feat(ai-moderation): enhance context handling with structured XML blocks and user profiles

.gitignore

@@ -12,7 +12,7 @@ worktrees/
 .worktrees/
 services/frontend/frontend/dist/
 target/
-
+nix/
 # Gitea CI runner logs
 .gitea/workflows/*.log
 

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

@@ -25,6 +25,7 @@ import {
   buildConversationContext,
   buildLocationContext,
 } from "./conversationContext.js";
+import { buildConversationContextBlock } from "./moderationBuilders.js";
 import { runModerationAnalysis } from "./moderationOrchestrator.js";
 
 const logger = createChildLogger("ai-analysis-worker");
@@ -280,13 +281,11 @@ async function processBatch(job: {
     maxAgeMs: config.AI_ANALYSIS_CONTEXT_MAX_AGE_MS,
     gapMs: config.AI_ANALYSIS_CONTEXT_GAP_MS,
   });
-  const contextText = [
+  const contextBlock = buildConversationContextBlock({
-    buildLocationContext(messages),
+    location: buildLocationContext(messages),
-    contextLines.descriptor,
+    descriptor: contextLines.descriptor,
-    ...contextLines.lines,
+    lines: contextLines.lines,
-  ]
+  });
-    .filter((l) => l.trim().length > 0)
-    .join("\n");
 
   const targetIds = messages.map((m) => m.id);
   const contextIds = contextBefore.map((m) => m.id);
@@ -300,7 +299,7 @@ async function processBatch(job: {
   // when media is present), not N per-message calls.
   const moderationResult = await runModerationAnalysis({
     targets: messages,
-    contextText,
+    contextBlock,
     attachments,
   });
 
@@ -373,13 +372,11 @@ async function processIndividual(job: {
     maxAgeMs: config.AI_ANALYSIS_CONTEXT_MAX_AGE_MS,
     gapMs: config.AI_ANALYSIS_CONTEXT_GAP_MS,
   });
-  const contextText = [
+  const contextBlock = buildConversationContextBlock({
-    buildLocationContext([message]),
+    location: buildLocationContext([message]),
-    contextLines.descriptor,
+    descriptor: contextLines.descriptor,
-    ...contextLines.lines,
+    lines: contextLines.lines,
-  ]
+  });
-    .filter((l) => l.trim().length > 0)
-    .join("\n");
 
   const contextIds = contextBefore.map((m) => m.id);
   const attachments = await messageStore.getAttachmentsForMessages([
@@ -390,7 +387,7 @@ async function processIndividual(job: {
   try {
     const moderationResult = await runModerationAnalysis({
       targets: [message],
-      contextText,
+      contextBlock,
       attachments,
     });
 

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

@@ -6,7 +6,7 @@ import {
 } from "../message-capture/messageMetadata.js";
 import type { MessageRecord } from "../message-capture/types.js";
 import { sanitizeDiscordTokens } from "./discordTokens.js";
-import { resolveDisplayName } from "./moderationBuilders.js";
+import { escapeXml, resolveDisplayName } from "./moderationBuilders.js";
 
 const logger = createChildLogger("conversationContext");
 
@@ -124,8 +124,10 @@ export function formatMessageForPrompt(
   msg: MessageRecord,
   label: "context" | "target",
 ): string {
-  const content = sanitizeDiscordTokens(
+  const content = truncateContextLine(
-    renderDiscordMentions(msg.edited_content ?? msg.content, msg.metadata),
+    sanitizeDiscordTokens(
+      renderDiscordMentions(msg.edited_content ?? msg.content, msg.metadata),
+    ),
   );
   const timestamp = formatTimestamp(msg.created_at);
   const mediaEvidence = formatMediaEvidenceForPrompt(msg.metadata);
@@ -134,11 +136,26 @@ export function formatMessageForPrompt(
   return `[${label}] id=${msg.id} time=${timestamp} user=${resolveDisplayName(msg)}: ${content}${mediaSuffix}${refInfo}`;
 }
 
