From 61ee2e9717ae70c03069cca9a64fd873aa90c256 Mon Sep 17 00:00:00 2001
From: luvs01 <27862058+luvs01@users.noreply.github.com>
Date: Wed, 19 Aug 2026 11:47:36 +0900
Subject: [PATCH] fix(responses): strip unsupported cache retention

---
 src/adapters/openai-responses.ts           | 17 ++++++++++----
 tests/openai-responses-passthrough.test.ts | 27 ++++++++++++++++++++--
 2 files changed, 38 insertions(+), 6 deletions(-)

diff --git a/src/adapters/openai-responses.ts b/src/adapters/openai-responses.ts
index a8bbab8bfd..bd4959d797 100644
--- a/src/adapters/openai-responses.ts
+++ b/src/adapters/openai-responses.ts
@@ -909,15 +909,24 @@ function stripStatefulResponsesParams(body: unknown): unknown {
  * with `{"detail":"Unsupported parameter: …"}` (strict allowlist). Codex CLI never
  * sends these — it controls output length via `reasoning.effort` — but third-party
  * Responses API clients (GJC, SDK wrappers) include `max_output_tokens` per the
- * public spec. `metadata` is likewise absent from the allowlist. No-op when the
- * body carries neither field, keeping the common Codex path allocation-free.
+ * public spec. `metadata` is likewise absent from the allowlist. Codex can also
+ * send `prompt_cache_retention`, but the ChatGPT backend rejects it for models
+ * that do not expose extended retention instead of falling back to normal cache
+ * retention. No-op when the body carries none of these fields, keeping the common
+ * Codex path allocation-free. API-key Responses requests retain all three fields.
  */
 function stripUnsupportedForwardParams(body: unknown): unknown {
   if (!isPlainObject(body)) return body;
   const hasMot = Object.prototype.hasOwnProperty.call(body, "max_output_tokens");
   const hasMeta = Object.prototype.hasOwnProperty.call(body, "metadata");
-  if (!hasMot && !hasMeta) return body;
-  const { max_output_tokens: _mot, metadata: _meta, ...rest } = body;
+  const hasPromptCacheRetention = Object.prototype.hasOwnProperty.call(body, "prompt_cache_retention");
+  if (!hasMot && !hasMeta && !hasPromptCacheRetention) return body;
+  const {
+    max_output_tokens: _mot,
+    metadata: _meta,
+    prompt_cache_retention: _promptCacheRetention,
+    ...rest
+  } = body;
   return rest;
 }
 
diff --git a/tests/openai-responses-passthrough.test.ts b/tests/openai-responses-passthrough.test.ts
index c2e9d38a71..bff02a5aae 100644
--- a/tests/openai-responses-passthrough.test.ts
+++ b/tests/openai-responses-passthrough.test.ts
@@ -804,8 +804,13 @@ describe("OpenAI Responses passthrough sanitization", () => {
     expect(body.prompt_cache_key).toBe("project-cache-v1");
   });
 
-  test("preserves prompt_cache_retention in the raw Responses passthrough body", () => {
-    const adapter = createResponsesPassthroughAdapter(provider);
+  test("preserves prompt_cache_retention for OpenAI API-key Responses requests", () => {
+    const adapter = createResponsesPassthroughAdapter({
+      adapter: "openai-responses",
+      baseUrl: "https://api.openai.com/v1",
+      authMode: "key",
+      apiKey: "sk-test",
+    });
     const request = adapter.buildRequest({
       modelId: "gpt-5.5",
       context: { messages: [] },
@@ -822,6 +827,24 @@ describe("OpenAI Responses passthrough sanitization", () => {
     expect(body.prompt_cache_retention).toBe("24h");
   });
 
+  test("strips prompt_cache_retention from ChatGPT forward requests", () => {
+    const adapter = createResponsesPassthroughAdapter(provider);
+    const request = adapter.buildRequest({
+      modelId: "gpt-5.5",
+      context: { messages: [] },
+      stream: true,
+      options: {},
+      _rawBody: {
+        model: "gpt-5.5",
+        input: "hi",
+        prompt_cache_retention: "24h",
+      },
+    }, { headers: new Headers({ authorization: "Bearer token" }) });
+    const body = JSON.parse(request.body) as { prompt_cache_retention?: string };
+
+    expect(body).not.toHaveProperty("prompt_cache_retention");
+  });
+
   const expandedRawBody = {
     model: "gpt-5.5",
     previous_response_id: "resp_1",