fix(agent-core): only send prompt_cache_key to official OpenAI endpoints

.changeset/prompt-cache-key-official-openai-only.md

@@ -0,0 +1,6 @@
+---
+"@moonshot-ai/agent-core-v2": patch
+"@moonshot-ai/kimi-code": patch
+---
+
+Stop sending the OpenAI prompt cache key to custom OpenAI-compatible endpoints, which reject the unknown field with a 400 error; the key is still sent to the official OpenAI API.

packages/agent-core-v2/src/human/llm/requester/bases/openai-base-url.ts

@@ -0,0 +1,11 @@
+export function isOfficialOpenAIBaseUrl(baseUrl: string | undefined): boolean {
+  if (baseUrl === undefined) {
+    return true;
+  }
+  try {
+    const hostname = new URL(baseUrl).hostname;
+    return hostname === 'api.openai.com' || hostname.endsWith('.api.openai.com');
+  } catch {
+    return false;
+  }
+}

packages/agent-core-v2/src/human/llm/requester/bases/openai-responses/format.ts

@@ -11,6 +11,7 @@ import { encodeReasoningEffortFallback } from '#/llm/thinking';
 import type { TokenUsage } from '#/llm/usage';
 
 import { isContextOverflowErrorCode, isOpenAIInsufficientQuotaCode } from '../openai/format';
+import { isOfficialOpenAIBaseUrl } from '../openai-base-url';
 import { lowerMessage, type ResponsesInputItem } from './lower';
 
 type RawObject = Record<string, unknown>;
@@ -365,7 +366,9 @@ function resolveRequestKwargs(input: FormatRequestInput): Record<string, unknown
   } = input;
   let kwargs: Record<string, unknown> = {};
   if (cacheKey !== undefined) {
-    kwargs = trait?.cacheKey?.(cacheKey, ctx) ?? { prompt_cache_key: cacheKey };
+    kwargs =
+      trait?.cacheKey?.(cacheKey, ctx) ??
+      (isOfficialOpenAIBaseUrl(ctx.model.baseUrl) ? { prompt_cache_key: cacheKey } : {});
   }
   if (thinking !== undefined) {
     kwargs = applyThinking(kwargs, thinking, trait, ctx, (t) =>

packages/agent-core-v2/src/human/llm/requester/bases/openai/format.ts

@@ -37,6 +37,7 @@ import type { TokenUsage } from '#/llm/usage';
 import { lowerMessage, type OpenAIWireMessage } from './lower';
 import { extractToolMedia } from './patterns';
 import { DEFAULT_REASONING_KEY, extractReasoning } from './reasoning-key';
+import { isOfficialOpenAIBaseUrl } from '../openai-base-url';
 
 function responseFormatToOpenAI(format: ResponseFormat): Record<string, unknown> {
   if (format.type === 'json_object') {
@@ -176,7 +177,9 @@ function resolveRequestKwargs(input: FormatRequestInput): ResolvedRequestKwargs
   } = input;
   let kwargs: Record<string, unknown> = {};
   if (cacheKey !== undefined) {
-    kwargs = trait?.cacheKey?.(cacheKey, ctx) ?? { prompt_cache_key: cacheKey };
+    kwargs =
+      trait?.cacheKey?.(cacheKey, ctx) ??
+      (isOfficialOpenAIBaseUrl(ctx.model.baseUrl) ? { prompt_cache_key: cacheKey } : {});
   }
   let preserveThinking = false;
   if (thinking !== undefined) {

packages/agent-core-v2/src/human/test/llm/cache-key.test.ts

@@ -11,7 +11,7 @@ const model: LlmModel = {
   provider: 'test',
   model: 'test-model',
   capability: UNKNOWN_CAPABILITY,
-  baseUrl: 'https://example.test/v1',
+  baseUrl: 'https://api.openai.com/v1',
 };
 const messages: readonly Message[] = [createUserMessage('hi')];
 
@@ -121,6 +121,20 @@ describe('openai requester cacheKey', () => {
     expect(client.body()['extra_body']).toEqual({ trace_id: 't1' });
   });
 
+  it('omits prompt_cache_key for third-party OpenAI endpoints', async () => {
+    const client = stubOpenAIClient(chatCompletionChunks);
+    const requester = createOpenAIRequester(undefined, { clientFactory: client.clientFactory });
+    await requester.generate(
+      {
+        model: { ...model, baseUrl: 'https://integrate.api.nvidia.com/v1' },
+        cacheKey: 'session-1',
+      },
+      { messages },
+      { signal: new AbortController().signal },
+    );
+    expect(client.body()['prompt_cache_key']).toBeUndefined();
+  });
+
   it('lets a trait override the cache key params', async () => {
     const client = stubOpenAIClient(chatCompletionChunks);
     const requester = createOpenAIRequester(