+/** Max content chars per context line — a single huge paste (log dump,
+ *  copypasta) must not eat the whole conversation budget. */
+const CONTEXT_LINE_CONTENT_MAX_CHARS = 1500;
+
+/** Marker appended when a context line's content was cut. Distinct from the
+ *  target-content marker so the model knows which side was truncated. */
+export const CONTEXT_TRUNC_MARKER = "…[konteks dipotong: terlalu panjang]";
+
+/** Cap one context message's content to CONTEXT_LINE_CONTENT_MAX_CHARS. */
+export function truncateContextLine(content: string): string {
+  if (content.length <= CONTEXT_LINE_CONTENT_MAX_CHARS) return content;
+  return `${content.slice(0, CONTEXT_LINE_CONTENT_MAX_CHARS).trimEnd()}${CONTEXT_TRUNC_MARKER}`;
+}
+
 /**
- * Builds a one-line `<location_context>` source line for the batch — channel
+ * Builds a structured `<location_context .../>` element for the batch —
- * name, thread name and age-restriction flags from captured message metadata.
+ * channel/thread name and age-restriction flags from captured message
- * The LLM uses it to judge messages in the right channel context (e.g. a
+ * metadata. The LLM uses it to judge messages in the right channel context
- * thread about a specific topic, or an age-restricted channel).
+ * (e.g. a thread about a specific topic, or an age-restricted channel).
+ * Returns "" when no channel metadata was captured.
  */
 export function buildLocationContext(targets: MessageRecord[]): string {
   const target = targets[0];
@@ -155,26 +172,20 @@ export function buildLocationContext(targets: MessageRecord[]): string {
     };
     const ch = meta?.channel;
     if (!ch) return "";
-    const parts: string[] = [];
+    const attrs: string[] = [`channel_id="${escapeXml(target.channel_id)}"`];
-    parts.push(
+    if (ch.channelName)
-      `id=${target.channel_id}${
+      attrs.push(`channel_name="${escapeXml(ch.channelName)}"`);
-        ch.channelName ? ` name=${JSON.stringify(ch.channelName)}` : ""
-      }`,
-    );
     if (target.thread_id || ch.threadName) {
-      parts.push(
+      if (target.thread_id)
-        `thread=${target.thread_id}${
+        attrs.push(`thread_id="${escapeXml(target.thread_id)}"`);
-          ch.threadName ? ` thread_name=${JSON.stringify(ch.threadName)}` : ""
+      if (ch.threadName)
-        }`,
+        attrs.push(`thread_name="${escapeXml(ch.threadName)}"`);
-      );
-    }
-    if (typeof ch.nsfw === "boolean") {
-      parts.push(`nsfw=${ch.nsfw}`);
     }
+    if (typeof ch.nsfw === "boolean") attrs.push(`nsfw="${ch.nsfw}"`);
     if (typeof ch.ageRestricted === "boolean") {
-      parts.push(`age_restricted=${ch.ageRestricted}`);
+      attrs.push(`age_restricted="${ch.ageRestricted}"`);
     }
-    return `[location] ${parts.join(" ")}`;
+    return `<location_context ${attrs.join(" ")}/>`;
   } catch {
     return "";
   }

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

@@ -16,8 +16,10 @@ import { getChannelCulture } from "./channelCultureStore.js";
 import type { RetryState } from "./llmCaller.js";
 import { callModerationLLM } from "./llmCaller.js";
 import { prepareMediaMessage } from "./mediaAnalysisClient.js";
+import { buildUserProfilesBlock } from "./moderationBuilders.js";
 import { buildSystemPrompt as buildSystemPromptModular } from "./moderationPrompt.js";
 import { buildCorrectedFewShotExamples } from "./textBatchProcessor.js";
+import { getUserProfile } from "./userProfileStore.js";
 
 const log = createChildLogger("mediaBatchProcessor");
 
@@ -26,7 +28,7 @@ const log = createChildLogger("mediaBatchProcessor");
 // ---------------------------------------------------------------------------
 export async function runMediaBatch(
   targets: MessageRecord[],
-  contextText: string,
+  contextBlock: string,
   attachments: AttachmentRecord[] | undefined,
 ): Promise<{ results: AnalysisResult[]; raw: unknown }> {
   if (!targets.length) return { results: [], raw: null };
@@ -58,14 +60,32 @@ export async function runMediaBatch(
   const channelCulture = channelCultureObj?.culture_summary;
   const correctedExamples = await buildCorrectedFewShotExamples();
   const systemText = buildSystemPromptModular({
-    contextText,
     mode: "mixed",
     correctedExamples,
     channelCulture,
   });
 
