From 4725b55566e0812baf03046a711da35f4121f0bf Mon Sep 17 00:00:00 2001
From: Juergen Kunz <juergen@foss.global>
Date: Thu, 14 May 2026 19:59:30 +0000
Subject: [PATCH] fix(openai): strip unsupported ChatGPT prompt cache retention
 options while preserving prompt cache keys

---
 changelog.md                    |  7 +++++++
 test/test.openai-auth.node.ts   |  8 ++++++++
 ts/smartai.middleware.openai.ts | 29 +++++++++++++++++++----------
 3 files changed, 34 insertions(+), 10 deletions(-)

diff --git a/changelog.md b/changelog.md
index 7085741..6803529 100644
--- a/changelog.md
+++ b/changelog.md
@@ -4,6 +4,13 @@
 
 
 
+### Fixes
+
+- strip unsupported ChatGPT prompt cache retention options while preserving prompt cache keys (openai)
+  - Removes promptCacheRetention values before sending requests to the ChatGPT Codex backend.
+  - Keeps prompt_cache_key forwarding intact for OpenAI provider options.
+  - Only rewrites system prompts into top-level instructions when needed, avoiding unnecessary prompt changes.
+
 ## 2026-05-14 - 4.0.1
 
 ### Fixes
diff --git a/test/test.openai-auth.node.ts b/test/test.openai-auth.node.ts
index 3b45163..93e2e3d 100644
--- a/test/test.openai-auth.node.ts
+++ b/test/test.openai-auth.node.ts
@@ -203,6 +203,12 @@ tap.test('getModel uses ChatGPT Codex backend for OpenAI ChatGPT auth', async ()
       model,
       system: 'system prompt',
       prompt: 'hello',
+      providerOptions: {
+        openai: {
+          promptCacheKey: 'session-1',
+          promptCacheRetention: 'in_memory',
+        },
+      },
     });
 
     expect(capturedRequest?.url).toEqual('https://chatgpt.com/backend-api/codex/responses');
@@ -214,6 +220,8 @@ tap.test('getModel uses ChatGPT Codex backend for OpenAI ChatGPT auth', async ()
     expect(capturedBody.input).toEqual([
       { role: 'user', content: [{ type: 'input_text', text: 'hello' }] },
     ]);
+    expect(capturedBody.prompt_cache_key).toEqual('session-1');
+    expect(capturedBody.prompt_cache_retention).toEqual(undefined);
   } finally {
     globalThis.fetch = originalFetch;
   }
diff --git a/ts/smartai.middleware.openai.ts b/ts/smartai.middleware.openai.ts
index 0af1d59..6c32ec7 100644
--- a/ts/smartai.middleware.openai.ts
+++ b/ts/smartai.middleware.openai.ts
@@ -11,6 +11,10 @@ const getSystemInstructions = (prompt: LanguageModelV3CallOptions['prompt']): st
   return instructions.length > 0 ? instructions.join('\n') : undefined;
 };
 
+const hasUnsupportedChatGptCacheRetention = (options: JSONObject): boolean => {
+  return options.promptCacheRetention !== undefined || options.prompt_cache_retention !== undefined;
+};
+
 /**
  * ChatGPT's Codex backend requires top-level Responses API instructions.
  * The standard OpenAI provider otherwise serializes system prompts as input items.
@@ -20,25 +24,30 @@ export function createOpenAiChatGptInstructionsMiddleware(): LanguageModelV3Midd
     specificationVersion: 'v3',
     transformParams: async ({ params }) => {
       const instructions = getSystemInstructions(params.prompt);
-      if (!instructions) {
+      const providerOptions = params.providerOptions ?? {};
+      const openAiProviderOptions = providerOptions.openai ?? {};
+      const shouldApplyInstructions = !!instructions && !isNonEmptyString(openAiProviderOptions.instructions);
+      const shouldStripCacheRetention = hasUnsupportedChatGptCacheRetention(openAiProviderOptions);
+
+      if (!shouldApplyInstructions && !shouldStripCacheRetention) {
         return params;
       }
 
-      const providerOptions = params.providerOptions ?? {};
-      const openAiProviderOptions = providerOptions.openai ?? {};
-      if (isNonEmptyString(openAiProviderOptions.instructions)) {
-        return params;
+      const nextOpenAiProviderOptions: JSONObject = { ...openAiProviderOptions };
+      delete nextOpenAiProviderOptions.promptCacheRetention;
+      delete nextOpenAiProviderOptions.prompt_cache_retention;
+      if (shouldApplyInstructions) {
+        nextOpenAiProviderOptions.instructions = instructions;
       }
 
       return {
         ...params,
-        prompt: params.prompt.filter((message) => message.role !== 'system'),
+        prompt: shouldApplyInstructions
+          ? params.prompt.filter((message) => message.role !== 'system')
+          : params.prompt,
         providerOptions: {
           ...providerOptions,
-          openai: {
-            ...openAiProviderOptions,
-            instructions,
-          } as JSONObject,
+          openai: nextOpenAiProviderOptions,
         },
       } satisfies LanguageModelV3CallOptions;
     },