When using a custom OpenAI-compatible provider behind litellm (Open WebUI 0.8.12 with litellm as the backend), context overflow errors are not detected by isContextOverflow(), so overflow-triggered auto-compaction never fires.

Error message:

Error: 503 litellm.ServiceUnavailableError: litellm.MidStreamFallbackError: litellm.APIConnectionError: APIConnectionError: OpenAIException - Requested token count exceeds the model's maximum context length of 131072 tokens.

What's happening:

The error message contains "exceeds the model's maximum context length" but the existing patterns in @earendil-works/pi-ai/utils/overflow.ts appear to expect different phrasings — e.g. "exceeds the context window" or "maximum context length is … tokens". This particular wording from litellm/OpenAI ("exceeds the model's maximum context length of … tokens") doesn't match any of them. Additionally, the error is wrapped in a litellm ServiceUnavailableError / MidStreamFallbackError / APIConnectionError chain, prefixed with HTTP 503, which could also interfere with pattern matching if any patterns anchor to specific status codes or error prefixes.

Provider setup:

• Open WebUI 0.8.12 (litellm backend), api: "openai-completions"
• Model: glm-5.1-llmlb, contextWindow: 131072, maxTokens: 16384, reasoning: true
• pi 0.74.0

pi config:

{
  "compaction": {
    "enabled": true,
    "reserveTokens": 32000,
    "keepRecentTokens": 12000
  }
}

Model entry in models.json:

{
  "id": "glm-5.1-llmlb",
  "contextWindow": 131072,
  "maxTokens": 16384,
  "reasoning": true
}