+  // Gather user profiles ONCE for the whole batch and emit a deduplicated
+  // <user_profiles> map; per-message blocks (from prepareMediaMessage)
+  // reference it via <user_profile_ref>.
+  const profileByUser = new Map<string, string>();
+  for (const t of targets) {
+    if (profileByUser.has(t.user_id)) continue;
+    const profile = await getUserProfile(t.user_id);
+    profileByUser.set(t.user_id, profile?.profile_summary ?? "");
+  }
+  const userProfilesBlock = buildUserProfilesBlock(profileByUser);
+
   const messagesBlock = prepared.map((p) => p.messageBlock).join("\n");
-  const userContent = `<messages_to_analyze>\n${messagesBlock}\n</messages_to_analyze>`;
+  // Data/instruction separation: the system prompt is stable per mode — all
+  // per-batch context (profiles, conversation) lives in the USER payload,
+  // ordered oldest-first so targets come last.
+  const userBlocks = [
+    userProfilesBlock?.trimEnd() ?? "",
+    contextBlock?.trimEnd() ?? "",
+    `<messages_to_analyze>\n${messagesBlock}\n</messages_to_analyze>`,
+  ].filter((b) => b.trim().length > 0);
+  const userContent = userBlocks.join("\n\n");
 
   const perMsgTimeout = config.AI_LLM_MEDIA_ANALYSIS_TIMEOUT_MS ?? 60000;
   const batchTimeout = Math.min(

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

@@ -9,6 +9,7 @@ import { renderDiscordMentions } from "../message-capture/messageMetadata.js";
 import { messageStore } from "../message-capture/messageStore.js";
 import type { MessageRecord } from "../message-capture/types.js";
 import { sanitizeDiscordTokens } from "./discordTokens.js";
+import { sanitizeAiContent } from "./prompts/output.js";
 
 /** Simple XML-escaping for content text. */
 export function escapeXml(s: string): string {
@@ -19,6 +20,101 @@ export function escapeXml(s: string): string {
     .replace(/"/g, """);
 }
 
+// ---------------------------------------------------------------------------
+// Conversation context block — structured data for the USER message.
+//
+// All per-batch context lives in the USER message (not the SYSTEM prompt) so
+// the system prompt is stable per mode (cacheable on routers/providers) and
+// the role boundary is clean: instructions in SYSTEM, data in USER.
+// ---------------------------------------------------------------------------
+
+/** Outer char cap for the assembled `<conversation_context>` inner text. */
+export const CONVERSATION_CONTEXT_MAX_CHARS = 40_000;
+
+/**
+ * Wraps per-batch context data into structured XML blocks for the USER
+ * message:
+ *
+ *   <location_context channel_id="..." channel_name="..." nsfw="..."/>
+ *   <conversation_context>
+ *     [conversation_flow] status=ongoing context_msgs=12 dropped=0
+ *     [context] id=... time=... user=...: isi pesan
+ *     ...
+ *   </conversation_context>
+ *
+ * Empty blocks are omitted entirely (never emit a hollow `<conversation_context>`
+ * with no content). The inner text is AI/user-derived and passed through
+ * `sanitizeAiContent` (CDATA + XML-escape) to block prompt injection.
+ */
+export function buildConversationContextBlock(input: {
+  /** Pre-built `<location_context .../>` string (or ""). */
+  location?: string;
+  /** `[conversation_flow]` descriptor line from buildConversationContext. */
+  descriptor?: string;
+  /** `[context]` lines, oldest → newest. */
+  lines: string[];
+}): string {
+  const blocks: string[] = [];
+  const location = input.location?.trim();
+  if (location) blocks.push(location);
+
+  const inner = [input.descriptor ?? "", ...input.lines]
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0)
+    .join("\n");
+  if (inner) {
+    blocks.push(
+      `<conversation_context>\n${sanitizeAiContent(inner, CONVERSATION_CONTEXT_MAX_CHARS)}\n</conversation_context>`,
+    );
+  }
+  return blocks.join("\n");
+}
+
+// ---------------------------------------------------------------------------
+// Per-message content bounds — protects the LLM token budget from a single
+// huge paste (stack traces, log dumps, copypasta). Truncation is explicit so
+// the model never mistakes the cut for a real message boundary.
+// ---------------------------------------------------------------------------
+
+/** Max characters of a message's content sent to the LLM `<content>` payload. */
+export const AI_CONTENT_MAX_CHARS = 4000;
+
+/** Marker appended when a message is longer than AI_CONTENT_MAX_CHARS. */
+export const AI_CONTENT_TRUNC_MARKER = "\n…[pesan dipotong: terlalu panjang]";
+
+/** Truncate a message's content for the LLM `<content>` payload. */
+export function truncateForAi(content: string): string {
+  if (content.length <= AI_CONTENT_MAX_CHARS) return content;
+  return `${content.slice(0, AI_CONTENT_MAX_CHARS)}${AI_CONTENT_TRUNC_MARKER}`;
+}
+
+// ---------------------------------------------------------------------------
+// User profile deduplication — a batch can contain many messages from the
+// same user. Instead of repeating the (up to 3000-char) profile summary on
+// every message, emit a single <user_profiles> map per batch and reference
+// entries per message with <user_profile_ref user_id="..."/>.
+// ---------------------------------------------------------------------------
+
+/** Build a deduplicated `<user_profiles>` map block, keyed by Discord user id. */
+export function buildUserProfilesBlock(
+  profiles: ReadonlyMap<string, string>,
+): string {
+  const entries = Array.from(profiles.entries()).filter(
+    ([, text]) => text.trim().length > 0,
+  );
+  if (entries.length === 0) return "";
+  const lines = entries.map(
+    ([userId, text]) =>
+      `  <user_profile user_id="${escapeXml(userId)}">${sanitizeAiContent(text)}</user_profile>`,
+  );
+  return `<user_profiles>\n${lines.join("\n")}\n</user_profiles>`;
+}
+
+/** Per-message reference tag pointing at an entry in the `<user_profiles>` map. */
+export function buildUserProfileRef(userId: string): string {
+  return `<user_profile_ref user_id="${escapeXml(userId)}"/>`;
+}
+
 /**
  * Returns the real text content for AI analysis, stripping fallback text
  * that getDisplayContent() synthesized ("[Attachment: ...]", "[Sticker: ...]",

