openclaw - 💡(How to fix) Fix feat(gateway): allow per-agent replyToCurrent=false to disable auto-attaching assistant text as reply

Motivation

We hit two recurring pain points with the current "assistant text → auto-reply to inbound" behavior:

1. Dispatcher / cron-triggered agents: a cron systemEvent fires DARGO, DARGO calls message thread-reply to dispatch to PM, and the gateway then attaches DARGO's own "dispatched" assistant text as a reply to the cron message → produces orphan threads in Feishu.
2. Strict "no narration" agents: we enforce internally (red line R12) that DARGO must only produce channel output via the message tool. Currently this is policed by prompt + skill + tripwire scripts, and recurrence is still ~5–6 violations per week.

Proposal

Add a per-account opt-out, e.g.:

{
  "channels": {
    "feishu": {
      "accounts": {
        "dargo": {
          "replyToCurrent": false  // default true; opt-in opt-out
        }
      }
    }
  }
}

When false, plain assistant text is not auto-attached to the inbound message as a reply. Only explicit message tool calls produce channel output. Plain text still appears in the session transcript for debugging.

Equivalent agent-side knob also welcome:

{
  "outputContract": ["message-tool-only"]
}

Use cases

• DARGO subagent / cron flow: pure execution, no need to reply to inbound
• Agents that must obey strict "tool-only output" contracts (audit, dispatcher, watchdog roles)

Alternative

• Status quo: instruct agent via prompt + skill to "only use the message tool". Recurrence high (we logged 5+ R-D515 violations in May 2026 alone).

Risks

• Would break casual conversational agents → must be opt-in per agent, default true
• Need to clarify behavior for tool_use blocks that contain accompanying text (likely: same rule — drop text unless message tool is called)

Origin

Draft from workspace-dargo/state/openclaw-fr.md. Filed by the DARGO system architect agent on behalf of @90yuwang.