fix(discord): keep realtime partial transcript tail UTF-16 safe (#104896)

* fix(discord): keep realtime partial transcript tail UTF-16 safe

* refactor(discord): isolate realtime transcript merging

---------

Co-authored-by: Peter Steinberger <steipete@gmail.com>

extensions/discord/src/voice/realtime-transcript.test.ts

@@ -0,0 +1,30 @@
+import { describe, expect, it } from "vitest";
+import { mergeRealtimePartialTranscript } from "./realtime-transcript.js";
+
+describe("mergeRealtimePartialTranscript", () => {
+  it("returns previous transcript when the next chunk is blank", () => {
+    expect(mergeRealtimePartialTranscript("hello", "   ")).toBe("hello");
+  });
+
+  it("replaces with the growing chunk when it extends the previous prefix", () => {
+    expect(mergeRealtimePartialTranscript("hel", "hello world")).toBe("hello world");
+  });
+
+  it("appends when the next chunk is not a continuation of previous", () => {
+    expect(mergeRealtimePartialTranscript("hello", " there")).toBe("hello there");
+  });
+
+  it("does not split a surrogate pair at the tail cap boundary", () => {
+    const tail = "x".repeat(239);
+    const next = `${"y".repeat(50)}🦞${tail}`;
+
+    expect(mergeRealtimePartialTranscript("", next)).toBe(tail);
+  });
+
+  it("keeps an intact surrogate pair that sits just inside the cap", () => {
+    const tail = `🦞${"w".repeat(238)}`;
+    const next = `${"z".repeat(10)}${tail}`;
+
+    expect(mergeRealtimePartialTranscript("", next)).toBe(tail);
+  });
+});

extensions/discord/src/voice/realtime-transcript.ts

@@ -0,0 +1,13 @@
+// Discord plugin module owns realtime transcript accumulation.
+import { sliceUtf16Safe } from "openclaw/plugin-sdk/text-utility-runtime";
+
+const PARTIAL_TRANSCRIPT_MAX_CHARS = 240;
+
+export function mergeRealtimePartialTranscript(previous: string, next: string): string {
+  const trimmed = next.trim();
+  if (!trimmed) {
+    return previous;
+  }
+  const merged = trimmed.startsWith(previous) ? trimmed : `${previous}${next}`;
+  return sliceUtf16Safe(merged, -PARTIAL_TRANSCRIPT_MAX_CHARS);
+}

extensions/discord/src/voice/realtime.ts

@@ -53,6 +53,7 @@ import {
 } from "./audio.js";
 import { formatVoiceLogPreview } from "./log-preview.js";
 import { formatVoiceIngressPrompt } from "./prompt.js";
+import { mergeRealtimePartialTranscript } from "./realtime-transcript.js";
 import { loadDiscordVoiceSdk } from "./sdk-runtime.js";
 import {
   logVoiceVerbose,
@@ -88,7 +89,6 @@ 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_WAKE_ACKS = ["Yeah.", "Mm-hmm.", "Got it.", "One sec."];
-const DISCORD_REALTIME_PARTIAL_TRANSCRIPT_MAX_CHARS = 240;
 const REALTIME_PCM16_BYTES_PER_SAMPLE = 2;
 const DISCORD_RAW_PCM_FRAME_BYTES = 3_840;
 const DISCORD_REALTIME_OUTPUT_PREROLL_FRAMES = 25;
@@ -314,15 +314,6 @@ function normalizeControlSpeechText(text: string): string {
   return text.toLowerCase().replace(/\s+/g, " ").trim();
 }
 
-function mergeRealtimePartialTranscript(previous: string, next: string): string {
-  const trimmed = next.trim();
-  if (!trimmed) {
-    return previous;
-  }
-  const merged = trimmed.startsWith(previous) ? trimmed : `${previous}${next}`;
-  return merged.slice(-DISCORD_REALTIME_PARTIAL_TRANSCRIPT_MAX_CHARS);
-}
-
 function resolveDiscordRealtimeWakeNames(params: {
   config: DiscordRealtimeVoiceConfig;
   cfg: OpenClawConfig;