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

@@ -35,7 +35,13 @@ const log = createChildLogger("moderationOrchestrator");
 // ---------------------------------------------------------------------------
 export interface ModerationInput {
   targets: MessageRecord[];
-  contextText: string;
+  /**
+   * Pre-built XML context block for the USER message (from
+   * `buildConversationContextBlock`): `<location_context .../>` +
+   * `<conversation_context>...</conversation_context>`. Kept out of the
+   * system prompt so it stays stable/cacheable per mode.
+   */
+  contextBlock: string;
   attachments?: AttachmentRecord[];
 }
 
@@ -62,7 +68,7 @@ export interface ModerationOutput {
 export async function runModerationAnalysis(
   input: ModerationInput,
 ): Promise<ModerationOutput> {
-  const { targets, contextText, attachments } = input;
+  const { targets, contextBlock, attachments } = input;
 
   initSearxngCache(config.REDIS_URL);
   if (!targets.length) throw new Error("No targets provided for analysis");
@@ -320,10 +326,10 @@ export async function runModerationAnalysis(
   // Run both paths in parallel
   const [textBatchResult, mediaBatchResult] = await Promise.all([
     textOnlyTargets.length > 0
-      ? runTextOnlyBatch(textOnlyTargets, contextText)
+      ? runTextOnlyBatch(textOnlyTargets, contextBlock)
       : Promise.resolve({ results: [] as AnalysisResult[], raw: null }),
     mediaTargets.length > 0
-      ? runMediaBatch(mediaTargets, contextText, attachments)
+      ? runMediaBatch(mediaTargets, contextBlock, attachments)
       : Promise.resolve({ results: [] as AnalysisResult[], raw: null }),
   ]);
 

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

@@ -31,8 +31,12 @@ Struktur wajib:
   ]
 }
 
+Instruksi per field:
+- "message_id": WAJIB sama persis dengan id di input. Setiap <message> di <messages_to_analyze> menghasilkan SATU hasil. Jangan gabungkan beberapa pesan, jangan lewati, jangan karang id.
+- "evidence": kutipan PERSIS frasa yang melanggar (maks 1 baris). Pelanggaran di gambar/sticker → kutip deskripsi Media analysis. Pelanggaran lewat balasan/referensi → sebut konteks pesan yang dibalas. Boleh tambah label sumber, mis. [media analysis] / [web_search] / [reply]. Kosong jika clean.
+
 ## PERSONALITY & MEMORI — Profil Pengguna dan Kultur Channel
-Data konteks tersedia: <user_profile> (ringkasan kepribadian pengguna) dan <channel_culture> (topik/vibe channel).
+Data konteks tersedia: <user_profiles> (peta ringkasan kepribadian, di pesan USER), <user_reputation> (skor trust), dan <channel_culture> (topik/vibe channel). Setiap <message> dapat memuat <user_profile_ref user_id="..."/> yang menunjuk ke entri di peta <user_profiles>.
 Gunakan untuk personalisasi analysis, tapi:
 - Profil adalah KONTEKS, bukan bukti. Profil mencurigakan ≠ flag; profil bersih ≠ loloskan pelanggaran.
 - Perubahan perilaku mencolok (biasanya teknis tiba-tiba provokatif) layak dicatat di analysis.
@@ -66,7 +70,7 @@ CRITICAL:
 - 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.
+- GUNAKAN <user_profile_ref>/<user_profiles> untuk personalisasi analysis — jadikan analysis terasa seperti sistem "mengenal" pengguna.
 - Jika perilaku pesan menyimpang dari profil yang diketahui, CATAT dalam analysis sebagai informasi kontekstual yang relevan.
 - JANGAN paksa referensi profil jika tidak relevan — analysis natural lebih baik dari yang dipaksakan.`;
 

