Failure Loop

Catches the same tool or command failing repeatedly in a short session window. After enough failures, Relay warns or blocks the next identical attempt.

Part of Behavior Guards. Thresholds are configurable in Console.

What it catches

• Repeated failed shell, file, or MCP calls with the same intent
• Clears when a matching attempt succeeds, the window expires, or you Acknowledge on Guards
• Not the same as Retry Thrash (which still counts successful identical polls)

Example

An agent debugs a flaky CI step and keeps running the same npm test or pytest call. Every run fails. Without Guards, the session burns tokens. With Guards, Relay warns, then Soft asks or Enforce blocks, and an incident nests under Failure Loop on Guards.

Try it

Relay → Guards → Failure Loop — enable, Configure thresholds, run Sample, then Acknowledge or succeed once with the same command.