Trail every OpenAI call

One wrapper on your OpenAI client, and every chat completion your system makes becomes a signed, append-only loop event. No proxy in your request path, no gateway to operate, no call sites to change.

pip install looptail   # openai SDK stays yours, any version

import openai
import looptail
from looptail import instrument

looptail.init(app="checkout-assistant")
client = instrument.openai(openai.OpenAI())

# chat.completions.create and responses.create are now recorded
response = client.chat.completions.create(
    model="your-model",
    messages=[{"role": "user", "content": "Where is my order?"}],
)

What gets recorded

Each call appends one observe event to the app's trail file: provider and API, model id, truncated input and output previews, prompt_tokens /completion_tokens from the usage object, the finish reason, and duration. Exceptions are recorded and re-raised — your error handling is untouched. Records are hash-chained and Ed25519-signed at write time, so the file is tamper-evident: anyone can verify it without trusting you.

npx @looptail/cli verify --app checkout-assistant

✔ 8,341 loop events
✔ chain intact · signatures valid

From logs to evidence

Observability tools will happily show you these calls on a dashboard. What they don't give you is a record that stands up later — when a customer disputes what the assistant said, when a security review asks how changes are controlled, when theEU AI Act asks for automatic lifetime logging. The trail is that record: complete per decision, linked to theevaluations and changes that follow, and exportable as adated, self-verifying evidence pack. Local-first — no account needed until you want hosted sync and signed receipts.