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

@@ -39,7 +39,6 @@ Gambar/sticker/embed/preview link sudah DIDESKRIPSIKAN vision model sebelum batc
 // ---------------------------------------------------------------------------
 
 export interface BuildSystemPromptOptions {
-  contextText: string;
   /** Prompt mode — determines which sections are included. */
   mode: PromptMode;
   /** @deprecated Use `mode` instead. */
@@ -59,7 +58,6 @@ export interface BuildSystemPromptOptions {
 
 export function buildSystemPrompt(options: BuildSystemPromptOptions): string {
   const {
-    contextText,
     mode,
     includeMediaInstructions,
     correction,
@@ -105,15 +103,34 @@ export function buildSystemPrompt(options: BuildSystemPromptOptions): string {
   }
 
   parts.push(
-    `## Konteks Pengguna\nSetiap pesan mungkin memiliki tag <user_reputation>. Tag ini hanya indikator **referensi**, bukan bukti pelanggaran. Nilai trust_score yang rendah bukan alasan untuk memflag pesan yang bersih. Nilai trust_score yang tinggi bukan alasan untuk mengabaikan pelanggaran nyata. **Setiap pesan harus dinilai berdasarkan isinya sendiri.**`,
+    `## Blok Data di Pesan USER\n` +
+      `Semua data dinamis per-batch dikirim di pesan USER — system prompt ini TIDAK memuat data batch:\n` +
+      `- <location_context .../> = metadata channel/thread (channel_id, channel_name, thread_name, nsfw, age_restricted).\n` +
+      `- <conversation_context> = obrolan SEBELUM pesan target. Baris "[context]" di dalamnya BUKAN yang dinilai.\n` +
+      `- <user_profiles> = peta ringkasan kepribadian per user_id; setiap <message> merujuk lewat <user_profile_ref user_id="..."/>.\n` +
+      `- <web_searches> / <web_content> = bukti web (lihat "Web Sebagai Bukti Utama").\n` +
+      `- <messages_to_analyze> = pesan-pesan TARGET yang WAJIB dinilai.`,
+  );
+
+  parts.push(
+    `## Konteks Pengguna (Referensi, Bukan Bukti)\n` +
+      `Konteks per pengguna hanya indikator **referensi** untuk personalisasi analisis, BUKAN bukti pelanggaran:\n` +
+      `- <user_reputation trust_score="..."> = histori moderasi pengguna. Skor rendah BUKAN alasan memflag pesan bersih; skor tinggi BUKAN alasan mengabaikan pelanggaran nyata.\n` +
+      `- <user_profiles> (di pesan USER) = peta ringkasan kepribadian per user_id. <user_profile_ref user_id="..."/> dalam sebuah pesan menunjuk ke peta itu. Tanpa ref = tidak ada profil untuk pengguna tersebut.\n` +
+      `- Profil berguna untuk mengenali penyimpangan perilaku mencolok (mis. pengguna teknis tiba-tiba provokatif), tapi JANGAN memflag atau meloloskan hanya karena profil.\n` +
+      `**Setiap pesan dinilai berdasarkan isinya sendiri.**`,
+  );
+
+  parts.push(
+    `## Framing: Konteks vs Target\n` +
+      `- Baris dalam <conversation_context> berformat "[context] id=... time=<ISO> user=<nama>: isi", diurutkan paling lama → paling baru. Baris pertama biasanya "[conversation_flow] status=... context_msgs=... dropped=..." — metadata sistem tentang status percakapan (ongoing/sparse/cold_start), BUKAN pesan yang dinilai.\n` +
+      `- <messages_to_analyze> berisi pesan-pesan TARGET yang WAJIB dinilai. Hasilkan SATU hasil per message_id — jangan menggabungkan beberapa pesan, jangan melewati, jangan mengarang id.\n` +
+      `- Setiap target dinilai berdasarkan isinya sendiri; konteks percakapan memengaruhi interpretasi, bukan menggantikan isi pesan.\n` +
+      `- Marker "…[pesan dipotong: terlalu panjang]" = konten TARGET sengaja dipotong; marker "…[konteks dipotong: terlalu panjang]" = konten pesan KONTEKS dipotong. Nilai dari bagian yang terlihat; pemotongan BUKAN pelanggaran dan BUKAN teknik evasi.`,
   );
 
