diff --git a/electron/services/agent/cache.ts b/electron/services/agent/cache.ts
index 6d5359f3..b9a3135c 100644
--- a/electron/services/agent/cache.ts
+++ b/electron/services/agent/cache.ts
@@ -2,6 +2,7 @@ import { createHash } from 'crypto'
 import type { ProviderOptions, SystemModelMessage } from '@ai-sdk/provider-utils'
 import type { ToolSet } from 'ai'
 import { isArkBaseURL } from './arkContextFetch'
+import { isNvidiaInferenceBaseURL } from './promptCacheCompat'
 import type { AgentReasoningEffort, AgentRunInput } from './types'
 
 export interface AgentPromptParts {
@@ -146,7 +147,10 @@ export function buildProviderOptions(input: AgentRunInput, promptCacheKey: strin
     } else {
       // @ai-sdk/openai-compatible 只校验 camelCase 标准项；厂商扩展字段要用请求体原名透传。
       // 这里恢复 AI SDK 6 时代 OpenAI-compatible 服务常用的 prompt_cache_key 行为。
-      option.prompt_cache_key = promptCacheKey
+      // 英伟达推理接口（nvidia.com）不认该字段，命中时跳过注入，避免 400。见 issue #353。
+      if (!isNvidiaInferenceBaseURL(input.providerConfig.baseURL)) {
+        option.prompt_cache_key = promptCacheKey
+      }
     }
     if (Object.keys(option).length > 0) {
       const keys = new Set(['openai'])
diff --git a/electron/services/agent/promptCacheCompat.ts b/electron/services/agent/promptCacheCompat.ts
new file mode 100644
index 00000000..b80e5733
--- /dev/null
+++ b/electron/services/agent/promptCacheCompat.ts
@@ -0,0 +1,34 @@
+/**
+ * prompt cache 兼容性辅助：集中处理「哪些端点不支持 `prompt_cache_key`」。
+ *
+ * 背景：英伟达的 OpenAI 兼容推理接口（integrate.api.nvidia.com / api.nvidia.com）
+ * 不认 OpenAI 的 `prompt_cache_key` 扩展字段，带上会直接 400
+ * （Unsupported parameter(s): prompt_cache_key，见 issue #353）。
+ *
+ * 该字段会从两条路径进入请求体，必须同时拦截：
+ *  1. provider.ts 的 transformRequestBody（直接写请求体）
+ *  2. cache.ts 的 buildProviderOptions（经 providerOptions 透传）
+ * 本模块是唯一真相来源，两处都 import 它，避免逻辑漂移。
+ *
+ * 零依赖：可被单元测试单独导入，无需安装 Electron 依赖树。
+ */
+
+/** 判断 baseURL 是否属于英伟达推理接口。 */
+export function isNvidiaInferenceBaseURL(baseURL?: string): boolean {
+  return !!baseURL && /(^|\.)nvidia\.com(\/|$)/i.test(baseURL)
+}
+
+/**
+ * 仅对支持 prompt cache 的 OpenAI 兼容端点注入 `prompt_cache_key`；
+ * 英伟达推理接口不支持该字段，命中时跳过注入，避免 400。
+ * 纯函数，导出供单元测试。
+ */
+export function injectOpenAICompatiblePromptCacheKey(
+  args: Record<string, any>,
+  promptCacheKey?: string,
+  baseURL?: string,
+): Record<string, any> {
+  if (!promptCacheKey || args.prompt_cache_key) return args
+  if (isNvidiaInferenceBaseURL(baseURL)) return args
+  return { ...args, prompt_cache_key: promptCacheKey }
+}
diff --git a/electron/services/agent/provider.ts b/electron/services/agent/provider.ts
index 3dd7f97f..bf06d52b 100644
--- a/electron/services/agent/provider.ts
+++ b/electron/services/agent/provider.ts
@@ -14,6 +14,7 @@ import { CODEX_SUBSCRIPTION_DUMMY_API_KEY, createCodexSubscriptionFetch, getCode
 import { withOpenAICompatibleStreamSanitizer } from '../ai/openaiCompatibleStreamSanitizer'
 import { withGoogleExplicitCache } from './googleCacheFetch'
 import { isArkBaseURL, withArkContextCache } from './arkContextFetch'
+import { injectOpenAICompatiblePromptCacheKey } from './promptCacheCompat'
 import type { AgentProviderConfig } from './types'
 
 export type AgentLanguageModelOptions = {
@@ -100,11 +101,6 @@ function withAnthropicSanitizer(baseFetch: typeof globalThis.fetch | undefined):
   }) as typeof globalThis.fetch
 }
 
