openai_agents: tool wrapper is re-applied on every turn

[fuchengwarrenzhu]

How do you use Sentry?

Sentry SaaS (sentry.io)

Version

2.68.1

Steps to Reproduce

We ran into the tool-rebuild problem that #6961 fixed. Our agent crash-looped on UserError: ToolSearchTool() requires at least one searchable Responses surface, because defer_loading came back False on every tool. Moving to 2.67.0 fixed it for us, and the switch to RunHooks in #6994 looks like the right call.

While reading that fix I noticed one thing still worth reporting in patches/tools.py. It now sets on_invoke_tool in place:

tool.on_invoke_tool = create_wrapped_invoke(tool, original_on_invoke)

get_all_tools returns the agent's own tool objects and runs once per turn, so the wrapper goes on again on each turn and the previous one stays underneath it.

That file is only wired up for openai-agents < 0.3.2, so this repro calls _get_all_tools directly with a stub standing in for the original:

import asyncio, sentry_sdk
sentry_sdk.init(dsn="https://a@b.ingest.sentry.io/1", traces_sample_rate=1.0)
from sentry_sdk.integrations.openai_agents.patches.tools import _get_all_tools
from agents import Agent, FunctionTool

async def invoke(ctx, args): return "ok"

tool = FunctionTool(
    name="t", description="d",
    params_json_schema={"type": "object", "properties": {}, "additionalProperties": False},
    on_invoke_tool=invoke,
)
agent = Agent(name="a", tools=[tool])

async def original(a, c): return list(a.tools)   # stands in for get_all_tools

async def main():
    for turn in range(1, 6):
        await _get_all_tools(original, agent, None)
        f, depth = agent.tools[0].on_invoke_tool, 0
        while hasattr(f, "__wrapped__"):
            f, depth = f.__wrapped__, depth + 1
        print(f"after turn {turn}: wrapper depth = {depth}")

asyncio.run(main())

Expected Result

A tool gets wrapped once, however many turns a run takes, and one tool call produces one gen_ai.execute_tool span.

Actual Result

after turn 1: wrapper depth = 1
after turn 2: wrapper depth = 2
after turn 3: wrapper depth = 3
after turn 4: wrapper depth = 4
after turn 5: wrapper depth = 5

A tool call on turn N opens N nested gen_ai.execute_tool spans, and the wrapping stays on the caller's Agent after the run.

_patch_run_hooks in patches/runner.py already guards against this with _sentry_is_patched. The same flag on the tool would cover it.

Tested on Python 3.13 with openai-agents 0.17.5.

[linear-code]

PY-2753

[alexander-alderman-webb]

Thank you @fuchengwarrenzhu.

Yes, I'm aware of this problem. This happened before our fix in 2.67.0.
I'll carve out time to fix this for you.