Summary

The OpenAI SDK v1+ returns Pydantic model instances (e.g., ResponseFunctionToolCall) which are not directly JSON serializable. When these objects are passed to truncate_messages_by_size(), the json.dumps() call fails with a TypeError:

TypeError: Object of type ResponseFunctionToolCall is not JSON serializable

The _normalize_data() helper already exists in this file and properly handles Pydantic objects by calling .model_dump(), but it wasn't being used in truncate_messages_by_size().

The Fix

This PR adds normalization at the start of truncate_messages_by_size() to ensure all Pydantic objects are converted to JSON-compatible dicts before serialization:

# Normalize messages to ensure JSON serialization works
# (handles Pydantic objects from OpenAI SDK v1+)
messages = _normalize_data(messages, unpack=False)

Reproduction

When using the OpenAI SDK's Responses API with tool calls and Sentry's OpenAI integration enabled:

from openai import OpenAI
from openai.types.responses import ResponseFunctionToolCall
client = OpenAI()
response = client.responses.create(
 model="gpt-4o",
 input=[
 {"role": "user", "content": "What's the weather?"},
 {
 "role": "assistant", 
 "content": [ResponseFunctionToolCall(
 type="function_call",
 id="call_123",
 call_id="call_123",
 name="get_weather",
 arguments='{"location": "NYC"}'
 )]
 }
 ]
)

This crashes with the TypeError because Sentry's integration tries to serialize the messages for tracing.

Fixes #5350