Git provider

Github Cloud

System Info

model used: o4-mini with no credit (for raising openai.RateLimitError in completion)
deployment type: cli

Bug details

Description

relates #1802

• Current retry decorator configuration in AIHandler classes causes RateLimitError to be incorrectly retried
• The retry condition retry_if_exception_type(openai.APIError) catches RateLimitError since it inherits from APIError
• This breaks the intended behavior where rate limit errors should not be retried

# pr_agent/algo/ai_handlers/litellm_ai_handler.py

@retry(
    retry=retry_if_exception_type((openai.APIError, openai.APIConnectionError, openai.APITimeoutError)), # No retry on RateLimitError
    stop=stop_after_attempt(OPENAI_RETRIES)
)
async def chat_completion(self, ...):
    # ...

Current Behavior (Problematic)

• used o4-mini API with no credit, for raising openai.RateLimitError in completion
  • 
• RateLimitError gets retried 5 times despite being logged at error level
• The @retry decorator's openai.APIError condition absorbs all scenarios including rate limits

Issue Details

• The retry decorator comment explicitly states # No retry on RateLimitError
• However, since RateLimitError inherits from APIError:

  # openai/_exceptions.py
   
  class APIError(OpenAIError):
  ...
  class APIStatusError(APIError):
  ...
  class RateLimitError(APIStatusError):
      status_code: Literal[429] = 429
  ...

• The retry condition retry_if_exception_type(openai.APIError) catches RateLimitError instances
• This causes the retry mechanism to incorrectly retry rate limit errors up to OPENAI_RETRIES times

Expected Behavior

• RateLimitError should be raised immediately without retry attempts
• Only non-rate-limit APIError subtypes should trigger retries