openclaw - 💡(How to fix) Fix Bug: No automatic retry with fallback when primary model returns overloaded_error

Bug: No Automatic Retry with Fallback on Provider Overload

OpenClaw version: 2026.5.19-beta.1 (ba9034b) Affected: All embedded agents on minimax provider Channel: Telegram (agent:neon:telegram:direct:2101884310) Date: 2026-05-20

Expected Behavior

1. Agent calls primary model (MiniMax-M2.7)
2. Primary returns overloaded_error
3. OpenClaw retries with fallback model (MiniMax-M2.5)
4. If fallback succeeds → deliver reply, session continues normally
5. If fallback also fails → surface error to user, reset session gracefully

Result: User gets a reply, session stays alive.

Actual Behavior

1. Agent calls primary model (MiniMax-M2.7)
2. Primary returns overloaded_error immediately
3. No retry with fallback — system waits for primary to timeout
4. 120-second timeout elapses
5. Session is killed by takeover error (Issue #84542)
6. User sees "Something went wrong. Please try again or use /new."

Result: User sees an error, session dies, manual recovery needed.

Timeline (CEST = UTC+2)

Key Logs

Primary overload — no retry:

{"event":"embedded_run_agent_end","isError":true,"failoverReason":"overloaded","model":"MiniMax-M2.7","providerRuntimeFailureKind":"timeout","providerErrorType":"overloaded_error"}

Note: failoverReason="overloaded" is set but no actual fallback retry was attempted.

Timeout (120s after overload):

{"subsystem":"fetch-timeout","elapsedMs":120002}
{"subsystem":"provider-transport-fetch","name":"TimeoutError","elapsedMs":120421}

Configuration

Fallback is configured but not used as retry:

{
  "models": {
    "providers": {
      "minimax": {
        "models": [
          { "id": "MiniMax-M2.7", "contextWindow": 200000 },
          { "id": "MiniMax-M2.5", "contextWindow": 200000 }
        ]
      }
    }
  }
}

Fallback order: minimax/MiniMax-M2.7 → minimax/MiniMax-M2.5

The fallback model is configured and available. When primary returns overloaded_error, retry with fallback should happen automatically.

Root Cause

overloaded_error is a transient server-side condition. OpenClaw logs failoverReason="overloaded" but does not attempt to retry with the fallback model. The error handling treats the overload as terminal rather than transient.

Severity

High — Every MiniMax overload incident results in session death. With automatic retry, these would recover transparently and users would not notice.

Labels to apply: bug, type/provider, severity/high