fix: update stale CI test expectations

packages/ai/CHANGELOG.md

@@ -4,6 +4,7 @@
 
 ### Fixed
 
+- Fixed OpenAI-compatible prompt cache tests to cover proxies that explicitly disable long cache retention.
 - Stopped sending `tools: []` on OpenAI-compatible, Anthropic, OpenAI Responses, OpenAI Codex Responses, and Azure OpenAI Responses requests when no tools are active (e.g. `pi --no-tools`). DashScope/Aliyun Qwen (OpenAI-compatible) rejects empty tools arrays with `"[] is too short - 'tools'"` (HTTP 400); the field is now omitted unless the conversation has tool history (the existing LiteLLM/Anthropic-proxy workaround).
 - Fixed `supportsXhigh()` to recognize DeepSeek V4 Pro, preserving `xhigh` reasoning requests so they map to DeepSeek's `max` effort ([#3662](https://github.com/badlogic/pi-mono/issues/3662))
 - Fixed OpenAI-compatible DeepSeek V4 model replay to include empty `reasoning_content` on assistant messages when needed, preventing OpenRouter DeepSeek V4 sessions from failing after responses without reasoning deltas ([#3668](https://github.com/badlogic/pi-mono/issues/3668))

packages/ai/test/openai-completions-prompt-cache.test.ts

@@ -132,9 +132,10 @@ describe("openai-completions prompt caching", () => {
 		expect(payload?.prompt_cache_retention).toBeUndefined();
 	});
 
-	it("omits prompt cache fields for non-OpenAI base URLs", async () => {
+	it("omits prompt cache fields for non-OpenAI base URLs without compatible long retention", async () => {
 		const model = createModel({
 			baseUrl: "https://proxy.example.com/v1",
+			compat: { supportsLongCacheRetention: false },
 		});
 		const { payload } = await captureRequest({ cacheRetention: "long", sessionId: "session-proxy" }, model);