Incident: A timeout retried 1,384 times in a minute

Module 5 · Incident file · About 35 minutes

What happened

On 2 May 2026 a user, ThePuma312, opened issue #76293 on OpenClaw, an open-source agent runtime: "LLM idle timeout triggers unbounded retry loop, causing massive cost spikes" (version 2026.4.10). A model call had hit the error "LLM idle timeout (60s): no response from model", on Anthropic's claude-sonnet-4-6; the reporter saw the same with a local Ollama model. In their words: "Single timeout triggers a tight retry loop that fires until something external stops it."

They counted the damage from their own session logs. On 25 April 2026 the loop "generated 761 API call entries within 60 seconds". On 29 April it "generated 1,384 API call entries within 60 seconds", all stamped with the same minute. Their estimate, not billing data, was $20 to $30 in those 60 seconds. They named what was missing: "no maximum retry count, no exponential backoff, no circuit breaker". And they warned that a user on automatic top-ups "discovers it days later when reviewing usage."

The fix came the next day. Pull request #76345 by adhirajhangal, "fix(agents): circuit-break consecutive LLM idle timeouts to cap retry-loop cost", was merged by steipete on 3 May 2026 and closed the issue. Its description says: "Default cap is 5."