Auto-instrumentation via node hook breaks OpenAI helper APIs (parse/withResponse/stream helpers)

[lforst]

Summary

Under node --import braintrust/hook.mjs, several OpenAI SDK helper APIs stop working even though the equivalent lower-level calls still work.

What is broken

These helper paths fail in auto-instrumented mode:

• chat.completions.parse(...)
• .withResponse()-based request helpers
• sync stream helper paths like chat.completions.stream(...)

In practice, the auto-instrumented e2e scenario had to fall back to plain create(...) / streamed create(...) calls to keep coverage running.

Why this matters

This means the auto-instrumented path is not API-compatible with normal SDK usage for some common OpenAI helpers. A user can write valid OpenAI code that works normally, then see it break only when enabling Braintrust auto-instrumentation.

Repro

The aligned e2e coverage in e2e/scenarios/openai-auto-instrumentation-node-hook reproduces this. One concrete failure we hit was:

• _thenUnwrap is not a function from chat.completions.parse(...) under the node hook

Expected

The node-hook auto-instrumented path should preserve normal behavior for OpenAI helper APIs, not require falling back to lower-level request methods.