[Bug]: NO_REPLY leaks into webchat responses (silent token not filtered) #3340
Description
Summary:
In webchat, the internal silent token NO_REPLY is appearing as a visible assistant message. This should be suppressed (Clawdbot docs state it must be filtered from outgoing payloads).
Environment:
- Clawdbot version: 2026.1.24-3 (from config)
- Channel: webchat
- Model: deepseek-chat
Expected:
NO_REPLY should be treated as a silent token and never rendered in webchat output.
Actual:
User sees NO_REPLY as a visible assistant message in webchat.
Notes / context:
NO_REPLYshows up in session transcripts (expected), but it’s being delivered to the webchat UI.- Docs say silent replies are filtered from outgoing payloads.
Attempted local patch (did NOT fix):
I patched the gateway chat streamer to drop NO_REPLY before sending chat deltas/final:
File: dist/gateway/server-chat.js
import { isSilentReplyText, SILENT_REPLY_TOKEN } from "../auto-reply/tokens.js";
const emitChatDelta = (...) => {
...
if (isSilentReplyText(text, SILENT_REPLY_TOKEN)) return;
...
};
const emitChatFinal = (...) => {
const text = ...
if (isSilentReplyText(text, SILENT_REPLY_TOKEN)) {
broadcast("chat", { state: "final", message: undefined, ... });
return;
}
...
};
After restart, NO_REPLY still appeared in webchat, so the leak may occur earlier (e.g., message normalization or delivery mirror), or via a different stream.
Logs/examples:
- I can provide a session snippet where the assistant message is exactly "NO_REPLY".
Request:
Please ensure NO_REPLY is stripped from webchat outputs consistently (both streaming and final). Ideally apply the same silent-token filtering used elsewhere in the auto-reply pipeline.