User description

Description

This PR implements 3 changes to properly exclude RateLimitError from retry logic in all AIHandler.chat_completion()

Changes

1. Applied consistent @retry library:
   • All AIHandler classes now use from tenacity import retry instead of the built-in import retry library
     • OpenAIHandler at pr_agent/algo/ai_handlers/openai_ai_handler.py
     • LangChainOpenAIHandler at pr_agent/algo/ai_handlers/langchain_ai_handler.py
2. Excluded RateLimitError from @retry conditions:
   • Previous APIError condition was absorbing RateLimitError instances, preventing retry breakpoint

   # AS-IS (Problematic)
   @retry(
       retry=retry_if_exception_type((openai.APIError, openai.APIConnectionError, openai.APITimeoutError)), # No retry on RateLimitError
       stop=stop_after_attempt(OPENAI_RETRIES)
   )
   def foo():
     raise openai.RateLimitError  # Oops... This gets retried because RateLimitError inherits from APIError :-(

   • So, I added retry_if_not_exception_type to include APIError while explicitly excluding RateLimitError

   # TO-BE
   @retry(
       retry=retry_if_exception_type(openai.APIError) & retry_if_not_exception_type(openai.RateLimitError),
       stop=stop_after_attempt(OPENAI_RETRIES),
   )

3. Applied consistent exception handling pattern:
   • Unified RateLimitError(error) -> APIError(warn) -> Exception(warn) pattern consistently
     • Applied pattern in fix: reorder exception handling in LiteLLMAIHandler.chat_completion() #1803
   • Removed redundant APITimeoutError handling, since it inherits from APIError

closes #1807

Test Results

AS-IS (Problematic)

• 
• RateLimitError incorrectly retried 5 times despite error-level logging

TO-BE (Fixed)

1. RateLimitError scenario

   • 
   • RateLimitError now exits immediately after single attempt as intended

2. APIError (non-RateLimit) scenario

   • 
   • Other APIError types correctly retry 5 times as expected

3. Normal execution

   • 
   • Standard operation works as expected

PR Type

Bug fix, Enhancement

Description

• Excludes RateLimitError from retry logic in all AIHandler chat_completion methods

  • Uses retry_if_not_exception_type(openai.RateLimitError) with tenacity for precise control

• Standardizes retry and exception handling across OpenAI, LiteLLM, and LangChain handlers

  • Ensures consistent logging and error escalation patterns

• Removes redundant or overly broad exception handling

  • Cleans up catch blocks and error messages for clarity

Changes walkthrough 📝

Need help?

Type /help how to ... in the comments thread for any questions about Qodo Merge usage.

Check out the documentation for more information.