Summary

OpenClaw needs a native "listen-only" or "silent ingest" mode for group chats where the agent should see all messages (for context/indexing) but only invoke the LLM when explicitly mentioned or asked a direct question.

Currently, the only way to achieve this is requireMention: false + a systemPrompt instructing the agent to reply NO_REPLY for non-relevant messages. This wastes tokens on every single message just to produce a "stay silent" response.

Problem

With requireMention: false, every message in the group triggers a full LLM call. For a busy group with 50+ messages/day, this means:

• Significant token waste — each "NO_REPLY" decision costs input tokens (system prompt + context) + output tokens
• Unnecessary API calls — hitting rate limits faster
• Higher latency for actual requests — queue congestion from silent processing

Proposed Solution

Add native config-level support for selective response:

{
  "channels": {
    "telegram": {
      "groups": {
        "-123456789": {
          "listenMode": "all",
          "respondMode": "mention",
          "silentIngest": {
            "enabled": true,
            "hooks": ["session-memory"]
          }
        }
      }
    }
  }
}

Key concepts:

1. listenMode: "all" — Agent receives all messages into session context (current requireMention: false behavior)

2. respondMode: "mention" | "all" | "allowFrom" — Controls when the LLM is actually invoked:

   • "mention" — only when @mentioned or replied to
   • "all" — every message (current default when requireMention: false)
   • "allowFrom" — only for messages from specific sender IDs

3. silentIngest — Messages that don't trigger LLM invocation are still:

   • Added to session context (so agent has full history when invoked)
   • Processed by configured hooks (e.g., session-memory for indexing)
   • NOT sent to the LLM API

Additional enhancement: Per-sender response permissions

Currently allowFrom is binary — a sender either triggers the agent or is ignored entirely. A more granular model would allow:

{
  "allowFrom": {
    "293894843": { "canCommand": true, "canTriggerResponse": true },
    "355322273": { "canCommand": true, "canTriggerResponse": true },
    "999999999": { "canTriggerResponse": false }
  }
}

This enables scenarios like "team members can command the bot, but client messages are only observed."

Use Cases

1. Client observation — Bot in client chat, listens to everything, only responds when team @mentions it
2. Internal work chat — Bot indexes all discussions into knowledge graph, responds only when asked
3. Meeting notes — Bot captures all messages but doesn't interrupt conversation flow
4. Cost optimization — Reduce API calls by 80-90% in busy groups

Current Workaround

{
  "requireMention": false,
  "systemPrompt": "You receive ALL messages. Only respond when @mentioned or asked directly. For ALL other messages: reply with exactly NO_REPLY."
}

This works but wastes tokens on every message and relies on prompt compliance (soft enforcement).

Environment

• OpenClaw 2026.2.9 (stable)
• Telegram groups with mixed listen/respond needs

Status update (2026-02-17)

Implemented in #15841 (open PR).

The delivered design is more generalized than the initial request:

• shared silent-ingest pipeline
• shared group ingest policy resolver
• explicit message_ingest hook contract
• Telegram + Signal config/schema support

This aligns with current OpenClaw architecture (centralized policy resolution + thin channel adapters + typed hook contracts).