Recorder

OpenAI API

Wrap each call. usage_from reads both Chat Completions and Responses usage, including reasoning and cached tokens.

1. Copy harness_recorder.py into your project.
2. Wrap the shared function that calls the model.

with run.agent("classify", prompt_ref="prompts/classifier@3", model="openai/gpt-6-luna") as step:
    resp = client.responses.create(model="gpt-6-luna", input=ticket_text)
    step.model_call(**usage_from(resp))

When streaming, ask for usage in the final chunk (OpenAI only sends it when requested) and record it then.

Read the full guide on GitHub