Problem

When a sub-agent fails (bad config, missing dependency, logic error), there is no mechanism to prevent the parent from immediately spawning it again. This creates cascade failures where the same broken sub-agent is retried indefinitely, burning tokens and time.

OpenClaw tracks sub-agent completion status and has configurable runTimeoutSeconds, but there is no circuit breaker that says "this sub-agent has failed N times in a row — stop trying."

Evidence

• Cascade failures where a broken sub-agent was repeatedly spawned, each time failing the same way
• Token spend accumulates with each failed attempt (model boot cost, context injection, partial execution)

Workaround

I built cc-failure-guard.py, a PreToolUse hook (~80 lines Python) that:

1. Intercepts SubagentStart events before execution
2. Checks if a failure marker file exists in the target workspace
3. If present → blocks the spawn with feedback explaining the failure
4. The marker must be manually cleared before the sub-agent can be retried

Proposed Solution

Add a native circuit breaker to the sub-agent spawn system:

1. Track consecutive failures per sub-agent (by agent ID or task signature)
2. Configurable threshold: after N consecutive failures, block further spawns
3. Cooldown period: after threshold is hit, block for a configurable duration
4. User override: allow manual reset or forced spawn that bypasses the breaker
5. Notification: alert the user when a circuit breaker trips

{
  "agents": {
    "defaults": {
      "subagents": {
        "circuitBreaker": {
          "enabled": true,
          "maxConsecutiveFailures": 3,
          "cooldownMinutes": 30,
          "notifyOnTrip": true
        }
      }
    }
  }
}

Impact

Medium. Prevents token waste and user frustration from cascade failures. Especially valuable for automated/unattended pipelines.

Environment

• OpenClaw 2026.4.10 (npm, macOS)