feat: add multimodal analysis support to LLM moderation client by processing image attachments

2026-05-17 23:56:04 +07:00
parent 059d569566
commit 51dc1f8869
21 changed files with 443 additions and 83 deletions
@@ -2,11 +2,7 @@ export class AppError extends Error {
  public code: string;
  public statusCode: number;

-  constructor(
-    message: string,
-    code: string,
-    statusCode: number = 500,
-  ) {
+  constructor(message: string, code: string, statusCode: number = 500) {
    super(message);
    this.code = code;
    this.statusCode = statusCode;
@@ -39,10 +35,7 @@ export class VoiceConnectionError extends AppError {
 export class ValidationError extends AppError {
  public details?: Record<string, string[]>;

-  constructor(
-    message: string,
-    details?: Record<string, string[]>,
-  ) {
+  constructor(message: string, details?: Record<string, string[]>) {
    super(message, "VALIDATION_ERROR", 400);
    this.details = details;
    this.name = "ValidationError";
@@ -74,7 +74,12 @@ async function initializeApp() {
  }

  client.on("debug", (msg) => {
-    if (msg.includes("[VOICE") || msg.includes("[ffmpeg") || msg.toLowerCase().includes("error") || msg.toLowerCase().includes("stream")) {
+    if (
+      msg.includes("[VOICE") ||
+      msg.includes("[ffmpeg") ||
+      msg.toLowerCase().includes("error") ||
+      msg.toLowerCase().includes("stream")
+    ) {
      logger.info({ debugMsg: msg }, "Discord Client Debug");
    } else if (config.VERBOSE) {
      logger.debug({ debugMsg: msg }, "Discord Client Debug");
@@ -17,7 +17,10 @@ import { createMusicPlayer } from "./musicPlayer";
 export interface MediaControllerDependencies {
  isVoiceConnected?: () => boolean;
  isBrowserStreaming?: () => boolean;
-  resolveMediaSource?: (source: string, mode?: MediaMode) => Promise<ResolvedMediaSource>;
+  resolveMediaSource?: (
+    source: string,
+    mode?: MediaMode,
+  ) => Promise<ResolvedMediaSource>;
  musicPlayer?: MusicPlayer;
  screenController?: ScreenShareController;
  onStateChange?: (state: MediaState) => void;
@@ -91,7 +94,10 @@ export class MediaController {
    // reject to avoid stealing the shared player. If this controller started
    // the screenPlayback, stop it and proceed.
    if (this.screenPlayback || this.dependencies.screenController?.isActive()) {
-      if (this.dependencies.screenController?.isActive() && !this.screenPlayback) {
+      if (
+        this.dependencies.screenController?.isActive() &&
+        !this.screenPlayback
+      ) {
        throw new AppError("Another media mode is active", "MEDIA_BUSY", 409);
      }
      this.screenPlayback?.stop();
@@ -20,7 +20,7 @@ export function createMediaResolver(

  return async function resolve(
    input: string,
-    mode: MediaMode = "music"
+    mode: MediaMode = "music",
  ): Promise<ResolvedMediaSource> {
    const source = input.trim();
    if (!source) {
@@ -34,17 +34,19 @@ export function createMediaResolver(
    const url = parseUrl(source);
    if (url && isYouTubeUrl(url)) {
      const metadata = await ytdlp.getMetadata(source);
-      const directUrl = mode === "screen" 
-        ? await ytdlp.getDirectVideoUrl(source)
-        : await ytdlp.getDirectAudioUrl(source);
+      const directUrl =
+        mode === "screen"
+          ? await ytdlp.getDirectVideoUrl(source)
+          : await ytdlp.getDirectAudioUrl(source);
      return { source: directUrl, title: metadata.title, kind: "youtube" };
    }

    if (url && isSpotifyTrackUrl(url)) {
      const result = await playDlResolver.resolveSpotifyTrack(source);
-      const directUrl = mode === "screen"
-        ? await ytdlp.getDirectVideoUrl(result.url)
-        : await ytdlp.getDirectAudioUrl(result.url);
+      const directUrl =
+        mode === "screen"
+          ? await ytdlp.getDirectVideoUrl(result.url)
+          : await ytdlp.getDirectAudioUrl(result.url);
      return { source: directUrl, title: result.title, kind: "spotify" };
    }

