Consider extracting the GPT-5 model detection logic into a shared helper function. This logic (using modelId.toLowerCase().includes('gpt-5') || modelId.toLowerCase().includes('gpt5')) appears in both createMessage and completePrompt, and centralizing it would improve maintainability.

^{This comment was generated because it violated a code review rule: irule_tTqpIuNs8DV0QFGj.}

The model detection logic could be more precise. Currently, modelId.toLowerCase().includes("gpt-5") would match unintended models like "not-gpt-5000". Consider using a more specific pattern:

This detection logic is duplicated from line 111. Would it be cleaner to extract this into a helper method to maintain DRY principles? Something like:

private isGPT5Model(modelId: string): boolean {
  const modelLower = modelId.toLowerCase()
  return modelLower.startsWith("gpt-5") || modelLower.startsWith("gpt5") || modelLower === "gpt5"
}

Is there a way to avoid using @ts-ignore here? Could we extend the OpenAI types or create a custom interface that includes max_completion_tokens to maintain type safety? For example:

interface GPT5RequestOptions extends Omit<OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming, 'max_tokens'> {
  max_completion_tokens?: number
  max_tokens?: never
}

Great test coverage! Consider adding edge cases like mixed case variations ("GpT-5", "gPt5") or models with additional suffixes ("gpt-5-32k", "gpt-5-vision") to ensure the detection works correctly for all possible GPT-5 model names.

Consider adding a comment explaining why GPT-5 models require this special handling, perhaps with a link to relevant Azure/OpenAI documentation. This would help future maintainers (including future me) understand the context behind this workaround.