OpenAI Responses API Does Not Persist Instructions Across Chained Turns

OpenAI's Responses API allows developers to chain conversation turns using a previous_response_id field, but this mechanism carries forward only conversation state, not top-level instructions. This means output format rules, tool constraints, or safety policies set in an earlier turn are silently absent in follow-up requests unless explicitly repeated. A developer writing for DEV Community highlighted this counterintuitive boundary and demonstrated how accidental omission can cause an application to respond outside its expected policy contract. To address this, the author built a request constructor that enforces non-empty instructions on every policy-bound turn and rejects conflicting state fields at the application level. The approach treats instruction repetition as an application invariant rather than an API requirement, making policy enforcement explicit and deliberate across all turns.