The original title is: "Configuring OpenClaw's Model Fallback Chain: Lessons from a 3 AM Agent Outage"

Short summary

A developer shares a real incident where their OpenClaw agent stalled for 33 minutes after a primary model 503, and explains how to configure per-call model fallback chains to prevent silent failures. Key settings include maxRetries, retryDelayMs, and timeoutMs — aggressive values prevent long outages during overnight runs. The author also recommends logging fallback events and setting up alerts so you know when your agent is running degraded.

• •OpenClaw's fallback chain retries per-call, not per-session, preserving context across model switches
• •Misconfigured retry delays (5s × 5 retries) caused a 33-minute stall; aggressive settings (1s × 2 retries) fixed it
• •Log and alert on fallback events to detect unstable primary models early

Generated with AI, which can make mistakes.