fix(closes OPEN-8565): prevent multiple patching in trace_litellm function to avoid duplicate traces

[viniciusdsmello]

Pull Request

Summary

Fix duplicate traces issue in the LiteLLM tracer by making trace_litellm() idempotent.

Changes

• Added _litellm_traced module-level flag to track if patching has been applied
• Added guard in trace_litellm() to prevent multiple patching of litellm.completion

Context

When trace_litellm() is called multiple times (e.g., accidentally placed inside a request handler instead of at initialization), each call wraps the already-wrapped litellm.completion function, creating a chain of wrappers. This causes each LLM completion to generate multiple traces on the Openlayer platform.

Root cause: trace_litellm() captures litellm.completion as original_completion and replaces it with a traced version. On subsequent calls, the "original" is actually the previously wrapped version, leading to nested tracing.

Solution: Add a module-level flag that ensures the patching only happens once, regardless of how many times trace_litellm() is called.

Testing

• Manual testing
• Unit tests
• Postman CI/CD
• Other (please specify)

Manual testing performed:

1. Verified trace_litellm() can be called multiple times without error
2. Confirmed only a single trace is generated per LLM completion after the fix
3. Verified debug logs correctly indicate when patching is applied vs skipped