfidence(llmResult.confidence),
       recommendedAction: llmResult.recommendedAction ?? "none",
-      toxicityScore: llmResult.toxicityScore ?? 0,
-      harmScore: llmResult.harmScore ?? 0,
-      jailbreakScore: llmResult.jailbreakScore ?? 0,
-      safetyScore: llmResult.safetyScore ?? 0,
-      explanation:
-        llmResult.explanation?.trim() ||
-        (llmResult.status === "clean"
-          ? "Tidak ada indikasi pelanggaran."
-          : "Pesan terindikasi melanggar kebijakan (analisis AI)."),
+      score: llmResult.score ?? 0,
+      analysis:
+        llmResult.analysis?.trim() ||
+        buildFallbackAnalysis(message, llmResult.status ?? "clean"),
     };
   } catch (error) {
     const errorMsg = error instanceof Error ? error.message : String(error);
@@ -242,6 +234,28 @@ function normalizeConfidence(raw: number | undefined | null): number {
   return 0.7;
 }
 
+/**
+ * Content-aware fallback when the LLM returns an empty `analysis` field.
+ * Quotes the actual message content so the log still explains WHAT was said
+ * instead of a bare template like "Tidak ada indikasi pelanggaran."
+ */
+function buildFallbackAnalysis(
+  message: MessageRecord,
+  status: string,
+): string {
+  const raw = (message.edited_content ?? message.content ?? "").trim();
+  const snippet = raw.length > 120 ? `${raw.slice(0, 120).trimEnd()}…` : raw;
+
+  if (status === "clean") {
+    return snippet
+      ? `Tidak ada indikasi pelanggaran. Pesan: "${snippet}" dinilai wajar dalam konteks percakapan.`
+      : "Tidak ada indikasi pelanggaran. Isi pesan dinilai wajar dalam konteks percakapan.";
+  }
+  return snippet
+    ? `Pesan terindikasi melanggar kebijakan: "${snippet}".`
+    : "Pesan terindikasi melanggar kebijakan (analisis AI).";
+}
+
 function buildFallbackResult(
   messageId: string,
   reason: string,
@@ -254,11 +268,8 @@ function buildFallbackResult(
     severity: "none",
     confidence: 0,
     recommendedAction: "review",
-    toxicityScore: 0,
-    harmScore: 0,
-    jailbreakScore: 0,
-    safetyScore: 0,
-    explanation: reason,
+    score: 0,
+    analysis: reason,
   };
 }
 
@@ -319,14 +330,14 @@ async function processBatch(job: {
     result: {
       status: result.status,
       flags: JSON.stringify(result.flags),
-      score: result.toxicityScore,
-      analysis: result.explanation,
+      score: result.score,
+      analysis: result.analysis,
       categories: result.categories,
       severity: result.severity,
       confidence: result.confidence,
       recommendedAction: result.recommendedAction,
       analyzedAt: Date.now(),
-      error: result.status === "error" ? result.explanation : null,
+      error: result.status === "error" ? result.analysis : null,
     },
   }));
 

services/discord-gateway/src/modules/ai-moderation/autoDeleteLogger.ts

@@ -33,6 +33,7 @@ export async function logDeletionToChannel(
       const severity = message.ai_severity ?? "none";
       const categories =
         message.ai_categories ?? message.ai_moderation_flags ?? "—";
+      const reason = message.ai_analysis ?? "—";
       const snippet = (message.edited_content ?? message.content).substring(
         0,
         200,
@@ -42,6 +43,7 @@ export async function logDeletionToChannel(
           `**Status:** ${message.ai_status}\n` +
           `**Severitas:** ${severity}\n` +
           `**Kategori:** ${categories}\n` +
+          `**Alasan:** ${reason}\n` +
           `**Isi:** ${snippet}\n` +
           `**Waktu:** <t:${Math.floor(Date.now() / 1000)}:R>`,
       );

