openclaw - 💡(How to fix) Fix globalCircuitBreakerThreshold is per-tool-type, not session-global — allows cross-tool loop evasion

Problem

The globalCircuitBreakerThreshold setting is implemented as a per-tool-type counter, not a session-level global counter. This allows agents to evade the circuit breaker by switching tools, leading to unbounded loops.

Real-world incident

• Date: 2026-05-08
• Session: feishu DM, model qwen/qwen3.5-plus
• What happened: Agent entered a loop calling wecom_mcp → blocked after 5 calls → switched to exec → blocked after 5 calls → switched to memory_search → blocked after 5 calls → switched again... repeated across 4+ tools.
• Result: ~60+ LLM calls, ~750k input tokens consumed, user had to manually /stop
• Root cause: globalCircuitBreakerThreshold resets its counter for each tool type, so switching tools resets the count

Expected behavior

Per documentation: globalCircuitBreakerThreshold is described as "global no-progress breaker threshold". The word "global" implies it should count across all tools within a session, not per-tool-type.

Actual behavior

Each tool type has its own independent counter. Switching tools resets the count:

Proposed fix

Track no-progress counts at the session level, not per-tool-type:

{
  "tools": {
    "loopDetection": {
      "globalCircuitBreakerThreshold": 5  // total across ALL tools in session
    }
  }
}

When the session-level counter reaches the threshold, block all tool calls (not just the current tool) and force the agent to produce a text response.

Workaround

Lower the threshold to 2-3 to reduce the blast radius of each tool switch, but this does not solve the fundamental issue.

Environment

• OpenClaw 2026.5.6
• Ubuntu 24.04
• Model: qwen/qwen3.5-plus