Fix #377: Resolve OpenAI client streaming/non-streaming parser mixing

[pxkundu]

Fix #377: Resolve OpenAI client streaming/non-streaming parser mixing

Problem

The OpenAI client gets "stuck" in streaming or non-streaming mode after switching between stream=True and stream=False calls. This happens because the self.response_parser instance variable is modified during each call and persists across subsequent calls, causing the wrong parser to be used.

Root Cause

The issue occurs in these lines:

• self.response_parser = self.streaming_response_parser_sync (for streaming)
• self.response_parser = self.non_streaming_response_parser (for non-streaming)

The instance variable assignment means:

1. Call 1: generator(stream=True) → Sets self.response_parser = streaming_parser
2. Call 2: generator(stream=False) → Sets self.response_parser = non_streaming_parser
3. Call 3: generator(stream=True) → But parse_chat_completion() still uses the old self.response_parser

Solution

• ✅ Replace instance variable assignment with dynamic parser selection
• ✅ Determine correct parser based on completion type in parse_chat_completion()
• ✅ Add type-specific logic: Response → non-streaming, AsyncIterable → async streaming, Iterable → sync streaming
• ✅ Exclude basic types (str, bytes, dict) from streaming detection to avoid false positives
• ✅ Remove problematic self.response_parser assignments

Benefits

• 🔄 Dynamic Mode Switching: Clients can reliably switch between streaming/non-streaming
• 🎯 Type-Based Selection: Parser chosen based on actual completion type, not previous calls
• 🛡️ Robust Detection: Excludes basic iterables (strings, dicts) that aren't streaming responses
• 🧹 Cleaner Logic: Eliminates confusing instance variable mutations

Testing

• ✅ Verified dynamic parser selection for all completion types
• ✅ Tested mode switching: streaming → non-streaming → streaming
• ✅ Confirmed fallback behavior for unknown types
• ✅ Ensured basic types (strings) use non-streaming parser
• ✅ Validated backward compatibility with existing functionality

Code Changes

1. Modified parse_chat_completion(): Dynamic parser selection based on completion type
2. Updated sync call path: Removed self.response_parser assignment
3. Updated async call path: Removed self.response_parser assignment
4. Enhanced type detection: Added exclusions for basic types

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• Improves reliability for production use cases

Impact

This fix resolves a critical issue that made the OpenAI client unreliable when users needed to dynamically switch between streaming and non-streaming modes in the same application session.

Fixes #377