-function injectOpenAICompatiblePromptCacheKey(args: Record<string, any>, promptCacheKey?: string): Record<string, any> {
-  if (!promptCacheKey || args.prompt_cache_key) return args
-  return { ...args, prompt_cache_key: promptCacheKey }
-}
-
 export function createLanguageModel(config: AgentProviderConfig, options: AgentLanguageModelOptions = {}): LanguageModel {
   const { providerKind, name, apiKey, baseURL, model, headers, proxyUrl } = config
   const fetch = createProxyFetch(proxyUrl)
@@ -141,7 +137,7 @@ export function createLanguageModel(config: AgentProviderConfig, options: AgentL
     includeUsage: true,
     // 火山方舟端点：system 前缀自动走 context 缓存，见 arkContextFetch.ts
     fetch: withOpenAICompatibleStreamSanitizer(compatibleFetch),
-    transformRequestBody: (args) => injectOpenAICompatiblePromptCacheKey(args, options.promptCacheKey),
+    transformRequestBody: (args) => injectOpenAICompatiblePromptCacheKey(args, options.promptCacheKey, baseURL),
   }).chatModel(model)
 }
 
diff --git a/scripts/test-nvidia-prompt-cache.ts b/scripts/test-nvidia-prompt-cache.ts
new file mode 100644
index 00000000..537daf1a
--- /dev/null
+++ b/scripts/test-nvidia-prompt-cache.ts
@@ -0,0 +1,33 @@
+import assert from 'node:assert/strict'
+import { injectOpenAICompatiblePromptCacheKey, isNvidiaInferenceBaseURL } from '../electron/services/agent/promptCacheCompat.ts'
+
+// ---- isNvidiaInferenceBaseURL ----
+assert.equal(isNvidiaInferenceBaseURL('https://integrate.api.nvidia.com/v1'), true, 'integrate.api.nvidia.com/v1 应判为英伟达')
+assert.equal(isNvidiaInferenceBaseURL('https://api.nvidia.com/v1'), true, 'api.nvidia.com/v1 应判为英伟达')
+assert.equal(isNvidiaInferenceBaseURL('https://integrate.api.nvidia.com'), true, '无路径的 integrate.api.nvidia.com 也应判为英伟达')
+assert.equal(isNvidiaInferenceBaseURL('https://api.openai.com/v1'), false, 'OpenAI 不应判为英伟达')
+assert.equal(isNvidiaInferenceBaseURL('https://api.deepseek.com'), false, 'DeepSeek 不应判为英伟达')
+assert.equal(isNvidiaInferenceBaseURL('https://ark.cn/v1'), false, '火山方舟不应被误判为英伟达（方舟需保留 prompt_cache_key）')
+assert.equal(isNvidiaInferenceBaseURL(undefined), false, '无 baseURL 不应判为英伟达')
+
+// ---- injectOpenAICompatiblePromptCacheKey：NVIDIA 必须跳过，其它端点保留 ----
+const baseArgs = { model: 'x', messages: [] }
+
+const nvidiaResult = injectOpenAICompatiblePromptCacheKey(baseArgs, 'cache-key', 'https://integrate.api.nvidia.com/v1')
+assert.equal('prompt_cache_key' in nvidiaResult, false, 'NVIDIA 请求体不得含 prompt_cache_key')
+assert.equal(nvidiaResult, baseArgs, 'NVIDIA 命中 skip 分支时应原样返回（不新增字段、不复制）')
+
+const openaiResult = injectOpenAICompatiblePromptCacheKey(baseArgs, 'cache-key', 'https://api.openai.com/v1')
+assert.equal(openaiResult.prompt_cache_key, 'cache-key', 'OpenAI 应保留 prompt_cache_key')
+
+const deepseekResult = injectOpenAICompatiblePromptCacheKey(baseArgs, 'cache-key', 'https://api.deepseek.com')
+assert.equal(deepseekResult.prompt_cache_key, 'cache-key', 'DeepSeek 应保留 prompt_cache_key')
+
+const arkResult = injectOpenAICompatiblePromptCacheKey(baseArgs, 'cache-key', 'https://ark.cn/v1')
+assert.equal(arkResult.prompt_cache_key, 'cache-key', '火山方舟应保留 prompt_cache_key（context 缓存依赖它）')
+
+// 未提供 key 时原样返回（不新增字段）
+assert.equal(injectOpenAICompatiblePromptCacheKey(baseArgs, undefined, 'https://integrate.api.nvidia.com/v1'), baseArgs, '无 key 时原样返回')
+assert.equal(injectOpenAICompatiblePromptCacheKey(baseArgs, undefined, 'https://api.openai.com/v1'), baseArgs, '无 key 时原样返回')
+
+console.log('nvidia prompt_cache_key compat tests passed')