openclaw - 💡(How to fix) Fix Heartbeat + auth fallback + message tool creates cascading delivery-mirror feedback loop [1 comments, 2 participants]

Summary

When an agent's heartbeat fires while the primary model provider has invalidated auth tokens (all profiles in cooldown), the fallback chain eventually reaches an alternative provider. That provider's LLM sees session context about "empty messages" and uses the message tool to send replies directly to the user's messaging channel. This creates a self-reinforcing feedback loop where each heartbeat cycle generates another user-facing message, which then becomes context for the next heartbeat, ad infinitum.

Steps to Reproduce

1. Configure an agent with:

   • heartbeat.every: "30m"
   • heartbeat.model: "openai-codex/gpt-5.5" (or any provider that can enter auth cooldown)
   • tools.profile: "full" (which includes the message tool)
   • WhatsApp channel connected

2. Cause the primary model's auth tokens to become invalidated (e.g., token expiry, provider outage)

3. Wait for heartbeat cycles to fire

What Happens

heartbeat fires (every 30m)
  → prompt injected into agent main session
  → primary model fails (auth error)
  → all auth profiles enter cooldown
  → fallback chain exhausted: gpt-5.5 → gpt-5.4-mini → gpt-5.4-nano → all fail
  → falls back to alternative provider (e.g., oc/kimi-k2.6)
  → alternative LLM sees session context about "empty messages"
  → LLM uses `message` tool to send "I'm receiving empty messages..." to user's WhatsApp
  → user receives unexpected message
  → next heartbeat: context now includes the auto-generated message
  → alternative LLM sends ANOTHER message ("I'll stop responding now")
  → loop continues indefinitely

Evidence

Real-world cascade observed over ~7 hours on 2026-04-28, all via delivery-mirror entries correlating with web-outbound WhatsApp deliveries:

Log evidence for the 02:34 cascade:

23:34:14 — lane task error: lane=session:agent:harry-vip:main
           error="FailoverError: No available auth profile for openai-codex
           (all in cooldown or unavailable)."

23:34:49 — message tool result:
           channel=whatsapp, to=+554799810540
           messageId=3EB058C2A090A42C5BEEE7

Session transcript shows empty user messages ("text": "") being injected into the main session at each heartbeat trigger, which the fallback LLM interprets as "the user is sending empty audio."

Proposed Fixes

1. Heartbeat isolation: Heartbeat-triggered runs should not have access to the message tool (or any outbound delivery tool) unless explicitly configured. The heartbeat's output should be introspection-only by default.

2. Empty-message loop detection: The delivery-mirror or auto-reply system should detect when it's generating repeated "empty message" fallback texts and apply a cutoff after N consecutive such deliveries.

3. Auth-agnostic heartbeat: Consider allowing heartbeats to use a different (more reliable) model than the main conversation model, so auth failures on the primary provider don't block heartbeat processing.

Environment

• OpenClaw version: 2026.4.25-da6bdffc3d96
• Node: 25.9.0
• Platform: Darwin arm64 (macOS)