Problem

When Anthropic returns overloaded_error (HTTP 529), OpenClaw surfaces it as a terminal error ("Subagent claude timed out" / session error) instead of triggering the configured model fallback chain.

Expected behavior

Per docs/concepts/model-failover.md, auth/rate-limit/timeout errors should trigger cooldown on the current auth profile, then rotation to fallback models. The overloaded_error should be classified as a failover-worthy transient error.

Observed behavior

• Multiple sessions simultaneously hit overloaded_error from Anthropic
• Auth profile errorCount stays at 0 (cooldown not triggered)
• Fallback model (openai-codex/gpt-5.3-codex) never attempted
• Sub-agent sessions surface "timed out" to the parent
• Cron jobs show lastRunStatus: error with no retry

Config

{
  "model": {
    "primary": "anthropic/claude-opus-4-6",
    "fallbacks": ["openai-codex/gpt-5.3-codex"]
  }
}

Error payload from Anthropic

{
  "type": "error",
  "error": {
    "details": null,
    "type": "overloaded_error",
    "message": "Overloaded"
  }
}

Suggested fix

Classify overloaded_error (and HTTP 529) as a failover-worthy transient error alongside existing rate-limit/timeout handling. Apply the same cooldown + profile rotation + model fallback chain.

Environment

• OpenClaw 2026.3.2 (85377a2)
• macOS 26.3 (arm64)
• Node 25.6.1