Python: fix: prevent repeating instructions in continued Responses API conversations

[Eduard van Valkenburg (eavanvalkenburg)]

Summary

Fixes #3498

This PR prevents instructions from being repeatedly sent to the OpenAI Responses API when continuing a conversation.

Changes

• Modified _prepare_options() in packages/core/agent_framework/openai/_responses_client.py to prepend instructions only on the first turn (when no conversation_id/response_id exists).
• When a conversation/response ID is present, instructions are skipped since they already exist in server-side conversation context.
• Applies to both OpenAI and Azure Responses API clients (Azure inherits from OpenAI implementation).
• Consolidated continuation regression tests into the existing responses suite file instead of a standalone test file.

Testing

• Added/updated regression tests in packages/core/tests/openai/test_openai_responses_client.py covering:
  • First turn includes instructions.
  • Continuation with resp_ format ID skips instructions.
  • Continuation with conv_ format ID skips instructions.
  • First-turn then continuation behavior in a single conversation flow.
• uv run ruff format packages/core/tests/openai/test_openai_responses_client.py
• uv run ruff check packages/core/tests/openai/test_openai_responses_client.py
• uv run python -m pytest -q packages/core/tests/openai/test_openai_responses_client.py -k "instructions_sent_first_turn_then_skipped_for_continuation or instructions_not_repeated_for_continuation_ids or instructions_included_without_conversation_id"

Behavior

Before: Instructions were sent with every API call, causing repeated system messages and extra token usage.

After: Instructions are sent only on the first turn. Subsequent turns with conversation/response IDs do not prepend instructions again.

[Mark Wallace (markwallace-microsoft)]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework/openai
_responses_client.py	613	79	87%	294–297, 301–302, 305–306, 312–313, 318, 331–337, 358, 366, 389, 552, 555, 610, 614, 616, 618, 620, 696, 706, 711, 754, 831, 848, 861, 917, 1010, 1015, 1019–1021, 1025–1026, 1049, 1118, 1140–1141, 1156–1157, 1175–1176, 1307–1308, 1324, 1326, 1405–1413, 1508, 1563, 1578, 1614–1615, 1617–1619, 1633–1635, 1645–1646, 1652, 1667
TOTAL	20473	3250	84%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
3922	225 💤	0 ❌	0 🔥	1m 7s ⏱️

[Copilot]

Pull request overview

This PR fixes a token waste issue where instructions were being repeatedly sent to the OpenAI Responses API on every turn of a conversation, even though the API maintains server-side state and only needs instructions on the first turn.

Changes:

• Modified _prepare_options in OpenAIResponsesClient to conditionally prepend instructions only when no conversation_id exists (first turn)
• Added comprehensive regression tests to verify instructions are sent on first turn but skipped on continuations
• Fix applies to both OpenAI and Azure Responses API clients through inheritance

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
python/packages/core/agent_framework/openai/_responses_client.py	Added logic to skip instruction prepending when conversation_id (resp_* or conv_*) is present, preventing repeated system messages
python/packages/core/tests/openai/test_responses_instructions_continuation.py	New test file with 4 comprehensive tests covering first turn, continuation with resp_ ID, continuation with conv_ ID, and baseline without conversation_id