Summary

Discord channels occasionally surface internal LLM tool-call artifacts that should never be visible to end-users. The leaked content includes:

• NO_REPLY
• to=functions.memory_search
• commentary
• Raw tool-call JSON arguments (e.g. {"query":"...","maxResults":5})
• recipient_name / parameters envelope fields

Other channels (Telegram, WhatsApp, Signal) do not exhibit this issue.

Expected Behavior

1. NO_REPLY should be consumed silently — never sent to any channel.
2. commentary, to=functions.*, tool-call JSON, and other internal envelope fields should be stripped before delivery.
3. Only natural-language content should reach end-users.

Actual Behavior

Discord intermittently delivers raw internal messages as visible chat messages in the channel.

Root Cause Analysis

Traced through the delivery pipeline in v2026.3.8:

1. Discord missing from PLAIN_TEXT_SURFACES (src/infra/outbound/sanitize-text.ts)

const PLAIN_TEXT_SURFACES = new Set([
  "whatsapp", "signal", "sms", "irc", "telegram", "imessage", "googlechat"
  // ← "discord" is absent
]);

normalizePayloadsForChannelDelivery() calls sanitizeForPlainText() only for surfaces in this set. Discord payloads skip sanitization entirely.

2. sanitizeTextContent() does not strip tool-routing traces (src/auto-reply/extract-text.ts)

function sanitizeTextContent(text) {
  return stripThinkingTagsFromText(stripDowngradedToolCallText(stripMinimaxToolCallXml(text)));
}

This strips thinking tags and Minimax XML, but does not catch:

• NO_REPLY
• to=functions.*
• commentary
• Bare JSON tool arguments

3. sendDiscordText() has no pre-send filter (src/infra/send/discord.ts)

The Discord send path serializes and posts the text chunk directly — no final sanitization gate.

4. NO_REPLY suppression has a gap (src/infra/outbound/normalize.ts)

In normalizeReplyPayloadsForDelivery, silent payloads are only suppressed when mergedMedia.length === 0. Payloads with media attached can pass through with NO_REPLY still in the text.

High-Probability Trigger Scenarios

• Inter-session / cross-agent message relay
• Subagent completion auto-forwarding
• Intermediate tool-call messages during multi-step reasoning
• NO_REPLY return branches

Suggested Fix

Add a universal outbound sanitizer in normalizePayloadsForChannelDelivery (or equivalent) that runs for all channels, not just plain-text surfaces:

const INTERNAL_TRACE_RE = /^(?:NO_REPLY|to=functions\.\S+.*|commentary|recipient_name|parameters)$/m;
const TOOL_JSON_RE = /^\s*\{\s*"(?:query|maxResults|tool_name|function_call|name)"\s*:/m;

function stripInternalTraces(text) {
  if (!text) return text;
  const trimmed = text.trim();
  if (trimmed === "NO_REPLY") return null;
  if (INTERNAL_TRACE_RE.test(trimmed)) return null;
  if (TOOL_JSON_RE.test(trimmed) && trimmed.length < 500) return null;
  return text
    .replace(/^NO_REPLY\s*/gm, "")
    .replace(/^to=functions\.\S+.*$/gm, "")
    .replace(/^commentary\s*$/gm, "")
    .replace(/^recipient_name\s*$/gm, "")
    .replace(/^parameters\s*$/gm, "")
    .replace(/\n{3,}/g, "\n\n")
    .trim() || null;
}

Also enhance sanitizeTextContent() to strip the same patterns at the extraction layer.

Environment

• OpenClaw: 2026.3.8 (3caab92)
• Platform: macOS (darwin)
• Channels affected: Discord
• Channels working correctly: Telegram, WhatsApp, Signal