@@ -62,9 +64,10 @@ export function createMediaResolver(

    if (!url && !looksLikeUrl(source)) {
      const result = await playDlResolver.searchYouTube(source);
-      const directUrl = mode === "screen"
-        ? await ytdlp.getDirectVideoUrl(result.url)
-        : await ytdlp.getDirectAudioUrl(result.url);
+      const directUrl =
+        mode === "screen"
+          ? await ytdlp.getDirectVideoUrl(result.url)
+          : await ytdlp.getDirectAudioUrl(result.url);
      return { source: directUrl, title: result.title, kind: "search" };
    }

@@ -1,7 +1,4 @@
-import {
-  Streamer,
-  playPreparedStream,
-} from "../streaming";
+import { Streamer, playPreparedStream } from "../streaming";
 import { AppError } from "../errors";
 import { createChildLogger } from "../logger";
 import { discordPlayer } from "../player";
@@ -23,8 +20,14 @@ export interface ScreenShareControllerDependencies {
  getDirectVideoUrl?: (source: string) => Promise<string>;
  streamer: Streamer;
  useTranscoder?: boolean;
-  onBeforeStreamStart?: (guildId: string, channelId: string) => Promise<void> | void;
-  onAfterStreamEnd?: (guildId: string, channelId: string) => Promise<void> | void;
+  onBeforeStreamStart?: (
+    guildId: string,
+    channelId: string,
+  ) => Promise<void> | void;
+  onAfterStreamEnd?: (
+    guildId: string,
+    channelId: string,
+  ) => Promise<void> | void;
  onStreamStart?: () => void;
  onStreamEnd?: () => void;
 }
@@ -3,6 +3,7 @@ import { initializeDatabase } from "../database/drizzle.ts";
 import { buildConversationPromptMessages } from "./conversationContext.ts";
 import { runModerationAnalysis } from "./llmModerationClient.ts";
 import {
+  getAttachmentsForMessages,
  getConversationContextBefore,
  updateMessageAIAnalysis,
 } from "./messageStore.ts";
@@ -66,9 +67,15 @@ async function processAnalysisRequest({
      maxTokens: MAX_CONTEXT_TOKENS,
    });

+    const targetIds = messages.map((m) => m.id);
+    const contextIds = contextBefore.map((m) => m.id);
+    const allMessageIds = [...targetIds, ...contextIds];
+    const attachments = await getAttachmentsForMessages(allMessageIds);
+
    const result = await runModerationAnalysis({
      targets: messages,
      contextText: promptMessages.join("\n"),
+      attachments,
    });

    const rows: MessageRecord[] = [];
@@ -146,7 +146,7 @@ async function runAnalysisInWorker(
  return new Promise((resolve, reject) => {
    const worker = new Worker(
      new URL("./aiAnalysisWorker.ts", import.meta.url),
-      { execArgv: process.execArgv }
+      { execArgv: process.execArgv },
    );

    worker.once("message", (response: AnalysisWorkerResponse) => {
@@ -1,7 +1,7 @@
 import { config } from "../config.ts";
 import { createChildLogger } from "../logger.ts";
 import { retryWithBackoff } from "../retry.ts";
-import type { AnalysisResult, MessageRecord } from "./types";
+import type { AnalysisResult, AttachmentRecord, MessageRecord } from "./types";

 const log = createChildLogger("llmModerationClient");

@@ -174,6 +174,7 @@ export function parseModerationResponse(
 interface ModerationInput {
  targets: MessageRecord[];
  contextText: string;
+  attachments?: AttachmentRecord[];
 }

 interface ModerationOutput {
@@ -188,7 +189,7 @@ interface ModerationOutput {
 export async function runModerationAnalysis(
  input: ModerationInput,
 ): Promise<ModerationOutput> {
-  const { targets, contextText } = input;
+  const { targets, contextText, attachments } = input;

  if (!targets.length) {
    throw new Error("No targets provided for analysis");
@@ -220,6 +221,88 @@ Each result must have:

 Return ONLY valid JSON, no other text.`;

