openclaw - 💡(How to fix) Fix Feature request: configurable OVERLOAD_FAILOVER_BACKOFF_POLICY [1 comments, 2 participants]

Problem

The hardcoded OVERLOAD_FAILOVER_BACKOFF_POLICY uses an initial backoff of 250ms and a max of 1500ms. This is too aggressive for agents with large system prompts (e.g. ~40K tokens), where API overload events are more likely and token processing time is longer.

What happens in practice

When Anthropic returns an overload response during normal operation, OpenClaw retries with 250ms → 500ms → 1000ms → 1500ms backoffs. For a large-prompt agent, the retry itself can also hit overload (the system hasn't recovered yet), causing a cascading failure loop — a restart death spiral where the agent continuously crashes and restarts, never successfully making a request.

Current workaround

Manually patching the dist files:

# In dist/reply-*.js, dist/compact-*.js, dist/plugin-sdk/dispatch-*.js
const OVERLOAD_FAILOVER_BACKOFF_POLICY = {
  initialMs: 2000,   # was 250
  maxMs: 15000,      # was 1500
  factor: 2,
  jitter: .2
};

This workaround is fragile — it gets silently overwritten on every npm update openclaw, requiring manual re-patching after every update.

Feature Request

Expose overloadFailoverBackoffPolicy as a configurable option in agent config (e.g., clawdbot.json):

{
  "overloadFailoverBackoffPolicy": {
    "initialMs": 2000,
    "maxMs": 15000,
    "factor": 2,
    "jitter": 0.2
  }
}

Requirements

• Keep 250ms as the default for backward compatibility — this change should be opt-in
• Allow per-agent override so agents with large system prompts can use longer backoffs
• Alternatively, auto-scale the backoff based on estimated prompt token count

Impact

Agents with large system prompts (~40K+ tokens) are significantly more vulnerable to overload death spirals. The 250ms default was likely designed for lightweight agents and doesn't account for the variance in prompt complexity across different deployments.

Environment

• OpenClaw version: latest (npm)
• Affected agent sizes: ~40K token system prompts
• Workaround patch applied to: dist/reply-*.js, dist/compact-*.js, dist/plugin-sdk/dispatch-*.js