openclaw - 💡(How to fix) Fix [Feature]: Gateway-level suppression of intermediate text output (pre-tool narration) [1 comments, 2 participants]

Problem

In multi-agent Slack deployments, language models frequently emit intermediate text before or between tool calls — "Let me check...", "I'll look into that...", process narration. This text is posted to the channel as a visible message, creating noise.

Why prompt engineering is insufficient

• System prompts say "don't narrate" but compete with training defaults (models are trained to explain their reasoning)
• When AGENTS.md says "don't narrate routine calls" and SOUL.md says "speak when needed", the model resolves ambiguity toward speech — the cost of silence-when-expected is perceived as higher than speech-when-unwanted
• blockStreamingDefault: on prevents inter-chunk streaming leaks but does NOT suppress deliberate text output
• thinkingDefault: low provides private reasoning space but the model still voluntarily emits text

This is fundamentally fighting the grain of a language model. Producing language is what it does. The fix should be structural, not behavioural.

Proposed solution

A gateway-level config option that suppresses text output that precedes tool calls:

{
  "agents": {
    "main": {
      "suppressIntermediateText": true
    }
  }
}

Behaviour:

• When suppressIntermediateText: true, any text block that is followed by a tool call is dropped (not posted to the channel)
• Only the final text after all tool calls complete is delivered
• If the entire response is text (no tool calls), it is delivered normally
• This is per-agent, allowing different behaviour for different bots

Alternative: per-channel output mode

{
  "channels": {
    "slack": {
      "outputMode": "result-only"
    }
  }
}

This would be channel-scoped rather than agent-scoped, useful for group channels where narration is never wanted but allowing it in DMs.

Impact

In our 5-agent deployment, intermediate narration accounts for ~30-40% of visible messages in group channels. Each narration message:

• Adds cognitive load for humans monitoring channels
• Can trigger other agents to respond (cascade risk)
• Wastes channel space with no informational value

Related

• #50984 — Thread session routing (separate but related multi-agent coordination issue)
• #5470 — replyToMode: off threading bug (closed, local patch applied)

Environment

• OpenClaw 2026.1.24-3
• Slack (Socket Mode)
• 5 agents across 2 servers
• blockStreamingDefault: on, thinkingDefault: low already configured