tracerazor-openai 0.2.0

Drop-in TraceRazor-instrumented OpenAI client. Zero-friction token efficiency auditing.

tracerazor-openai

Drop-in replacement for openai.OpenAI with automatic TraceRazor token-efficiency auditing. Zero friction — change one import line and every chat.completions.create() call is captured as a reasoning step.

Install

pip install tracerazor-openai

You'll also need the tracerazor binary on your PATH (or a running tracerazor-server).

Use

from tracerazor_openai import OpenAI

client = OpenAI(agent_name="support-bot")

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Refund order ORD-9182"}],
)

# ... more calls ...

report = client.audit()
print(report.summary())

Constructor accepts every kwarg the plain openai.OpenAI does plus:

Kwarg	Default	Meaning
agent_name	"openai-agent"	Name used in the TraceRazor report
tracer	None	Append to an existing Tracer instead of creating one
server	None	URL of a running tracerazor-server (default: local binary)

How it works

The wrapper uses composition — every attribute of the underlying openai.OpenAI client is proxied via __getattr__. Only chat.completions.create() is intercepted to record the call as a reasoning step. If instrumentation fails for any reason (parsing, network, etc.) the original response is still returned unchanged.

License

Apache-2.0