+  // Check for image attachments to support multimodal analysis
+  const imageAttachments = (attachments || []).filter(
+    (att) =>
+      (att.uploaded_url || att.discord_url) && att.type.startsWith("image/"),
+  );
+
+  let messageContent:
+    | string
+    | Array<{ type: string; text?: string; image_url?: { url: string } }>;
+  if (imageAttachments.length > 0) {
+    const contentParts: Array<{
+      type: string;
+      text?: string;
+      image_url?: { url: string };
+    }> = [];
+
+    // Download and convert all images to base64 data URLs
+    for (const att of imageAttachments) {
+      try {
+        const urlToUse = att.uploaded_url || att.discord_url;
+        log.info(
+          { attachmentId: att.id, url: urlToUse },
+          "Downloading attachment for base64 encoding",
+        );
+        const res = await fetch(urlToUse);
+        if (res.ok) {
+          const buffer = await res.arrayBuffer();
+          const base64Str = Buffer.from(buffer).toString("base64");
+          const dataUrl = `data:${att.type};base64,${base64Str}`;
+
+          contentParts.push({
+            type: "image_url",
+            image_url: {
+              url: dataUrl,
+            },
+          });
+
+          contentParts.push({
+            type: "text",
+            text: `\n[Image Attachment for Message ID: ${att.message_id}, Filename: ${att.filename}]`,
+          });
+        } else {
+          log.warn(
+            { attachmentId: att.id, status: res.status },
+            "Failed to fetch attachment image",
+          );
+        }
+      } catch (err) {
+        log.warn(
+          {
+            attachmentId: att.id,
+            error: err instanceof Error ? err.message : String(err),
+          },
+          "Error base64 encoding attachment",
+        );
+      }
+    }
+
+    contentParts.push({
+      type: "text",
+      text: prompt,
+    });
+
+    messageContent = contentParts;
+  } else {
+    // If no image is present, send a transparent 1x1 dummy PNG to satisfy multimodal omni requirements
+    const dummyPng =
+      "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg==";
+    messageContent = [
+      {
+        type: "image_url",
+        image_url: {
+          url: dummyPng,
+        },
+      },
+      {
+        type: "text",
+        text: prompt,
+      },
+    ];
+  }
+
  const result = await retryWithBackoff(
    async () => {
      const controller = new AbortController();
@@ -243,10 +326,14 @@ Return ONLY valid JSON, no other text.`;
              messages: [
                {
                  role: "user",
-                  content: prompt,
+                  content: messageContent,
                },
              ],
-              temperature: 0.3,
+              temperature: 0.6,
+              top_p: 0.95,
+              max_tokens: 65536,
+              reasoning_budget: 16384,
+              chat_template_kwargs: { enable_thinking: true },
            }),
          },
        );
@@ -1,4 +1,14 @@
-import { and, asc, desc, eq, isNull, or, type SQL, sql } from "drizzle-orm";
+import {
+  and,
+  asc,
+  desc,
+  eq,
+  inArray,
+  isNull,
+  or,
+  type SQL,
+  sql,
+} from "drizzle-orm";
 import { getDatabase } from "../database/drizzle.ts";
 import { attachmentsTable, messagesTable } from "../database/schema.ts";
 import { createChildLogger } from "../logger.ts";
@@ -605,3 +615,27 @@ export async function getPendingConversationKeys(
    throw error;
  }
 }
+
+export async function getAttachmentsForMessages(
+  messageIds: string[],
+): Promise<AttachmentRecord[]> {
+  try {
+    if (messageIds.length === 0) return [];
+    const database = db();
+    const rows = await database
+      .select()
+      .from(attachmentsTable)
+      .where(inArray(attachmentsTable.message_id, messageIds));
+
+    return rows as AttachmentRecord[];
+  } catch (error) {
+    logger.error(
+      {
+        messageIds,
+        error: error instanceof Error ? error.message : String(error),
+      },
+      "Failed to get attachments for messages",
+    );
+    throw error;
+  }
+}
@@ -41,7 +41,10 @@ export class Streamer {
    this.dankStreamer = new DankStreamer(client);
  }

