Bug Report: Get duplicated LLM spans in AgentScope applications

[Cirilla-zmh]

Describe your environment

OS: MacOS
Python version: 3.12
Package version: 0.3.0.dev

What happened?

We always get 4 LLM span for one react iteration. Just see:

Steps to Reproduce

1. Download source code of copaw: https://github.com/agentscope-ai/CoPaw
2. Instrument copaw with loongsuite-python-agent, just do:
   a. clone loongsuite-python-agent
   b. install instrumentation packages: pip install loongsuite-python-agent/loongsuite-distro & pip install loongsuite-python-agent/util/opentelemetry-util-genai & pip install loongsuite-python-agent/instrumentation-loongsuite/loongsuite-instrumentation-agentscope
   c. do from opentelemetry.instrumentation.auto_instrumentation import sitecustomize before copaw is initialize
3. Start CoPaw following README: https://github.com/agentscope-ai/CoPaw?tab=readme-ov-file#quick-start
4. Ask CoPaw to complete a simple task with a tool, like "what's on my desktop".

Expected Result

Only one LLM span should be record for each react iteration.

Actual Result

You will gain a trace with duplicated LLM spans.

Additional context

No response

Would you like to implement a fix?

None

Tip

_{React with 👍 to help prioritize this issue. Please use comments to provide useful context, avoiding +1 or me too, to help us triage it. Learn more here.}

[Cirilla-zmh]

This issue can be caused by either of two reasons:

1. In cases where there are multiple layers of wrappers/proxies, if each wrapper/proxy implements the ChatModelBase class, then each layer will enter the __call__ method once, which can result in duplicate LLM spans.

class RetryChatModel(ChatModelBase):
    """Transparent retry wrapper around any :class:`ChatModelBase`.

    The wrapper delegates every call to the underlying *inner* model and
    retries on transient errors with exponential back-off.  Streaming
    responses are also covered: if the stream fails mid-consumption the
    entire request is retried from scratch.
    """

    def __init__(self, inner: ChatModelBase) -> None:
        super().__init__(model_name=inner.model_name, stream=inner.stream)
        self._inner = inner

2. When both AgentScope and the OpenAI/DashScope plugin are installed, they may each emit LLM spans for the same LLM invocation, which can also lead to duplicate spans.

[Cirilla-zmh]

Let’s provide some mechanisms in genai-util to resolve this issue.