Problem

When using sessions_spawn, the sub-agent announce step runs after completion and posts a summary to the requester chat. The only way to suppress this is for the sub-agent model to reply exactly ANNOUNCE_SKIP during the announce step.

The problem: The sub-agent model frequently ignores the ANNOUNCE_SKIP instruction in the task, even when explicitly stated. This means unverified, potentially fabricated summaries get posted directly to the chat before the main agent can review them.

Current workaround

Include When prompted to announce, reply exactly: ANNOUNCE_SKIP in the task. This works ~60% of the time — models sometimes follow it, sometimes don't.

Requested feature

A config option to suppress announce at the framework level:

{
  agents: {
    defaults: {
      subagents: {
        announcePolicy: "skip"  // "skip" | "always" | "model" (default)
      }
    }
  }
}

Or a per-spawn parameter:

sessions_spawn({ task: "...", announce: false })

Why this matters

• Sub-agent output is unverified — fabricated data, file paths, false claims
• The main agent should always be the gatekeeper between sub-agent and user
• Prompt-based enforcement has inherent unreliability
• Code/config enforcement is the only way to guarantee behavior

Environment

• OpenClaw version: 2026.1.24-1
• Model: Claude Opus 4.5 (main) / Sonnet 4 (sub-agents)
• Channel: Telegram