-  async createSession(guildId: string, channelId: string): Promise<StreamSession> {
+  async createSession(
+    guildId: string,
+    channelId: string,
+  ): Promise<StreamSession> {
    await this.dankStreamer.joinVoice(guildId, channelId);

    let stopped = false;
@@ -62,7 +65,10 @@ export class Streamer {
    return {
      connection: {} as any,
      stream: {} as any,
-      play: async (source: string | Readable, options: StreamPlayOptions = {}) => {
+      play: async (
+        source: string | Readable,
+        options: StreamPlayOptions = {},
+      ) => {
        if (stopped) return;

        let targetSource: string | Readable = source;
@@ -75,7 +81,12 @@ export class Streamer {
        const bitrateStr = String(options.bitrate ?? 8000).replace(/k$/i, "");
        const bitrateVideo = parseInt(bitrateStr, 10) || 8000;

-        console.log("[Streamer] Starting screen share for source:", typeof targetSource === "string" ? targetSource.slice(0, 50) + "..." : "ReadableStream");
+        console.log(
+          "[Streamer] Starting screen share for source:",
+          typeof targetSource === "string"
+            ? targetSource.slice(0, 50) + "..."
+            : "ReadableStream",
+        );
        const { command, output } = dankPrepareStream(targetSource, {
          encoder: Encoders.software({
            x264: { preset: (options.presetH26x as any) ?? "ultrafast" },
@@ -100,7 +111,7 @@ export class Streamer {

        const webOutput = new PassThrough();
        const discordOutput = new PassThrough();
-        
+
        output.pipe(webOutput);
        output.pipe(discordOutput);

@@ -21,7 +21,10 @@ export class Transcoder {
  restartTimer: NodeJS.Timeout | null = null;
  maxRestarts = 6;

-  constructor(private source: string, private opts: TranscoderOptions = {}) {}
+  constructor(
+    private source: string,
+    private opts: TranscoderOptions = {},
+  ) {}

  start(): { command: ChildProcess; output: Readable } {
    const fps = this.opts.fps ?? 30;
@@ -99,13 +102,19 @@ export class Transcoder {

  scheduleRestart() {
    if (this.restartAttempts >= this.maxRestarts) {
-      logger.error({ attempts: this.restartAttempts }, "transcoder reached max restart attempts");
+      logger.error(
+        { attempts: this.restartAttempts },
+        "transcoder reached max restart attempts",
+      );
      return;
    }
    const delay = Math.min(30000, 1000 * Math.pow(2, this.restartAttempts));
    this.restartAttempts += 1;
    transcoderRestartsCounter.inc();
-    logger.info({ delay, attempt: this.restartAttempts }, "scheduling transcoder restart");
+    logger.info(
+      { delay, attempt: this.restartAttempts },
+      "scheduling transcoder restart",
+    );
    this.restartTimer = setTimeout(() => {
      try {
        this.start();
@@ -129,9 +138,9 @@ export class Transcoder {
      clearTimeout(this.restartTimer);
      this.restartTimer = null;
    }
-      if (this.proc && !this.proc.killed) {
-        return new Promise<void>((resolve) => {
-          this.proc?.once("exit", () => resolve());
+    if (this.proc && !this.proc.killed) {
+      return new Promise<void>((resolve) => {
+        this.proc?.once("exit", () => resolve());
        try {
          this.proc?.kill("SIGTERM");
        } catch {
@@ -141,7 +150,7 @@ export class Transcoder {
            resolve();
          }
        }
-          setTimeout(() => resolve(), 5000);
+        setTimeout(() => resolve(), 5000);
      }).then(() => {
        this.proc = null;
        this.output = null;
@@ -151,7 +160,10 @@ export class Transcoder {
  }
 }

-export function prepareTranscoder(source: string, options: TranscoderOptions = {}) {
+export function prepareTranscoder(
+  source: string,
+  options: TranscoderOptions = {},
+) {
  const t = new Transcoder(source, options);
  const { command, output } = t.start();
  return { transcoder: t, command, output };
@@ -3,13 +3,16 @@ import fs from "fs";

