Bug

Version: 2026.3.1
Component: Reply pipeline / streaming partial suppression

Description

When an agent reply starts with NO_REPLY (the silent reply token), the partial text NO leaks to webchat as a visible streaming delta before the full suppression logic catches it.

Root Cause

handlePartialForTyping in reply-XaR8IPbY.js calls isSilentReplyPrefixText on each streaming delta. The function has an early-exit guard:

if (!normalized.includes('_')) return false;

This means "NO" is not recognized as a prefix of NO_REPLY (because it has no underscore yet), so it passes through and gets pushed to webchat as a typing signal. By the time "NO_" arrives and would be caught, "NO" has already been rendered.

Reproduction

1. Configure an agent that replies with only NO_REPLY (the silent token)
2. Observe webchat — NO briefly appears as a streamed partial before being suppressed
3. Final message is correctly suppressed; only the partial leaks

Expected Behavior

All prefixes of NO_REPLY that consist solely of uppercase letters (i.e., N, NO) should also be suppressed in handlePartialForTyping, even before the underscore arrives.

Suggested Fix

In isSilentReplyPrefixText, relax the includes('_') guard to also match pure-alpha prefixes that are a prefix of the token up to (but not including) the first _:

function isSilentReplyPrefixText(text, token = SILENT_REPLY_TOKEN) {
  if (!text) return false;
  const normalized = text.trimStart().toUpperCase();
  if (!normalized) return false;
  if (/[^A-Z_]/.test(normalized)) return false;
  return token.toUpperCase().startsWith(normalized);
}

Simply removing the includes('_') guard fixes the leak while keeping all existing behavior intact.

Impact

• Webchat users see a spurious NO message when an agent uses the silent reply token
• Affects any streaming webchat session on 2026.3.1