OpenAI Integration: `truncate_messages_by_size` fails on Pydantic SDK objects (e.g., `ResponseFunctionToolCall`)

[mecampbellsoup]

How do you use Sentry?

Sentry Saas (sentry.io)

Version

2.49.x (and likely all versions with the OpenAI integration)

Steps to Reproduce

1. Enable the OpenAI integration with sentry-sdk[openai]
2. Use the OpenAI Responses API with tool calls
3. When building the follow-up API call after tool execution, include SDK objects (like ResponseFunctionToolCall) in the input list - as shown in OpenAI's documentation

from openai.types.responses import ResponseFunctionToolCall

# After a tool call, building input for follow-up request
input_list = [
    {"role": "user", "content": "Please notify the team"},
    tool_call,  # ResponseFunctionToolCall - a Pydantic object
    {"type": "function_call_output", "call_id": tool_call.call_id, "output": "Success"},
]

# This triggers Sentry's instrumentation which calls json.dumps() on input_list
response = client.responses.create(model="gpt-4o", input=input_list)

Expected Result

Sentry's instrumentation should handle Pydantic/SDK objects gracefully, either by:

• Calling .model_dump() on them (as _normalize_data() already does)
• Using a fallback serialization strategy

Actual Result

TypeError: Object of type ResponseFunctionToolCall is not JSON serializable

The error originates from sentry_sdk/ai/utils.py:truncate_messages_by_size() which calls json.dumps(messages) directly without normalizing Pydantic objects first.

Root Cause

The SDK already has _normalize_data() in sentry_sdk/ai/utils.py that correctly handles Pydantic objects:

def _normalize_data(data: "Any", unpack: bool = True) -> "Any":
    # convert pydantic data (e.g. OpenAI v1+) to json compatible format
    if hasattr(data, "model_dump"):
        return _normalize_data(data.model_dump(), unpack=unpack)
    ...

However, truncate_messages_by_size() doesn't use it:

def truncate_messages_by_size(messages, ...):
    serialized_json = json.dumps(messages, separators=(",", ":"))  # No normalization!
    ...

Suggested Fix

Call _normalize_data() before json.dumps():

def truncate_messages_by_size(messages, ...):
    normalized = _normalize_data(messages, unpack=False)
    serialized_json = json.dumps(normalized, separators=(",", ":"))
    ...

Or apply the same fix in truncate_and_annotate_messages() before calling truncate_messages_by_size().

Additional Context

This is the same class of bug as #4923 (handling Omit/NotGiven types) - the OpenAI integration assumes inputs are plain Python dicts, but OpenAI's SDK uses Pydantic objects throughout.

OpenAI's own documentation shows passing SDK objects directly in follow-up requests, so this is a common pattern that Sentry's instrumentation should handle.

[linear]

PY-2043

[alexander-alderman-webb]

Thanks for raising the issue @mecampbellsoup. Please see my response in #5351 (comment).