openclaw - 💡(How to fix) Fix Suppress intermediate text output before tool calls in group channels [1 comments, 2 participants]

Problem

In multi-agent group channels, LLMs frequently emit "thinking out loud" text before tool calls — messages like "Let me check...", "Looking into this...", "I should set up...". This text is the model's deliberate output (not a streaming artifact), so blockStreamingDefault cannot suppress it.

Each intermediate text block gets posted as a separate channel message, creating noise and burning tokens. In a 5-bot setup, a single question can generate 10+ process-narration messages before any actual result appears.

What we've tried

1. blockStreamingDefault: "on" — only buffers streamed chunks, not complete text outputs preceding tool calls ❌
2. thinkingDefault: "low" — provides private reasoning space, but the model still voluntarily narrates ❌
3. AGENTS.md instructions ("no intermediate text") — competes with model training to be helpful/explanatory, doesn't stick reliably ❌

Proposed solution

Add a config option to suppress text that precedes tool calls:

{
  agents: {
    defaults: {
      suppressIntermediateText: true  // or per-channel
    }
  }
}

Behavior

When suppressIntermediateText: true:

• If the model's response contains text followed by tool calls, the pre-tool text is dropped (not sent to the channel)
• Only the final text after all tool calls complete is delivered as the visible response
• If the model's response is text-only (no tool calls), it's delivered normally

This matches how humans work: form the thought → do the work → deliver the result. The intermediate narration is process, not product.

Alternative: per-channel config

{
  channels: {
    slack: {
      channels: {
        "C123": {
          suppressIntermediateText: true  // group channels
        }
      }
    }
  }
}

Environment

• Clawdbot version: 2026.1.24-3+
• Channel: Slack (multi-agent group channels)
• Models: Claude (Anthropic), Gemini fallback