 async function run() {
  console.log("Starting prepareStream...");
-  const { command, output } = prepareStream("https://rr3---sn-2uuxa3vh-unte.googlevideo.com/videoplayback?expire=1779046518&ei=FsQJatGDGNqp9fwP4qz4SA&ip=180.252.24.35&id=o-APFvGry6yPgoap-1RT0pu59DxD-pcXC4oXtMQuCMtjOy&itag=18&source=youtube&requiressl=yes&xpc=EgVo2aDSNQ%3D%3D&cps=618&met=1779024918%2C&mh=VD&mm=31%2C29&mn=sn-2uuxa3vh-unte%2Csn-oguelnze&ms=au%2Crdu&mv=m&mvi=3&pcm2cms=yes&pl=20&rms=au%2Cau&initcwndbps=763750&bui=AbKmrwofOLw_tOID4kBHnWgaXP2wnDlEYmbyHyrnZk1n7vjMaQIuY046T9MhH0PuL9JGJwj6YlwCr2Uu&spc=96Xrv8WI7iTS7MOF7Dvg-8a3RT-sMI9ux49zUa4Pg6GHkzXExSS0&vprv=1&svpuc=1&mime=video%2Fmp4&rqh=1&cnr=14&ratebypass=yes&dur=19.063&lmt=1772437158054287&mt=1779024581&fvip=4&fexp=51565116%2C51565681&c=ANDROID_VR&txp=4530534&sparams=expire%2Cei%2Cip%2Cid%2Citag%2Csource%2Crequiressl%2Cxpc%2Cbui%2Cspc%2Cvprv%2Csvpuc%2Cmime%2Crqh%2Ccnr%2Cratebypass%2Cdur%2Clmt&sig=AHEqNM4wRgIhAJe1vu37ssUQQm3scVgXY7NYDx_frKW1AZ4gHRdcqsUlAiEAkKt6jxaCNvaEh6jag1OWheo5qQeu3ObfCCoQIZ9xnCA%3D&lsparams=cps%2Cmet%2Cmh%2Cmm%2Cmn%2Cms%2Cmv%2Cmvi%2Cpcm2cms%2Cpl%2Crms%2Cinitcwndbps&lsig=APaTxxMwRQIhAMkeJ6WrDFU7fTfSb6s_WbdDpn4J-4NqkfzKV3B_y1cgAiBJ7aExkhh-0hvIWwNorjDwoOkTIKIfmzx6o6Z3mxlazA%3D%3D", {
-    encoder: Encoders.software(),
-    width: 1280,
-    height: 720,
-    includeAudio: true,
-    minimizeLatency: false // Add this
-  });
+  const { command, output } = prepareStream(
+    "https://rr3---sn-2uuxa3vh-unte.googlevideo.com/videoplayback?expire=1779046518&ei=FsQJatGDGNqp9fwP4qz4SA&ip=180.252.24.35&id=o-APFvGry6yPgoap-1RT0pu59DxD-pcXC4oXtMQuCMtjOy&itag=18&source=youtube&requiressl=yes&xpc=EgVo2aDSNQ%3D%3D&cps=618&met=1779024918%2C&mh=VD&mm=31%2C29&mn=sn-2uuxa3vh-unte%2Csn-oguelnze&ms=au%2Crdu&mv=m&mvi=3&pcm2cms=yes&pl=20&rms=au%2Cau&initcwndbps=763750&bui=AbKmrwofOLw_tOID4kBHnWgaXP2wnDlEYmbyHyrnZk1n7vjMaQIuY046T9MhH0PuL9JGJwj6YlwCr2Uu&spc=96Xrv8WI7iTS7MOF7Dvg-8a3RT-sMI9ux49zUa4Pg6GHkzXExSS0&vprv=1&svpuc=1&mime=video%2Fmp4&rqh=1&cnr=14&ratebypass=yes&dur=19.063&lmt=1772437158054287&mt=1779024581&fvip=4&fexp=51565116%2C51565681&c=ANDROID_VR&txp=4530534&sparams=expire%2Cei%2Cip%2Cid%2Citag%2Csource%2Crequiressl%2Cxpc%2Cbui%2Cspc%2Cvprv%2Csvpuc%2Cmime%2Crqh%2Ccnr%2Cratebypass%2Cdur%2Clmt&sig=AHEqNM4wRgIhAJe1vu37ssUQQm3scVgXY7NYDx_frKW1AZ4gHRdcqsUlAiEAkKt6jxaCNvaEh6jag1OWheo5qQeu3ObfCCoQIZ9xnCA%3D&lsparams=cps%2Cmet%2Cmh%2Cmm%2Cmn%2Cms%2Cmv%2Cmvi%2Cpcm2cms%2Cpl%2Crms%2Cinitcwndbps&lsig=APaTxxMwRQIhAMkeJ6WrDFU7fTfSb6s_WbdDpn4J-4NqkfzKV3B_y1cgAiBJ7aExkhh-0hvIWwNorjDwoOkTIKIfmzx6o6Z3mxlazA%3D%3D",
+    {
+      encoder: Encoders.software(),
+      width: 1280,
+      height: 720,
+      includeAudio: true,
+      minimizeLatency: false, // Add this
+    },
+  );

