Summary

當主要模型回傳 overloaded 錯誤時（特別是 Anthropic 的 HTTP 529、Gemini/OpenAI 的「The AI service is temporarily overloaded. Please try again in a moment.」、503/502/504、api_error 等），目前 OpenClaw 不會正確觸發

Problem to solve

Related Issues

• [Bug]: Overloaded error does not trigger model fallback — session hangs indefinitely with no retry or UI escape #24378 [Bug] Overloaded error does not trigger model fallback
• Bug: HTTP 529 (Anthropic overloaded) does not trigger model fallback #28502 [Bug] HTTP 529 (Anthropic overloaded) does not trigger model fallback
• Treat HTTP 503 (and 502/504) as failover-eligible so model fallback triggers #20999 Treat HTTP 503 as failover-eligible
• Anthropic 'Internal server error' (api_error) doesn't trigger model fallback #23193 Anthropic 'Internal server error' doesn't trigger fallback
• Feature: Trigger model fallback on context length exceeded #9986 Context length exceeded should also fallback

Proposed solution

Related Issues

• [Bug]: Overloaded error does not trigger model fallback — session hangs indefinitely with no retry or UI escape #24378 [Bug] Overloaded error does not trigger model fallback
• Bug: HTTP 529 (Anthropic overloaded) does not trigger model fallback #28502 [Bug] HTTP 529 (Anthropic overloaded) does not trigger model fallback
• Treat HTTP 503 (and 502/504) as failover-eligible so model fallback triggers #20999 Treat HTTP 503 as failover-eligible
• Anthropic 'Internal server error' (api_error) doesn't trigger model fallback #23193 Anthropic 'Internal server error' doesn't trigger fallback
• Feature: Trigger model fallback on context length exceeded #9986 Context length exceeded should also fallback

Alternatives considered

No response

Impact

1. 把 primary model 設成 anthropic/claude-3-5-sonnet-20241022（或任何容易過載的模型）
2. 在vice is tempora設定好 fallbacks：

   agents:
     defaults:
       model:
         primary: anthropic/claude-3-5-sonnet-20241022
         fallbacks:
           - gemini/gemini-2.0-flash
           - moonshot/kimi-k2.5
           - openai/gpt-4o-mini

正常對話一段時間後，遇到 overloaded
結果：session 永久 hang 住，fallback 完全不執行，連 retry 都沒有
Expected Behavior（我建議的理想行為）
自動偵測所有常見 overloaded 錯誤（HTTP status + error message 雙重判斷）
立即觸發 fallback（依序嘗試下一個模型）
自動 retry 機制（同一個模型先 exponential backoff retry 3 次 → 再切 fallback）
UI 逃生門：在聊天視窗顯示臨時提示「AI 暫時過載，正在切換模型…」並支援 /retry 或 /model next 手動強制切換
可設定 fallback 順序、健康檢查間隔、單一模型最大 retry 次數

Evidence/examples

No response

Additional information

No response