Errors and limits

Treat every API response as part of the contract. Authenticate with the minimum scope, validate request input locally, distinguish client errors from transient service failures, and retry only operations that are safe to retry.

For long-running work, enforce time, output, budget, and concurrency limits before the request. Use idempotency for the same intended inference operation, and preserve lifecycle receipts for sandbox cleanup and reconciliation.

Next: OpenAI-compatible API.