Summary

In workflow-style sessions, user-visible replies sometimes include internal tool-call scaffolding instead of only the clean final result.

Example

User sees leaked content like:

<function_calls>
<invoke name="exec">
<parameter name="command">cd /Users/mikeroberts/.openclaw/workspace && node scripts/knowledge-store.mjs search --query "what skills matter most in the age of AI" --limit 4</parameter>
<parameter name="yieldMs">10000</parameter>
</invoke>
</function_calls>
<function_response>
Searching for: "what skills matter most in the age of AI"
...
</function_response>

Important note

The underlying workflow is often working correctly. In this case, the semantic search result quality was good. The problem is that the user-facing reply is polluted with raw internal tool envelopes and command text.

Expected

Only the final human-readable answer should be shown to the user.

Actual

Internal tool-call scaffolding is included in the user-visible response:

• <function_calls>
• <invoke ...>
• <function_response>
• raw command text
• intermediate internal execution output

Impact

• confusing and noisy UX
• exposes internal implementation details
• makes workflows feel broken even when results are correct
• damages trust in agent responses

Context

Observed in Telegram workflow/agent sessions such as Second Brain, but likely not limited to that workflow.