  const fileStream = fs.createWriteStream("/mnt/code/bete/test_out.nut");
  output.pipe(fileStream);
@@ -26,7 +29,9 @@ async function run() {
  });

  setTimeout(() => {
-    try { command.kill("SIGKILL"); } catch(e) {}
+    try {
+      command.kill("SIGKILL");
+    } catch (e) {}
    process.exit(0);
  }, 10000);
 }
@@ -3,24 +3,29 @@ import { demux } from "@dank074/discord-video-stream/dist/media/LibavDemuxer.js"

 async function run() {
  console.log("Starting prepareStream...");
-  const { command, output } = prepareStream("https://samplelib.com/preview/mp4/sample-5s.mp4", {
-    encoder: Encoders.software(),
-    width: 1280,
-    height: 720,
-    includeAudio: true,
-    minimizeLatency: false // Add this
-  });
+  const { command, output } = prepareStream(
+    "https://samplelib.com/preview/mp4/sample-5s.mp4",
+    {
+      encoder: Encoders.software(),
+      width: 1280,
+      height: 720,
+      includeAudio: true,
+      minimizeLatency: false, // Add this
+    },
+  );

  try {
    const { video, audio } = await demux(output, { format: "nut" });
    console.log("DEMUX VIDEO:", !!video);
    console.log("DEMUX AUDIO:", !!audio);
-  } catch(e) {
+  } catch (e) {
    console.error("DEMUX ERR:", e);
  }

  setTimeout(() => {
-    try { command.kill("SIGKILL"); } catch(e) {}
+    try {
+      command.kill("SIGKILL");
+    } catch (e) {}
    process.exit(0);
  }, 10000);
 }
@@ -3,12 +3,15 @@ import { demux } from "@dank074/discord-video-stream/dist/media/LibavDemuxer.js"
 import { Encoders } from "@dank074/discord-video-stream/dist/media/encoders/index.js";

 async function run() {
-  const { command, output } = prepareStream("http://commondatastorage.googleapis.com/gtv-videos-bucket/sample/BigBuckBunny.mp4", {
-    encoder: Encoders.software(),
-    width: 1280,
-    height: 720,
-    includeAudio: true
-  });
+  const { command, output } = prepareStream(
+    "http://commondatastorage.googleapis.com/gtv-videos-bucket/sample/BigBuckBunny.mp4",
+    {
+      encoder: Encoders.software(),
+      width: 1280,
+      height: 720,
+      includeAudio: true,
+    },
+  );

  const { video, audio } = await demux(output, { format: "nut" });
  console.log("Video found:", !!video);
@@ -225,7 +225,10 @@ describe("MediaController", () => {
    expect(screenController.start).toHaveBeenCalledWith(
      "https://youtu.be/video",
    );
-    expect(resolveMediaSource).toHaveBeenCalledWith("https://youtu.be/video", "screen");
+    expect(resolveMediaSource).toHaveBeenCalledWith(
+      "https://youtu.be/video",
+      "screen",
+    );
    expect(state).toMatchObject({ playing: true, activeMode: "screen" });
  });

@@ -71,10 +71,14 @@ describe("createMusicPlayer", () => {
      ],
      { stdio: ["ignore", "pipe", "pipe"] },
    );
-    expect(discordPlayer.playStream).toHaveBeenCalledWith(proc.stdout, "music", {
-      inputType: StreamType.Raw,
-      inlineVolume: true,
-    });
+    expect(discordPlayer.playStream).toHaveBeenCalledWith(
+      proc.stdout,
+      "music",
+      {
+        inputType: StreamType.Raw,
+        inlineVolume: true,
+      },
+    );
  });

