diff --git a/README.md b/README.md
index 0f7ea4f..8f2d291 100644
--- a/README.md
+++ b/README.md
@@ -594,7 +594,18 @@ choco install ffmpeg       # Windows
 
    `timeoutMs` raises the SDK's timeout, but with `openai` 5 or newer a request still fails after about 5 minutes
    without response headers: those versions use Node's built-in `fetch`, whose own 300-second limit applies first.
-   `openai` 4 honours `timeoutMs` beyond that.
+   `openai` 4 honours `timeoutMs` beyond that. To wait longer on `openai` 5+, pass an undici dispatcher through
+   `fetchOptions` (`npm i undici`):
+
+   ```js
+   import { Agent } from 'undici';
+
+   engines.openai({
+     baseUrl: 'http://localhost:8881/v1',
+     timeoutMs: 1_800_000,
+     fetchOptions: { dispatcher: new Agent({ headersTimeout: 0, bodyTimeout: 0 }) },
+   })
+   ```
 
 2. **Record** — Playwright runs the demo script in a real browser. The `narration` fixture records timestamps for each `mark()` call. Video is captured at native resolution.
 
diff --git a/skills/argo-guide/references/tts-engines.md b/skills/argo-guide/references/tts-engines.md
index 7ef85fc..35ed04e 100644
--- a/skills/argo-guide/references/tts-engines.md
+++ b/skills/argo-guide/references/tts-engines.md
@@ -86,7 +86,7 @@ engines.openai({
 ```
 
 - With `baseUrl` set, `OPENAI_API_KEY` is never read — an OpenAI key is not leaked to another server. Pass `apiKey` if the server needs one.
-- With `openai` >= 5, requests still fail after ~300 s without response headers (Node `fetch`'s `headersTimeout`), regardless of `timeoutMs`. `openai` 4 honours longer timeouts.
+- With `openai` >= 5, requests still fail after ~300 s without response headers (Node `fetch`'s `headersTimeout`), regardless of `timeoutMs`. `openai` 4 honours longer timeouts. On 5+, lift the limit with `fetchOptions: { dispatcher: new Agent({ headersTimeout: 0, bodyTimeout: 0 }) }` (`import { Agent } from 'undici'`).
 - `baseUrl` is recorded in `<demo>.meta.json` but is not part of the clip cache key — clear `.argo/<demo>/clips` after switching servers.
 
 ---
diff --git a/src/tts/engines/openai.ts b/src/tts/engines/openai.ts
index 3ad641f..6af6a6b 100644
--- a/src/tts/engines/openai.ts
+++ b/src/tts/engines/openai.ts
@@ -15,9 +15,16 @@ export interface OpenAIEngineOptions {
   baseUrl?: string;
   /**
    * Request timeout in milliseconds, for slow local servers. Default: the OpenAI SDK's (10 minutes).
-   * With `openai` >= 5, Node's built-in `fetch` still gives up after ~300 s without response headers.
+   * With `openai` >= 5, Node's built-in `fetch` still gives up after ~300 s without response headers;
+   * lift that with `fetchOptions`.
    */
   timeoutMs?: number;
+  /**
+   * Extra options merged into every `fetch` call the SDK makes (`openai` >= 5). To wait longer than ~300 s,
+   * pass an undici dispatcher with a larger `headersTimeout`:
+   * `fetchOptions: { dispatcher: new Agent({ headersTimeout: 0, bodyTimeout: 0 }) }` (`import { Agent } from 'undici'`).
+   */
+  fetchOptions?: Record<string, unknown>;
 }
 
 export class OpenAIEngine implements TTSEngine {
@@ -26,6 +33,7 @@ export class OpenAIEngine implements TTSEngine {
   private instructions?: string;
   private baseUrl?: string;
   private timeoutMs?: number;
+  private fetchOptions?: Record<string, unknown>;
 
   constructor(options?: OpenAIEngineOptions) {
     this.apiKey = options?.apiKey ?? '';
@@ -33,6 +41,7 @@ export class OpenAIEngine implements TTSEngine {
     this.instructions = options?.instructions;
     this.baseUrl = options?.baseUrl;
     this.timeoutMs = options?.timeoutMs;
+    this.fetchOptions = options?.fetchOptions;
   }
 
   describe(): TTSEngineMetadata {
@@ -67,6 +76,7 @@ export class OpenAIEngine implements TTSEngine {
       apiKey: this.resolveApiKey(),
       ...(this.baseUrl ? { baseURL: this.baseUrl } : {}),
       ...(this.timeoutMs != null ? { timeout: this.timeoutMs } : {}),
+      ...(this.fetchOptions ? { fetchOptions: this.fetchOptions } : {}),
     });
 
     // Request raw PCM so we can build an exact WAV without ffmpeg pipe artifacts
diff --git a/tests/tts/openai.test.ts b/tests/tts/openai.test.ts
index 5e4c484..37d1758 100644
--- a/tests/tts/openai.test.ts
+++ b/tests/tts/openai.test.ts
@@ -50,6 +50,12 @@ describe('OpenAIEngine', () => {
     });
   });
 
+  it('passes fetchOptions through to the client', async () => {
+    const dispatcher = { tag: 'agent' };
+    await new OpenAIEngine({ baseUrl: 'http://localhost:8881/v1', fetchOptions: { dispatcher } }).generate('Hello', {});
+    expect(clientOptions.mock.calls[0][0].fetchOptions).toEqual({ dispatcher });
+  });
+
   it('never sends OPENAI_API_KEY to another server', async () => {
     process.env.OPENAI_API_KEY = 'sk-env';
     await new OpenAIEngine({ baseUrl: 'http://localhost:8881/v1' }).generate('Hello', {});