Clearing a compaction session retains the previous response chain

[sylvesterkaczmarek]

Describe the bug

OpenAIResponsesCompactionSession.clear_session() clears the underlying history and cached compaction items, and resets _deferred_response_id, but it leaves _response_id and _last_unstored_response_id unchanged.

As a result, later compaction can reuse response-chain state from history the caller explicitly cleared. In previous_response_id mode, a stale response ID can still be sent to responses.compact after the session is empty.

Reproduction

session = OpenAIResponsesCompactionSession(
    session_id="test",
    underlying_session=mock_session,
    client=mock_client,
    should_trigger_compaction=lambda _context: False,
)

await session.run_compaction({"response_id": "resp-old", "store": True})
await session.clear_session()

await session.run_compaction(
    {"force": True, "compaction_mode": "previous_response_id"}
)

On current main, the second call proceeds instead of raising that no response ID is available for the cleared session. A focused regression reaches the mocked compact client after the clear.

Expected behavior

A successful clear_session() should reset all response-chain state owned by the wrapper: _response_id, _deferred_response_id, and _last_unstored_response_id, in addition to the cached history.

After clear, explicit previous_response_id compaction without a newly supplied response ID should fail before any client call. If the underlying clear fails, response-chain state should remain unchanged because the old history may still exist.

[Neonstudio28]

I implemented and tested a focused fix on the published fork branch fix/issue-4864-clear-response-chain, commit ce343f0. It clears _response_id, _deferred_response_id, and _last_unstored_response_id after a successful clear, preserves them if the underlying clear fails, and adds regression tests. The focused compaction test file passes (66 tests), plus Ruff, mypy, and pyright for the changed file.

[sylvesterkaczmarek]

Thanks for validating the same fix shape. There is already an open implementation PR for this issue, #4865, covering the response-chain reset and regressions. To avoid duplicate implementation/review work, please consolidate any follow-up there rather than opening a second PR.