Neutralize accidental @everyone/@here in outbound agent text

Agent-authored replies pass through sanitizeForOutbound before Discord
send, but it did not touch mass-mention tokens — so a reply that merely
*mentioned* "@everyone" while explaining a feature pinged the whole
channel. Add neutralizeMassMentions (zero-width space after @) to the
central sanitizer so normal reply/progress/edit text can never mass-ping;
intentional broadcasts (e.g. the disk-usage alert) use a dedicated path.
Leaves <@id> mentions and emails intact. Covered by unit tests.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/router-mass-mentions.test.ts

@@ -0,0 +1,45 @@
+import { describe, expect, it } from 'vitest';
+
+import { neutralizeMassMentions, sanitizeForOutbound } from './router.js';
+
+const ZWSP = '\u200b';
+
+describe('neutralizeMassMentions', () => {
+  it('neutralizes @everyone / @here so they no longer ping', () => {
+    expect(neutralizeMassMentions('@everyone hi')).toBe(`@${ZWSP}everyone hi`);
+    expect(neutralizeMassMentions('ping @here now')).toBe(
+      `ping @${ZWSP}here now`,
+    );
+    // The rendered text still reads the same to a human...
+    expect(neutralizeMassMentions('@everyone').replace(ZWSP, '')).toBe(
+      '@everyone',
+    );
+    // ...but the literal mention token is broken.
+    expect(neutralizeMassMentions('@everyone')).not.toBe('@everyone');
+  });
+
+  it('leaves user/role mentions, emails, and plain text untouched', () => {
+    expect(neutralizeMassMentions('<@216851709744513024> hello')).toBe(
+      '<@216851709744513024> hello',
+    );
+    expect(neutralizeMassMentions('mail me@example.com')).toBe(
+      'mail me@example.com',
+    );
+    expect(neutralizeMassMentions('everyone here knows')).toBe(
+      'everyone here knows',
+    );
+  });
+
+  it('is idempotent (already-neutralized text is unchanged)', () => {
+    const once = neutralizeMassMentions('@everyone and @here');
+    expect(neutralizeMassMentions(once)).toBe(once);
+  });
+});
+
+describe('sanitizeForOutbound applies mass-mention neutralization', () => {
+  it('breaks an accidental @everyone in normal agent output', () => {
+    const out = sanitizeForOutbound('알림: @everyone 디스크 부족');
+    expect(out).toContain(`@${ZWSP}everyone`);
+    expect(out).not.toContain('@everyone');
+  });
+});

src/router.ts

@@ -146,10 +146,24 @@ export function neutralizeStrayMarkdown(text: string): string {
 /** @deprecated Kept for back-compat; use neutralizeStrayMarkdown. */
 export const neutralizeStrayBackticks = neutralizeStrayMarkdown;
 
+/**
+ * Prevent accidental mass pings. Agent-authored text that literally contains
+ * `@everyone` / `@here` must never notify the whole channel — that is only ever
+ * intended via a dedicated broadcast path (e.g. the disk-usage alert), not via a
+ * normal reply/progress/edit message. A zero-width space (U+200B) after the `@`
+ * keeps the text readable ("@everyone" still reads the same) while stopping
+ * Discord from parsing it as a mention. Idempotent, and leaves `<@id>` user/role
+ * mentions and ordinary emails (`me@example.com`) untouched.
+ */
+export function neutralizeMassMentions(text: string): string {
+  return text.replace(/@(everyone|here)\b/g, '@\u200b$1');
+}
+
 /**
  * Sanitize raw agent output for internal use (storage, IPC, intermediate
- * channel buffers). Strips internal tags + tool-call leaks and redacts
- * secrets, but does NOT touch markdown delimiters.
+ * channel buffers). Strips internal tags + tool-call leaks, redacts secrets,
+ * and neutralizes accidental @everyone/@here pings, but does NOT touch markdown
+ * delimiters.
  *
  * Use this when the text will pass through another `formatOutbound` call
  * downstream (e.g., the Discord channel boundary). Applying the markdown
@@ -161,7 +175,8 @@ export function sanitizeForOutbound(rawText: string): string {
   if (!text) return '';
   text = stripToolCallLeaks(text);
   if (!text) return '';
-  return redactSecrets(text);
+  text = redactSecrets(text);
+  return neutralizeMassMentions(text);
 }
 
 /**