openclaw - 💡(How to fix) Fix Model fallback doesn't recover after cooldown — persists across sessions [1 comments, 2 participants]

Bug description

When minimax-portal/MiniMax-M2.7 triggers a format error and enters cooldown, OpenClaw correctly falls back to the next available model (volcengine-plan/ark-code-latest). However, after the cooldown expires, OpenClaw does NOT automatically recover the primary model. Instead:

1. The session continues using the fallback model indefinitely
2. New sessions (e.g., new Feishu DM threads) also use the fallback model instead of the configured primary
3. Only a manual model switch via /model or restarting the gateway restores the primary model

Steps to reproduce

1. Configure minimax-portal/MiniMax-M2.7 as the primary model with fallbacks: [minimax/MiniMax-M2.7, volcengine-plan/ark-code-latest, ...]
2. Use the primary model normally until a format error occurs (e.g., API returns an unexpected response format)
3. The model enters cooldown and OpenClaw falls back to volcengine-plan/ark-code-latest
4. Wait for the cooldown to expire (the cooldownUntil timestamp passes)
5. Observe that new messages still use volcengine-plan/ark-code-latest, not the primary model

Expected behavior

After a cooldown expires, OpenClaw should:

• Attempt to use the primary model again for new requests
• Or at minimum, attempt models in fallback order fresh, rather than sticky-falling-back to the last successful one

Diagnosis data

auth-state.json after the format error:

{
  "minimax-portal:default": {
    "errorCount": 1,
    "cooldownUntil": 1777728876071,
    "cooldownReason": "format",
    "lastFailureAt": 1777728846071
  }
}

After cooldown expires, lastGood still points to the working fallback, not the primary. The heuristic algorithm appears to prefer the most recently successful model over the configured primary.

System info

• OpenClaw 2026.4.26 (be8c246)
• macOS Darwin 24.6.0 (arm64)
• Node.js v25.9.0
• Model: minimax-portal/MiniMax-M2.7 (OAuth)
• Channel: Feishu DM

Suggested fix

Options (any of these would help):

1. Primary model always-first: After cooldown expires, always attempt the primary model first, not the last-used fallback
2. Session-level model reset: When a session is idle for >X minutes or on explicit model switch, re-evaluate the primary model instead of sticky-using the fallback
3. Configurable recovery policy: Add a config option like agents.defaults.model.recoveryPolicy: "primary_always" | "last_successful_preferred" so users can control fallback recovery behavior
4. Clearer diagnostics: Add a log line when cooldown expires and which model is being selected next, so users can understand why the model hasn't "recovered"

Workaround

Users must manually:

1. Edit ~/.openclaw/agents/main/agent/auth-state.json to reset errorCount: 0 and remove cooldownUntil
2. Restart the gateway
3. Run /model <primary-alias> to force-switch back

This is non-obvious and burdensome for end users.