To maintainers: This is a THREE LINE change which fixes a nasty bug. Also I opened this before I read CONTRIBUTING.md telling me not to open PRs. Sorry! Please don't ban me!

Summary

packages/ai/src/providers/azure-openai-responses.ts requests include: ["reasoning.encrypted_content"] (the stateless reasoning-replay path) but never sets store, so Azure runs in its default stateful mode. The sibling openai-responses and openai-codex-responses providers both set store: false alongside the same include; Azure is the only Responses provider that omits it.

Symptom

Intermittent 400 Item with id 'rs_...' not found on multi-turn conversations, which clears if you retry the same request.

Root cause

When a transcript is replayed, convertResponsesMessages re-emits each prior assistant reasoning item with its original rs_* id inlined in input (the item is round-tripped through thinkingSignature). In stateful mode Azure tries to resolve that rs_* id against its server-side store rather than using the inlined encrypted_content. Azure's Responses endpoint is load-balanced and stored items are tied to a specific response/backend with a retention window, so whether the lookup succeeds is non-deterministic — replays that land on a backend without the item fail, and retrying eventually routes to one that resolves (or regenerates), which is why "try again" clears it.

This bites hardest when the transcript is rebuilt from persisted messages each turn (no previous_response_id threading), so the replayed reasoning ids reference responses the serving backend no longer has.

Fix

Set store: false in buildParams. With store: false + include: ["reasoning.encrypted_content"], every request is self-contained: the reasoning travels inline as encrypted content and Azure never does a store lookup, so the call is deterministic. This matches the existing behavior of openai-responses and openai-codex-responses.

 	const params: ResponseCreateParamsStreaming = {
 		model: deploymentName,
 		input: messages,
 		stream: true,
 		prompt_cache_key: clampOpenAIPromptCacheKey(options?.sessionId),
+		store: false,
 	};

Test

Extended packages/ai/test/azure-openai-base-url.test.ts to assert the provider sends store: false. npm run check and the full packages/ai test suite pass (323 passed, e2e/keyed tests skipped).