  it("rejects playback when Discord is not connected", () => {
@@ -336,4 +336,100 @@ describe("runModerationAnalysis", () => {
      }),
    ).rejects.toThrow(/No content in LLM response/);
  });
+
+  it("sends multimodal payload when image attachments are present", async () => {
+    const mockResponse = {
+      choices: [
+        {
+          message: {
+            content: JSON.stringify({
+              results: [
+                {
+                  message_id: "m1",
+                  status: "clean",
+                  flags: [],
+                  score: 0.1,
+                  analysis: "OK",
+                },
+              ],
+            }),
+          },
+        },
+      ],
+    };
+
+    global.fetch = vi.fn().mockImplementation((url: string) => {
+      if (url.includes("picser.tech") || url.includes("discord.com")) {
+        return Promise.resolve({
+          ok: true,
+          arrayBuffer: async () => {
+            const buffer = Buffer.from("fake-image-bytes");
+            return buffer.buffer.slice(
+              buffer.byteOffset,
+              buffer.byteOffset + buffer.byteLength,
+            );
+          },
+        });
+      }
+      return Promise.resolve({
+        ok: true,
+        text: async () => JSON.stringify(mockResponse),
+        json: async () => mockResponse,
+      });
+    });
+
+    const mockAttachment = {
+      id: "a1",
+      message_id: "m1",
+      guild_id: "guild123",
+      channel_id: "channel123",
+      thread_id: null,
+      user_id: "user123",
+      filename: "test.png",
+      size: 500,
+      type: "image/png",
+      discord_url: "https://discord.com/attachment.png",
+      uploaded_url: "https://picser.tech/test.png",
+      upload_status: "uploaded" as const,
+      upload_error: null,
+      created_at: Date.now(),
+      uploaded_at: Date.now(),
+    };
+
+    const result = await runModerationAnalysis({
+      targets: [createMessageRecord()],
+      contextText: "test context",
+      attachments: [mockAttachment],
+    });
+
+    expect(result.results).toHaveLength(1);
+    expect(global.fetch).toHaveBeenCalled();
+
+    const fetchCalls = (global.fetch as any).mock.calls;
+    // Should be called twice: 1st for image download, 2nd for API completions
+    expect(fetchCalls.length).toBe(2);
+
+    // Verify 1st call (image download)
+    expect(fetchCalls[0][0]).toBe("https://picser.tech/test.png");
+
+    // Verify 2nd call (chat completions API)
+    const [, completionsOptions] = fetchCalls[1];
+    const body = JSON.parse(completionsOptions.body);
+
+    const userMessage = body.messages[0];
+    expect(userMessage.role).toBe("user");
+    expect(Array.isArray(userMessage.content)).toBe(true);
+    expect(userMessage.content[0].type).toBe("image_url");
+    expect(userMessage.content[0].image_url.url).toContain(
+      "data:image/png;base64,",
+    );
+    expect(userMessage.content[1].type).toBe("text");
+    expect(userMessage.content[1].text).toContain(
+      "Image Attachment for Message ID: m1",
+    );
+    expect(userMessage.content[2].type).toBe("text");
+    expect(userMessage.content[2].text).toContain(
+      "You are a content moderation assistant.",
+    );
+  });
 });
@@ -8,7 +8,9 @@ import { createChildLogger } from "../../src/logger";
 import {
  decodeCursor,
  encodeCursor,
+  getAttachmentsForMessages,
  getMessageById,
+  insertAttachment,
  insertMessage,
  listMessages,
  listReviewMessages,
@@ -72,21 +74,40 @@ describe("message query integration tests", () => {
          "ai_error" text
        )
      `);
+
+      // Create attachments table
+      await db.run(`
+        CREATE TABLE IF NOT EXISTS "attachments" (
+          "id" text PRIMARY KEY NOT NULL,
+          "message_id" text NOT NULL,
+          "guild_id" text NOT NULL,
+          "channel_id" text NOT NULL,
+          "thread_id" text,
+          "user_id" text NOT NULL,
+          "filename" text NOT NULL,
+          "size" integer NOT NULL,
+          "type" text NOT NULL,
+          "discord_url" text NOT NULL,
+          "uploaded_url" text,
+          "upload_status" text DEFAULT 'pending' NOT NULL,
+          "upload_error" text,
+          "created_at" integer NOT NULL,
+          "uploaded_at" integer
+        )
+      `);
    } catch (error) {
-      logger.debug(
-        { error },
-        "Messages table already exists or error creating it",
-      );
+      logger.debug({ error }, "Tables already exist or error creating them");
    }
  });

  beforeEach(async () => {
-    // Clear messages table before each test
+    // Clear tables before each test
    try {
      const db = getTestDatabase();
      await db.run(`DELETE FROM "messages"`);
+      await db.run(`DELETE FROM "attachments"`);
    } catch (error) {
-      logger.debug({ error }, "Could not clear messages table");
+      logger.debug({ error }, "Could not clear tables");
    }
  });

