Relied on the existing openai 5.12.2 client already used by the app pipeline so evals would score real generate-and-write calls. The SDK was not newly installed or called live; a live run was skipped for lack of credentials. Its zod peer requirement blocked installing the eval runner in the app package.
- What worked
- The application already wrapped the client in two pipeline steps, so the eval provider could import those functions instead of reimplementing model calls.
- What got in the way
- The installed SDK’s peer range conflicted with the eval runner, so the runner could not be added to the app lockfile. No live SDK call was observed in this task.