Bug: Streaming responses missing usage metadata for cost tracking

[imranarshad]

When using streaming mode, the ChatLiteLLM class doesn't provide usage metadata (token counts), making it impossible to track costs during streaming operations. Non-streaming mode works correctly and includes usage metadata.

Steps to Reproduce

from langchain_litellm import ChatLiteLLM
from langchain_core.messages import HumanMessage

# Test streaming (the broken case)
llm = ChatLiteLLM(model="gpt-4o", streaming=True)
message = [HumanMessage(content="Say hello")]

# Collect all chunks and check for usage metadata
chunks = list(llm.stream(message))
usage_found = any(hasattr(chunk, 'usage_metadata') and chunk.usage_metadata for chunk in chunks)
print(f"Streaming usage metadata found: {usage_found}")  # Returns False

# Test non-streaming (the working case)
llm.streaming = False
result = llm.invoke(message)
print(f"Non-streaming usage metadata: {result.usage_metadata is not None}")  # Returns True

Expected Behavior

Both streaming and non-streaming should provide usage metadata:

Streaming usage metadata found: True
Non-streaming usage metadata: True

Actual Behavior

Streaming usage metadata found: False
Non-streaming usage metadata: True

Environment

• langchain-litellm: 0.2.1 (latest)
• Python: 3.11+
• Provider: OpenAI (also affects Anthropic and others)
• OS: Any

Impact

• Impossible to track streaming costs in real-time
• Inconsistent API behavior between streaming/non-streaming modes
• Prevents usage-based billing and monitoring in production applications

Root Cause

1. Missing stream_options={"include_usage": True} parameter
2. Usage data not being extracted from streaming chunks
3. Usage metadata not attached to message chunks

---

This affects all streaming operations across different LLM providers and is critical for production cost monitoring.

[imranarshad]

🎉 FIXED - This issue has been completely resolved in PR #22!

This PR fixes not only the streaming usage metadata issue but also 5 additional critical bugs:

1. ✅ Missing Usage Metadata in Streaming (this issue)
2. ✅ Missing reasoning_content Support for thinking models (o1, Claude, Gemini)
3. ✅ Streaming Crashes with dictionary deltas
4. ✅ Tool Call Processing Failures with certain providers
5. ✅ Incomplete Usage Metadata missing cache/reasoning tokens
6. ✅ Async Streaming Reliability issues

🧪 Test Results - CONFIRMED WORKING

from langchain_litellm import ChatLiteLLM
from langchain_core.messages import HumanMessage

# ✅ FIXED: Streaming now includes usage metadata
llm = ChatLiteLLM(model="gpt-4o", streaming=True)
chunks = list(llm.stream([HumanMessage(content="Hello")]))
usage_found = any(hasattr(chunk, 'usage_metadata') and chunk.usage_metadata for chunk in chunks)
print(f"Streaming usage metadata found: {usage_found}")  # ✅ Now returns True

# Get usage details
final_chunk = chunks[-1]
print(f"Input tokens: {final_chunk.usage_metadata.input_tokens}")    # 12
print(f"Output tokens: {final_chunk.usage_metadata.output_tokens}")  # 5
print(f"Total tokens: {final_chunk.usage_metadata.total_tokens}")    # 17

The fix is comprehensive, tested, and ready for review! 🚀