@@ -579,4 +600,58 @@ describe("message query integration tests", () => {
      expect(retrieved?.ai_error).toBeNull();
    });
  });
+
+  describe("getAttachmentsForMessages", () => {
+    it("returns attachments matching given message IDs", async () => {
+      const msgId1 = "msg-att-1";
+      const msgId2 = "msg-att-2";
+
+      const attachment1 = {
+        id: "att-1",
+        message_id: msgId1,
+        guild_id: "guild-123",
+        channel_id: "channel-456",
+        thread_id: null,
+        user_id: "user-789",
+        filename: "test1.png",
+        size: 1024,
+        type: "image/png",
+        discord_url: "https://discord.com/test1.png",
+        uploaded_url: "https://picser.tech/test1.png",
+        upload_status: "uploaded" as const,
+        upload_error: null,
+        created_at: Date.now(),
+        uploaded_at: Date.now(),
+      };
+
+      const attachment2 = {
+        id: "att-2",
+        message_id: msgId2,
+        guild_id: "guild-123",
+        channel_id: "channel-456",
+        thread_id: null,
+        user_id: "user-789",
+        filename: "test2.png",
+        size: 2048,
+        type: "image/png",
+        discord_url: "https://discord.com/test2.png",
+        uploaded_url: "https://picser.tech/test2.png",
+        upload_status: "uploaded" as const,
+        upload_error: null,
+        created_at: Date.now(),
+        uploaded_at: Date.now(),
+      };
+
+      await insertAttachment(attachment1);
+      await insertAttachment(attachment2);
+
+      const result = await getAttachmentsForMessages([msgId1, msgId2]);
+      expect(result).toHaveLength(2);
+      const ids = result.map((r) => r.id).sort();
+      expect(ids).toEqual(["att-1", "att-2"].sort());
+
+      const emptyResult = await getAttachmentsForMessages([]);
+      expect(emptyResult).toHaveLength(0);
+    });
+  });
 });
@@ -38,7 +38,8 @@ describe("playTranscodedPreparedStream", () => {
  it("pipes transcoder output to session and broadcasts to web", async () => {
    // mock global broadcast
    const broadcasts: Buffer[] = [];
-    (globalThis as any).broadcastVideoToWeb = (chunk: Buffer) => broadcasts.push(Buffer.from(chunk));
+    (globalThis as any).broadcastVideoToWeb = (chunk: Buffer) =>
+      broadcasts.push(Buffer.from(chunk));

    const session = {
      connection: { channel: { id: "c" } },
@@ -52,7 +53,9 @@ describe("playTranscodedPreparedStream", () => {
      stop: vi.fn(),
    } as any;

-    await playTranscodedPreparedStream("http://example.test/stream", session, { fps: 30 });
+    await playTranscodedPreparedStream("http://example.test/stream", session, {
+      fps: 30,
+    });
    expect(session.play).toHaveBeenCalled();
    expect(broadcasts.length).toBeGreaterThanOrEqual(0);
  });
@@ -39,7 +39,10 @@ import { prepareTranscoder } from "../../src/streaming/transcoder";

 describe("Transcoder", () => {
  it("starts ffmpeg and returns output stream and command", () => {
-    const { transcoder, command, output } = prepareTranscoder("http://example.test/video", { fps: 24 });
+    const { transcoder, command, output } = prepareTranscoder(
+      "http://example.test/video",
+      { fps: 24 },
+    );
    expect(transcoder).toBeTruthy();
    expect(command).toBeTruthy();
    expect(output).toBeTruthy();
@@ -3,6 +3,8 @@
    "target": "ESNext",
    "module": "ESNext",
    "moduleResolution": "bundler",
+    "allowImportingTsExtensions": true,
+    "noEmit": true,
    "strict": true,
    "esModuleInterop": true,
    "skipLibCheck": true,