From e23624960e85a665745260696aba007797d5f7eb Mon Sep 17 00:00:00 2001
From: Lily Shen <115414357+lilyshen0722@users.noreply.github.com>
Date: Fri, 28 Aug 2026 18:57:12 -0700
Subject: [PATCH] fix(chat): normalize zero-width sentinel prefixes

---
 .../agentMessageService.chatNoise.test.js     | 23 +++++++++++++++++++
 backend/services/agentMessageService.ts       | 10 ++++++--
 2 files changed, 31 insertions(+), 2 deletions(-)

diff --git a/backend/__tests__/unit/services/agentMessageService.chatNoise.test.js b/backend/__tests__/unit/services/agentMessageService.chatNoise.test.js
index aac4c4363f..a3c918e190 100644
--- a/backend/__tests__/unit/services/agentMessageService.chatNoise.test.js
+++ b/backend/__tests__/unit/services/agentMessageService.chatNoise.test.js
@@ -139,6 +139,29 @@ describe('sanitizeAgentContent — NO_REPLY suppression and sanitization', () =>
     expect(AgentMessageService.sanitizeAgentContent('NO_REPLY')).toBe('');
   });
 
+  it('normalizes Unicode format-character prefixes before both sentinel checks', () => {
+    // U+200B defeats both the total-match and leading-run checks without the
+    // shared normalization below. Keep the class coupled too: a one-predicate
+    // fix would re-open the same private-reasoning leak.
+    [
+      '\u200B', '\u200C', '\u200D', '\uFEFF', '\u2060',
+      '\u061C', '\u200E', '\u2062', '\u00AD',
+    ].forEach((formatCharacter) => {
+      expect(AgentMessageService.sanitizeAgentContent(
+        `${formatCharacter}NO_REPLY\n\nprivate reasoning`,
+      )).toBe('');
+      expect(AgentMessageService.sanitizeAgentContent(`${formatCharacter}NO_REPLY`)).toBe('');
+    });
+  });
+
+  it('keeps zero-width joiners in substantive agent output', () => {
+    // U+200D is also in the sentinel-normalization class, but it is
+    // load-bearing in emoji. Sentinel decisions may normalize a copy only.
+    const family = '👨‍👩‍👧';
+    expect(AgentMessageService.sanitizeAgentContent(`${family} shipped the fix.`))
+      .toBe(`${family} shipped the fix.`);
+  });
+
   it('drops known bare runtime artifacts without swallowing terse replies', () => {
     expect(AgentMessageService.sanitizeAgentContent('RGCTX')).toBe('');
 
diff --git a/backend/services/agentMessageService.ts b/backend/services/agentMessageService.ts
index 4bd7d036e8..8ab44b4737 100644
--- a/backend/services/agentMessageService.ts
+++ b/backend/services/agentMessageService.ts
@@ -59,6 +59,11 @@ const ATTACH_CLAIM_SCAN_LIMIT = 2000;
 const BARE_RUNTIME_ARTIFACTS = new Set(['RGCTX']);
 
 const NO_REPLY_SENTINEL = 'NO_REPLY';
+// `trim()` does not remove Unicode format characters. Normalize a
+// decision-only copy before total-match and leading-sentinel checks so one
+// invisible prefix cannot bypass both suppression paths. Do not rewrite the
+// stored payload: U+200D, for example, joins emoji glyphs.
+const FORMAT_CHARACTERS = /\p{Cf}/gu;
 
 /**
  * Word-boundary test for the sentinel scan. Deliberately not `\w` via regex:
@@ -1835,12 +1840,13 @@ class AgentMessageService {
     const outerFence = raw.match(/^```[^\n]*\n([\s\S]*?)```\s*$/s);
     const stripped = outerFence ? outerFence[1] : raw;
     const trimmed = stripped.trim();
+    const sentinelContent = trimmed.replace(FORMAT_CHARACTERS, '');
 
     // Sentinels are total-match contracts: suppress only when the complete
     // reply consists of NO_REPLY tokens. Gateways have historically joined
     // silent blocks into "NO_REPLYNO_REPLY" (or separated duplicates with
     // whitespace), so retain that compatibility.
-    if (/^(?:NO_REPLY\s*)+$/.test(trimmed)) return '';
+    if (/^(?:NO_REPLY\s*)+$/.test(sentinelContent)) return '';
 
     // A substantive fully fenced reply is explicitly code-formatted even
     // though this sanitizer removes the outer transport fence before storage.
@@ -1863,7 +1869,7 @@ class AgentMessageService {
     // reads as silent there too. That is the intended reading of those rows —
     // and it does not apply to the AX-43 leaks themselves, which were stored
     // with the token already stripped.
-    if (opensWithBareSentinel(trimmed)) {
+    if (opensWithBareSentinel(sentinelContent)) {
       if (observe) {
         console.warn(
           `[agent-msg] suppressed a reply opening with a bare sentinel (intended silence, not an edit) from agent=${observe.agentName} instance=${observe.instanceId} pod=${observe.podId}: ${trimmed.slice(0, 120)}`,