   parts.push(OUTPUT_INSTRUCTIONS);
 
-  // XML-delimited context — prevents prompt injection
-  const delimitedContext = `<conversation_context>\n${sanitizeAiContent(contextText, 8000)}\n</conversation_context>`;
-  parts.push(delimitedContext);
-
   let base = parts.join("\n\n");
 
   if (correction) {

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

@@ -19,14 +19,14 @@ import { callModerationLLM } from "./llmCaller.js";
 import { analyzeSingleMediaImage } from "./mediaAnalysisClient.js";
 import {
   buildReferenceXml,
+  buildUserProfileRef,
+  buildUserProfilesBlock,
   escapeXml,
   getAnalysisContent,
   resolveDisplayName,
+  truncateForAi,
 } from "./moderationBuilders.js";
-import {
+import { buildSystemPrompt as buildSystemPromptModular } from "./moderationPrompt.js";
-  buildSystemPrompt as buildSystemPromptModular,
-  sanitizeAiContent,
-} from "./moderationPrompt.js";
 import { logModerationAnalysis } from "./responseLogger.js";
 import {
   extractSearchQueries,
@@ -74,7 +74,7 @@ export async function buildCorrectedFewShotExamples(): Promise<string> {
 // ---------------------------------------------------------------------------
 export async function runTextOnlyBatch(
   targets: MessageRecord[],
-  contextText: string,
+  contextBlock: string,
 ): Promise<{ results: AnalysisResult[]; raw: unknown }> {
   if (!targets.length) return { results: [], raw: null };
 
@@ -188,7 +188,8 @@ export async function runTextOnlyBatch(
     const batch = subBatches[i];
     const targetIds = batch.map((t) => t.id);
 
-    // User reputation + profiles
+    // User reputation + profiles (raw summary text — deduplicated into a
+    // single <user_profiles> map per batch; messages only reference it).
     const userContexts = new Map<string, string>();
     const userProfiles = new Map<string, string>();
     for (const msg of batch) {
@@ -201,14 +202,10 @@ export async function runTextOnlyBatch(
       }
       if (!userProfiles.has(msg.user_id)) {
         const profile = await getUserProfile(msg.user_id);
-        userProfiles.set(
+        userProfiles.set(msg.user_id, profile?.profile_summary ?? "");
-          msg.user_id,
-          profile
-            ? `<user_profile>${sanitizeAiContent(profile.profile_summary)}</user_profile>`
-            : "",
-        );
       }
     }
+    const userProfilesBlock = buildUserProfilesBlock(userProfiles);
 
     // ── URL images → multimodal vision evidence ─────────────────────────
     // The text batch fetches inline URLs; whenever one resolved to an image
@@ -280,7 +277,6 @@ export async function runTextOnlyBatch(
         : undefined;
       const correctedExamples = await buildCorrectedFewShotExamples();
       const systemText = buildSystemPromptModular({
-        contextText,
         mode: batchHasImageEvidence ? "mixed" : "text",
         correction,
         correctedExamples,
@@ -290,7 +286,7 @@ export async function runTextOnlyBatch(
       const messagesBlock = (
         await Promise.all(
           batch.map(async (msg) => {
-            const content = getAnalysisContent(msg);
+            const content = truncateForAi(getAnalysisContent(msg));
             const msgUrls = extractUrlsFromText(content);
             const urlContexts = msgUrls
               .map((url) => {
@@ -307,25 +303,36 @@ export async function runTextOnlyBatch(
               .map((line) => `\n${line}`)
               .join("");
             const userCtx = userContexts.get(msg.user_id) ?? "";
-            const userProfileCtx = userProfiles.get(msg.user_id) ?? "";
+            const userProfileRef = (userProfiles.get(msg.user_id) ?? "").trim()
+              ? buildUserProfileRef(msg.user_id)
+              : "";
             const refXml = await buildReferenceXml(msg);
-            return `<message id="${msg.id}" user="${resolveDisplayName(msg)}">\n  ${userCtx}${userProfileCtx ? `\n  ${userProfileCtx}` : ""}${refXml ? `\n  ${refXml}` : ""}\n  <content>${escapeXml(content)}</content>${webContext}${mediaEvidenceCtx}\n</message>`;
+            return `<message id="${msg.id}" user="${resolveDisplayName(msg)}">\n  ${userCtx}${userProfileRef ? `\n  ${userProfileRef}` : ""}${refXml ? `\n  ${refXml}` : ""}\n  <content>${escapeXml(content)}</content>${webContext}${mediaEvidenceCtx}\n</message>`;
           }),
         )
       ).join("\n");
 
