Refactor Agents SDK instrumentation.

[tcdent]

Thus implementation demonstrates a sustainable pattern for all instrumentation that gets included in AgentOps.

• Uses semantic conventions exclusively.
• All pertinent data is exported, including workflow input/output, tool use, and completions.
• Spans nest under the parent trace and are opened/closed in a lifecycle that is compatible with agent's built-in tracing module.
• Tooling for generating test data from Response objects from openai-agents and completions & responses from openai
• Example code that uses handoffs.
• Better span_names
  • agents.function should be something like agents.tool.<tool_name>
  • agents.response should have the agent name
• Agent span is closing early? Should include the tool and llm calls within it's start/end times.
• Completion data is getting merged with the agent. We should have separate spans for the agent making the action and then the llm call as a child span. This appears to be a limitation of the way openai_agents hands off to tracing; lots of things get categorized as response, but it works much better now.
• instrumentation.openai is incomplete. I created it to be able to instrument responses api calls, but ended up incorporating that into the instrumentation.openai_agents. This will be applied in a new PR.
• Attributes have tests, but the exporter/processor does not.

[the-praxs]

Thanks @tcdent for the PR! This is helpful in moving towards a more stable instrumentation for the OpenAI provider and Agents framework.

I have requested some changes primarily on the semantic conventions being used to record the attributes. This would prevent duplication of data and provide an ordered structure of extracting data from queries in the backend.

[tcdent]

@the-praxs Did you run this? Post your script and output below.

[the-praxs]

@the-praxs Did you run this? Post your script and output below.

https://gist.github.com/the-praxs/73fe232afb46c6cccb110dd4de9db0ba

[codecov]

Codecov Report

Attention: Patch coverage is 64.38172% with 265 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
agentops/instrumentation/openai_agents/exporter.py	46.73%	98 Missing ⚠️
...strumentation/openai_agents/attributes/response.py	36.08%	62 Missing ⚠️
...instrumentation/openai_agents/attributes/tokens.py	57.69%	55 Missing ⚠️
.../instrumentation/openai_agents/attributes/model.py	64.40%	21 Missing ⚠️
...rumentation/openai_agents/attributes/completion.py	87.34%	10 Missing ⚠️
...tops/instrumentation/openai_agents/instrumentor.py	84.09%	7 Missing ⚠️
...instrumentation/openai_agents/attributes/common.py	93.65%	4 Missing ⚠️
agentops/instrumentation/openai_agents/__init__.py	76.92%	3 Missing ⚠️
agentops/helpers/serialization.py	88.23%	2 Missing ⚠️
...gentops/instrumentation/openai_agents/processor.py	91.66%	2 Missing ⚠️
... and 1 more

📢 Thoughts on this report? Let us know!

[the-praxs]

Testing results here -

https://gist.github.com/the-praxs/251d08f2aca7810759c3b954305f5850

[the-praxs]

Tested streaming with examples/agents-examples/basic/stream_text.py and it works as expected.

[the-praxs]

LGTM AND YOLO