[codeministry]

Why Spring AI stays at 2.0.1

Spring AI 2.0.2 breaks stopping a streaming chat turn: four tests fail (ChatStreamTest.stopCancelsTheModelCallAndKeepsThePartialAnswerMarkedStopped, both stop-on-delete tests in ChatConversationTest, ChatTurnDeadlineTest.aTurnPastItsDeadlineIsStoppedAndFreesItsThread). In each case the model stub never sees its connection closed: the turn ends on our side, but the model call keeps running to the end.

Root cause: an upstream regression in openai-java 4.66.0

• Spring AI 2.0.2 bumps the OpenAI SDK from 4.49.0 to 4.80.0 (spring-projects/spring-ai@351c272, #7170). Spring AI's own streaming code did not change in a way that matters: OpenAiChatModel.internalStream still does sink.onDispose(response::close).
• openai-java 4.66.0 (openai/openai-java#1027, commit 1441ba1) reordered StreamResponse.close() in StreamHandler.kt: the transport response is now closed before the BufferedReader. Before, reader.close() ran first and waited on the reader's lock until the in-flight readLine() on the stream-handler thread returned, so the OkHttp body was closed with no read in progress.
• We cancel from another thread (takeUntilOther on the stop signal or the deadline), while the stream-handler thread is blocked in a socket read. Closing the OkHttp body concurrently makes ChunkedSource.close() → discard() → Okio AsyncTimeout.enter() throw IllegalStateException: Unbalanced enter/exit. That skips noNewExchanges()/responseBodyComplete(), so the connection is neither released nor closed, and the exception is swallowed by the whenComplete in AsyncStreamResponse.close().

Evidence

• Spring AI 2.0.2 held fixed, only com.openai:openai-java-core forced: bisected over every release from 4.49.0 to 4.80.0. 4.65.0 green, 4.66.1 red (#1027 is tagged in v4.66.0, which never reached Maven Central; 4.66.1 is the first published release with it); the same four tests flip, the other 16 stay green. Still red on today's 4.92.0, and StreamHandler.kt on openai-java main is unchanged since #1027.
• A throwaway probe wrapping the SDK's HttpClient logged the close() calls: with 4.65.0 the reader closes, then the response, and the stub sees the reset; with 4.80.0 and 4.92.0 response.close() throws IllegalStateException: Unbalanced enter/exit and the stub writes every remaining chunk.
• Plain OkHttp 4.12.0, no SDK: closing a ResponseBody while another thread is blocked reading it throws the same exception and leaves the connection in the pool; closing the reader first, or Call.cancel(), closes the socket at once.
• Cancelling on the stream-handler thread itself (e.g. take(3) inside onNext) still closes the connection with 4.80.0. A stop or a deadline cannot cancel on that thread: it is blocked waiting for the next chunk.

The app code is not at fault: disposing the Flux from ChatModel.stream() is Spring AI's way to cancel a call, and the tests state user-visible guarantees (stop ends the model call, deadline frees the thread, delete stops the streaming turn), so neither is changed.

What this PR changes

• gradle/libs.versions.toml: a comment next to springAi saying why it is held at 2.0.1 and when to retry.
• renovate.json: Spring AI gets its own pull request instead of riding in "backend (minor and patch)". Held inside the group, it turns the whole group red every week (as in Update backend (minor and patch) #40 and chore(deps): Renovate updates of 2026-10-26 combined #47); on its own, its CI result says whether the next release fixed it.

No code or test changes.

Verification

• ./gradlew -p backend check green on this branch.
• renovate-config-validator renovate.json: valid.