fix(ai): honor disabled Azure prompt caching

Author: code-yeongyu

packages/ai/src/api/azure-openai-responses.ts

@@ -294,7 +294,9 @@ function buildParams(
 		model: deploymentName,
 		input: messages,
 		stream: true,
-		prompt_cache_key: clampOpenAIPromptCacheKey(options?.sessionId),
+		...((options?.cacheRetention ?? model.cacheRetention) === "none"
+			? {}
+			: { prompt_cache_key: clampOpenAIPromptCacheKey(options?.sessionId) }),
 		store: false,
 	};
 

packages/ai/src/changes.md

@@ -26,6 +26,20 @@
 - Added `supportsRemoteCompactionV2` so verified OpenAI Responses proxies can explicitly advertise the native
   `compaction_trigger` request contract. Unknown custom proxies remain disabled by default.
 
+## 2026-07-27 - Honor disabled Azure Responses prompt caching
+
+### What changed and why
+
+- `api/azure-openai-responses.ts`: requests with `cacheRetention: "none"` now omit `prompt_cache_key`,
+  matching the OpenAI Responses adapter instead of silently enabling Azure prompt-cache affinity from the
+  session id.
+- `../test/azure-openai-base-url.test.ts`: pins both the existing 64-character cache-key clamp and the
+  disabled-cache omission path.
+
+### Expected merge conflict zones
+
+- LOW: `api/azure-openai-responses.ts` request payload construction.
+
 ## 2026-07-27 - Treat Anthropic policy blocks as classifier refusals
 
 ### What changed and why

packages/ai/test/azure-openai-base-url.test.ts

@@ -157,6 +157,18 @@ describe("azure-openai-responses base URL normalization", () => {
 		expect(azureMock.lastParams?.prompt_cache_key).toBe("x".repeat(64));
 	});
 
+	it("omits prompt_cache_key when caching is disabled", async () => {
+		const model = getModel("azure-openai-responses", "gpt-4o-mini");
+		await streamAzureOpenAIResponses(model, context, {
+			apiKey: "test-api-key",
+			azureBaseUrl: "https://my-resource.openai.azure.com",
+			cacheRetention: "none",
+			sessionId: "cache-disabled-session",
+		}).result();
+
+		expect(azureMock.lastParams).not.toHaveProperty("prompt_cache_key");
+	});
+
 	it("disables server-side response storage", async () => {
 		const model = getModel("azure-openai-responses", "gpt-4o-mini");
 		await streamAzureOpenAIResponses(model, context, {