Problem

When the model streams NO_REPLY as multiple tokens (e.g. "NO" then "_REPLY"), the lifecycle:end event can fire while the buffer only contains the partial prefix. isSilentReplyText() returns false for "NO", causing it to leak to connected clients as a real assistant message.

This is particularly visible on node clients (VS Code extensions, mobile nodes) that display every chat-final message as a bubble — users see random "NO" or "NO_" messages appear.

Fix

Add a partial-prefix check in emitChatFinal: if the buffered text is a strict prefix of SILENT_REPLY_TOKEN, is shorter than the full token, and contains only [A-Z_] characters, treat it as a silent reply and suppress it.

The check is deliberately conservative:

• Minimum 2 characters (avoids matching single letters)
• Must be strictly shorter than the full token (full match already handled by isSilentReplyText)
• Only uppercase letters and underscores (won't match real messages)

Testing

Tested in production with a VS Code node client (Pawr) over multiple days. No more leaked partial tokens.

Fixes #3340

Greptile Summary

Adds logic to suppress partial NO_REPLY tokens that leak when the model streams the token across multiple chunks and lifecycle:end fires before the complete token arrives. The fix checks if buffered text is a strict prefix of SILENT_REPLY_TOKEN before emitting chat-final events.

• Prevents partial tokens like "NO" or "NO_" from appearing as assistant messages in client UIs
• Conservative approach with minimum 2-char length and character class validation
• Complements existing isSilentReplyText() which handles complete tokens

Confidence Score: 4/5

• Safe to merge - addresses real production bug with conservative logic
• The fix correctly addresses the partial token leak issue with appropriate boundary checks. Logic has been tested in production. Minor style improvement suggested on regex flag but doesn't affect correctness
• No files require special attention

_{Last reviewed commit: abe3c10}

_{(3/5) Reply to the agent's comments like "Can you suggest a fix for this @greptileai?" or ask follow-up questions!}

🤖 AI-assisted (Claude Opus/Sonnet via OpenClaw agent). Fully tested in production.