Python: [Feature]: Support OpenAI instructions in Responses API

[PratikWayase]

Motivation & Context

The OpenAI Responses API supports an instructions parameter that applies instructions to a single request ephemerally, without persisting in the conversation state. Previously, the agent framework intercepted this parameter and injected it as a system message into the chat history. This caused the instructions to persist across all future turns in the session, defeating the purpose of the ephemeral parameter and making it impossible to do per-request steering (like temporarily asking the agent to refuse a request). This PR enables developers to use the native OpenAI instructions parameter as intended for temporary, per-turn overrides.

Description & Review Guide

• What are the major changes?
  • Removed instructions from the exclude_keys set in _prepare_options in _chat_client.py, allowing it to flow natively into the OpenAI API payload.
  • Removed the logic that prepended instructions as a system message to the messages array.
  • Removed the now-unused prepend_instructions_to_messages import.
  • Added the instructions: str field to the OpenAIChatOptions TypedDict for type safety and IDE autocompletion.
  • Updated test_openai_chat_client.py: removed outdated tests that expected system message injection, added a new parameterized test (test_instructions_passed_natively_not_as_system_message) to verify instructions are passed top-level and don't mutate the input messages, and updated test_get_response_with_all_parameters to reflect the new behavior.
• What is the impact of these changes?
  • instructions passed via options will now only apply to the current response. Continuation turns using conversation_id or previous_response_id will no longer inherit the temporary instructions.
• What do you want reviewers to focus on?
  • Verify that removing the manual prepend logic in _prepare_options correctly passes the instructions parameter to the underlying OpenAI SDK without side effects.
  • Ensure the test coverage adequately proves the ephemeral nature of the instructions across continuation turns.

Related Issue

Fixes #7056

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

Pull request overview

Enables native use of the OpenAI Responses API instructions parameter in the Python OpenAI chat client so per-request steering remains ephemeral (i.e., it no longer persists by being injected into conversation message history).

Changes:

• Stops converting options["instructions"] into a prepended system message; allows it to flow through as the top-level Responses API instructions field.
• Updates OpenAIChatOptions to include an instructions: str entry for type safety/autocomplete.
• Updates unit tests to assert instructions is passed top-level and that input is not mutated with an injected system message.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
python/packages/openai/agent_framework_openai/_chat_client.py	Removes system-message injection behavior and passes instructions natively to the Responses API.
python/packages/openai/tests/openai/test_openai_chat_client.py	Updates/rewrites tests to validate native instructions passthrough and absence of injected system messages.

[giles17]

Automated Code Review

Reviewers: 5 | Confidence: 93% | Result: All clear

Reviewed: Correctness, Security Reliability, Test Coverage, Failure Modes, Design Approach

---

Automated review by giles17's agents

[github-actions]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/openai/agent_framework_openai
_chat_client.py	1366	117	91%	312, 325, 675–679, 687–690, 696–700, 750–757, 759–761, 768–770, 828, 836, 859, 1076, 1135, 1137, 1139, 1141, 1207, 1221, 1301, 1311, 1316, 1359, 1471–1472, 1487, 1752, 1859, 1864–1865, 1948, 1958, 1985, 1991, 2001, 2007, 2012, 2018, 2023–2024, 2104, 2148, 2151–2154, 2168, 2178–2179, 2191, 2233, 2298, 2315, 2318, 2345–2347, 2386, 2403, 2406, 2468, 2475, 2512–2513, 2548, 2586–2587, 2605–2606, 2778–2779, 2797, 2883–2891, 3069, 3084, 3134, 3144–3146, 3173–3175, 3185–3186, 3192, 3207, 3340–3341
TOTAL	45562	4474	90%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
9406	33 💤	0 ❌	0 🔥	2m 34s ⏱️

[giles17]

Hey @PratikWayase thanks for the contribution! Can you please address the failing typing checks

[gnanirahulnutakki]

I verified the native Responses API passthrough against OpenAI Python SDK 2.45.0 and the current official API contract, then reproduced the PR's typing gate locally. The runtime behavior is sound; the inline comment identifies the isolated test-typing failure and a validated minimal fix.

[eavanvalkenburg]

do we need to do the same for other providers?

[PratikWayase]

do we need to do the same for other providers?

I kept this PR scoped strictly to the OpenAI Responses API fix since that's where the native instructions parameter is specifically supported. I can open a separate follow-up issue/PR to audit the other providers (Anthropic, Gemini) to ensure they don't have redundant fields or suboptimal instruction handling. Let me know if you'd prefer I tackle that here or in a follow-up

[gnanirahulnutakki]

Re-reviewed exact head 8a6a52c9b1b2 after the rebase. The original OpenAIChatOptions typing defect is resolved. I reran the four focused instructions/all-parameters runtime cases, all five package typing checkers (mypy, pyright, pyrefly, ty, and zuban), and strict package Pyright; everything passes.