services/discord-gateway/src/modules/ai-moderation/autoDeleteNotify.ts

@@ -20,8 +20,14 @@ export async function sendDeletionNotification(
   try {
     const targetUser = await client.users.fetch(message.user_id);
     if (targetUser) {
+      // Prefer the descriptive LLM analysis so the user understands WHY;
+      // fall back to category/flag labels when it is unavailable.
+      const analysis = (message.ai_analysis ?? "").trim();
       const reason: string =
-        message.ai_categories ?? message.ai_moderation_flags ?? "(unknown)";
+        (analysis.length > 240 ? `${analysis.slice(0, 240)}…` : analysis) ||
+        message.ai_categories ??
+        message.ai_moderation_flags ??
+        "(unknown)";
       await targetUser.send(
         `Pesan Anda di **${guildName}** telah dihapus oleh sistem moderasi otomatis.\n` +
           `Alasan: ${reason}\n` +

services/discord-gateway/src/modules/ai-moderation/conversationContext.ts

@@ -42,12 +42,44 @@ export function estimateTokens(text: string): number {
 }
 
 /**
- * Formats reference info for a message (reply/forward/crosspost)
+ * Formats reference info for a message (reply/forward/crosspost).
+ *
+ * The replied-to content is stored in the message metadata
+ * (`metadata.reference.content` / `repliedUsername`) at capture time, so
+ * include it here — the LLM can then explain WHAT the user is replying to
+ * instead of only seeing a raw message ID it cannot resolve.
  */
 function formatReferenceInfo(msg: MessageRecord): string {
   const parts: string[] = [];
+
+  let repliedContent: string | null = null;
+  let repliedUsername: string | null = null;
+  try {
+    const meta = JSON.parse(msg.metadata ?? "") as {
+      reference?: {
+        content?: string | null;
+        repliedUsername?: string | null;
+      } | null;
+    };
+    repliedContent = meta?.reference?.content ?? null;
+    repliedUsername = meta?.reference?.repliedUsername ?? null;
+  } catch {
+    // metadata malformed — fall back to ID-only reference
+  }
+
+  const repliedText = (repliedContent ?? "").trim();
+  const repliedSnippet = repliedText
+    ? sanitizeDiscordTokens(
+        repliedText.length > 200
+          ? `${repliedText.slice(0, 200)}…`
+          : repliedText,
+      )
+    : null;
+
   if (msg.is_reply && msg.reference_message_id) {
-    parts.push(`[reply_to: ${msg.reference_message_id}]`);
+    const who = repliedUsername ? ` oleh ${repliedUsername}` : "";
+    const what = repliedSnippet ? `: "${repliedSnippet}"` : "";
+    parts.push(`[reply_to: ${msg.reference_message_id}${who}${what}]`);
     if (msg.reference_channel_id) {
       parts.push(`(reply_channel: ${msg.reference_channel_id})`);
     }

services/discord-gateway/src/modules/ai-moderation/prompts/output.ts

@@ -153,9 +153,11 @@ Contoh buruk: "Pengirim bercanda tentang agama." (JANGAN menggunakan kata "berca
 
 CRITICAL:
 - JANGAN PERNAH menulis "Pesan hanya berisi..." atau "Pesan tidak mengandung..." sebagai analysis.
+- JANGAN PERNAH menulis "Tidak ada indikasi pelanggaran" atau frasa generik serupa sebagai analysis — wajib sebutkan TOPIK/ISI pesan secara spesifik apa yang sedang dibicarakan pengirim.
 - JANGAN PERNAH menulis template generik seperti "Pengirim mengirimkan sebuah file GIF tanpa pelanggaran". Kamu WAJIB mendeskripsikan isi visualnya secara spesifik berdasarkan Media analysis.
 - JANGAN PERNAH menyebutkan nama / username pengguna secara langsung. Selalu gunakan kata "Pengirim" atau "Pengguna".
 - Selalu sebutkan ISI KONTEN secara spesifik — apa yang dibicarakan, apa yang terlihat di gambar.
+- Jika pesan adalah BALASAN (reply) ke pesan lain, jelaskan konteks balasannya: apa yang sedang dibicarakan, siapa yang dibalas (tanpa nama, cukup peran/isi pesan yang dibalas), dan bagaimana tanggapan pengirim terhadapnya.
 - Gunakan informasi dari Media analysis untuk mendeskripsikan gambar.
 - Analisis harus MEMBERI KONTEKS, bukan hanya menyatakan status.
 - GUNAKAN <user_profile> untuk personalisasi analysis — jadikan analysis terasa seperti sistem "mengenal" pengguna.