From c238564d2707be5527e83b112025443401e7e2fa Mon Sep 17 00:00:00 2001
From: Vincent Koc <vincentkoc@ieee.org>
Date: Tue, 4 Aug 2026 22:27:43 +0800
Subject: [PATCH] fix(discord): count retained speech bytes

---
 extensions/discord/src/voice/manager.e2e.test.ts |  6 ++++--
 extensions/discord/src/voice/realtime.ts         | 15 +++++++++------
 2 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/extensions/discord/src/voice/manager.e2e.test.ts b/extensions/discord/src/voice/manager.e2e.test.ts
index 56fd0bd85256..16dc4ab27011 100644
--- a/extensions/discord/src/voice/manager.e2e.test.ts
+++ b/extensions/discord/src/voice/manager.e2e.test.ts
@@ -4117,7 +4117,7 @@ describe("DiscordVoiceManager", () => {
     expectUserMessageIncludes("third answer");
   });
 
-  it("terminates realtime voice when retained exact speech exceeds the character budget", async () => {
+  it("terminates realtime voice when retained Unicode speech exceeds the byte budget", async () => {
     const client = createClient();
     client.fetchChannel.mockImplementation(async (channelId: string) => {
       const guildId = channelId === "2001" ? "g2" : "g1";
@@ -4138,7 +4138,9 @@ describe("DiscordVoiceManager", () => {
     const connection = (entry as unknown as { connection: { destroy: ReturnType<typeof vi.fn> } })
       .connection;
     const bridgeParams = lastRealtimeBridgeParams();
-    const accepted = "a".repeat(32 * 1024);
+    const accepted = "😀".repeat(8 * 1024);
+    expect(accepted.length).toBe(16 * 1024);
+    expect(Buffer.byteLength(accepted, "utf8")).toBe(32 * 1024);
 
     await manager.join({ guildId: "g2", channelId: "2001" });
     const siblingRealtime = getSessionEntry(manager, "g2").realtime as unknown as {
diff --git a/extensions/discord/src/voice/realtime.ts b/extensions/discord/src/voice/realtime.ts
index 851b63508093..ed18f91ef12d 100644
--- a/extensions/discord/src/voice/realtime.ts
+++ b/extensions/discord/src/voice/realtime.ts
@@ -90,7 +90,7 @@ const DISCORD_REALTIME_DUPLICATE_ERROR_SUPPRESS_MS = 60_000;
 const DISCORD_REALTIME_CONTROL_SPEECH_DEDUPE_MS = 5_000;
 const DISCORD_REALTIME_OUTPUT_PLAYBACK_WATCHDOG_MARGIN_MS = 1_500;
 const DISCORD_REALTIME_MAX_RETAINED_EXACT_SPEECH_MESSAGES = 32;
-const DISCORD_REALTIME_MAX_RETAINED_EXACT_SPEECH_CHARS = 32 * 1024;
+const DISCORD_REALTIME_MAX_RETAINED_EXACT_SPEECH_BYTES = 32 * 1024;
 const DISCORD_REALTIME_CANCELLATION_RACE_DETAIL = "Cancellation failed: no active response found";
 const DISCORD_REALTIME_WAKE_ACKS = ["Yeah.", "Mm-hmm.", "Got it.", "One sec."];
 const discordRealtimeTalkPayload = () => ({});
@@ -999,12 +999,15 @@ export class DiscordRealtimeVoiceSession implements VoiceRealtimeSession {
     }
     const retainedMessages =
       this.queuedExactSpeechMessages.length + (this.activeExactSpeechMessage ? 1 : 0);
-    const retainedChars =
-      this.queuedExactSpeechMessages.reduce((total, message) => total + message.length, 0) +
-      (this.activeExactSpeechMessage?.length ?? 0);
+    const retainedBytes =
+      this.queuedExactSpeechMessages.reduce(
+        (total, message) => total + Buffer.byteLength(message, "utf8"),
+        0,
+      ) + Buffer.byteLength(this.activeExactSpeechMessage ?? "", "utf8");
+    const incomingBytes = Buffer.byteLength(text, "utf8");
     if (
       retainedMessages >= DISCORD_REALTIME_MAX_RETAINED_EXACT_SPEECH_MESSAGES ||
-      retainedChars + text.length > DISCORD_REALTIME_MAX_RETAINED_EXACT_SPEECH_CHARS
+      retainedBytes + incomingBytes > DISCORD_REALTIME_MAX_RETAINED_EXACT_SPEECH_BYTES
     ) {
       // Completed speech cannot be silently dropped. Overflow terminally retires
       // this session before late provider or playback events can drain stale work.
@@ -1019,7 +1022,7 @@ export class DiscordRealtimeVoiceSession implements VoiceRealtimeSession {
       this.clearOutputAudio("exact-speech-overflow");
       this.params.onTerminalError(
         new Error(
-          `Discord realtime exact speech overflow: retained=${retainedMessages} retainedChars=${retainedChars} incomingChars=${text.length}`,
+          `Discord realtime exact speech overflow: retained=${retainedMessages} retainedBytes=${retainedBytes} incomingBytes=${incomingBytes}`,
         ),
       );
       return;