       const searxngBlock =
         searxngResults.size > 0
-          ? `\n\n<web_searches>\n${Array.from(searxngResults.entries())
+          ? `<web_searches>\n${Array.from(searxngResults.entries())
               .map(
                 ([q, xml]) =>
                   `  <search_query query="${escapeXml(q)}">\n${xml}  </search_query>`,
               )
               .join("\n")}\n</web_searches>`
           : "";
+      // Data/instruction separation: the system prompt is stable per mode —
+      // all per-batch context (profiles, conversation, web evidence) lives in
+      // the USER payload, ordered oldest-first so targets come last.
+      const userBlocks = [
+        userProfilesBlock?.trimEnd() ?? "",
+        contextBlock?.trimEnd() ?? "",
+        searxngBlock,
+        `<messages_to_analyze>\n${messagesBlock}\n</messages_to_analyze>`,
+      ].filter((b) => b.trim().length > 0);
       return {
         system: systemText,
-        user: `${searxngBlock}\n\n<messages_to_analyze>\n${messagesBlock}\n</messages_to_analyze>`,
+        user: userBlocks.join("\n\n"),
       };
     };
 

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

@@ -37,16 +37,17 @@ import {
 } from "./mediaDownloader.js";
 import {
   buildReferenceXml,
+  buildUserProfileRef,
   escapeXml,
   getAnalysisContent,
   resolveDisplayName,
+  truncateForAi,
 } from "./moderationBuilders.js";
 import {
   buildCustomEmojiVisionPrompt,
   buildGeneralImageVisionPrompt,
   buildStickerTextOnlyWarning,
   buildStickerVisionPrompt,
-  sanitizeAiContent,
 } from "./moderationPrompt.js";
 import {
   extractSearchQueries,
@@ -367,7 +368,14 @@ export async function prepareMediaMessage(
   const rep = await initializeUserReputation(target.user_id, target.guild_id);
   const profile = await getUserProfile(target.user_id);
   const refXml = await buildReferenceXml(target);
+  // Profile is emitted ONCE per batch in a <user_profiles> map (see
+  // mediaBatchProcessor); here we only reference it to avoid repeating the
+  // full summary on every message of the same user.
+  const profileRef =
+    profile && profile.profile_summary?.trim()
+      ? buildUserProfileRef(target.user_id)
+      : "";
 
-  const messageBlock = `<message id="${escapeXml(target.id)}" user="${escapeXml(resolveDisplayName(target))}">\n  <user_reputation trust_score="${rep.trust_score}" />${profile ? `\n  <user_profile>${sanitizeAiContent(profile.profile_summary)}</user_profile>` : ""}${refXml ? `\n  ${refXml}` : ""}\n  <content>${escapeXml(content)}</content>${mediaContext ? ` ${escapeXml(mediaContext)}` : ""}${webContext}${mediaAnalysisContext}${searxngXml}\n</message>`;
+  const messageBlock = `<message id="${escapeXml(target.id)}" user="${escapeXml(resolveDisplayName(target))}">\n  <user_reputation trust_score="${rep.trust_score}" />${profileRef ? `\n  ${profileRef}` : ""}${refXml ? `\n  ${refXml}` : ""}\n  <content>${escapeXml(truncateForAi(content))}</content>${mediaContext ? ` ${escapeXml(mediaContext)}` : ""}${webContext}${mediaAnalysisContext}${searxngXml}\n</message>`;
   return { targetId, messageBlock };
 }

services/discord-gateway/tests/conversationContext.test.ts

@@ -6,7 +6,9 @@ import {
   buildConversationContext,
   buildLocationContext,
   formatMessageForPrompt,
+  truncateContextLine,
 } from "../src/modules/ai-moderation/conversationContext.js";
+import { buildConversationContextBlock } from "../src/modules/ai-moderation/moderationBuilders.js";
 import { extractOgMeta } from "../src/modules/ai-moderation/urlFetcher.js";
 import type { MessageRecord } from "../src/modules/message-capture/types.js";
 
@@ -162,10 +164,21 @@ describe("formatMessageForPrompt — server nickname (displayName)", () => {
     );
     expect(line).toContain("user=user_n2");
   });
+
+  it("truncates an oversized context message so one paste cannot eat the whole budget", () => {
+    const huge = "A".repeat(5000);
+    const line = formatMessageForPrompt(msg("n3", NOW - MIN, huge), "context");
+    expect(line).toContain("…[konteks dipotong: terlalu panjang]");
+    expect(line.length).toBeLessThan(2000);
+  });
+
+  it("keeps short context content intact", () => {
+    expect(truncateContextLine("pendek")).toBe("pendek");
+  });
 });
 
 describe("buildLocationContext — channel/thread/nsfw enrichment", () => {
-  it("renders channel name + thread name from captured metadata", () => {
+  it("renders a structured <location_context/> element from captured metadata", () => {
     const t = target();
     t.metadata = JSON.stringify({
       channel: {
@@ -176,10 +189,12 @@ describe("buildLocationContext — channel/thread/nsfw enrichment", () => {
       },
     });
     const line = buildLocationContext([t]);
-    expect(line).toContain("[location]");
+    expect(line).toContain("<location_context");
-    expect(line).toContain('name="general"');
+    expect(line).toContain('channel_id="c1"');
+    expect(line).toContain('channel_name="general"');
     expect(line).toContain('thread_name="tanya coding"');
-    expect(line).toContain("nsfw=false");
+    expect(line).toContain('nsfw="false"');
+    expect(line).toContain('age_restricted="false"');
   });
 
   it("returns empty when no metadata", () => {
@@ -187,6 +202,50 @@ describe("buildLocationContext — channel/thread/nsfw enrichment", () => {
   });
 });
 
+describe("buildConversationContextBlock — structured USER-message context", () => {
+  it("wraps location + descriptor + lines into XML blocks", () => {
+    const block = buildConversationContextBlock({
+      location: buildLocationContext(
+        (() => {
+          const t = target();
+          t.metadata = JSON.stringify({
+            channel: { channelName: "general", nsfw: false },
+          });
+          return [t];
+        })(),
+      ),
+      descriptor: "[conversation_flow] status=ongoing context_msgs=1 dropped=0",
+      lines: ["[context] id=a time=2027-01-01T00:00:00.000Z user=user_a: hai"],
+    });
+    expect(block).toContain("<location_context");
+    expect(block).toContain("<conversation_context>");
+    expect(block).toContain("[conversation_flow] status=ongoing");
+    expect(block).toContain("[context] id=a");
+    // location block comes before conversation block
+    expect(block.indexOf("<location_context")).toBeLessThan(
+      block.indexOf("<conversation_context>"),
+    );
+  });
+
+  it("omits the conversation block when there are no lines", () => {
+    const block = buildConversationContextBlock({
+      location: "",
+      descriptor: "",
+      lines: [],
+    });
+    expect(block).toBe("");
+  });
+
+  it("keeps only the location block when lines are empty but location exists", () => {
+    const block = buildConversationContextBlock({
+      location: '<location_context channel_id="c1"/>',
+      descriptor: "",
+      lines: [],
+    });
+    expect(block).toBe('<location_context channel_id="c1"/>');
+  });
+});
+
 describe("extractOgMeta — page title/site for <web_content>", () => {
   it("extracts og:title, og:description and og:site_name", () => {
     const html = `

services/discord-gateway/tests/placeholder.test.ts

@@ -1,6 +1,8 @@
 // ═══════════════════════════════════════════════════════════════════════════════
 // 1. AppError Hierarchy
 // ═══════════════════════════════════════════════════════════════════════════════
+
+import { afterEach, describe, expect, it, vi } from "vitest";
 import {
   AppError,
   ConfigError,
@@ -9,7 +11,6 @@ import {
   UnauthorizedError,
   ValidationError,
 } from "../src/shared/errors/index.js";
-import { afterEach, describe, expect, it, vi } from "vitest";
 
 describe("AppError subclasses", () => {
   it("AppError carries code, statusCode, and details", () => {