fix OpenAI-compatible request timeout

apps/vscode-e2e/package.json

@@ -18,18 +18,19 @@
 		"clean": "rimraf out .turbo"
 	},
 	"devDependencies": {
+		"@copilotkit/aimock": "1.35.0",
 		"@playwright/test": "1.62.1",
 		"@roo-code/config-eslint": "workspace:^",
 		"@roo-code/config-typescript": "workspace:^",
 		"@roo-code/types": "workspace:^",
-		"@copilotkit/aimock": "1.35.0",
 		"@types/mocha": "10.0.10",
 		"@types/node": "22.20.1",
 		"@types/vscode": "1.100.0",
 		"@vscode/test-electron": "2.5.2",
 		"dotenv-cli": "11.0.0",
 		"glob": "11.1.0",
 		"mocha": "11.2.2",
-		"rimraf": "6.0.1"
+		"rimraf": "6.0.1",
+		"undici": "^6.21.3"
 	}
 }

apps/vscode-e2e/src/suite/providers/deepseek-v4.test.ts

@@ -3,12 +3,20 @@ import * as assert from "assert"
 import * as fs from "fs/promises"
 import * as path from "path"
 import * as vscode from "vscode"
-
+import * as undici from "undici"
 import { RooCodeEventName, type ClineMessage } from "@roo-code/types"
 
 import { setDefaultSuiteTimeout } from "../test-utils"
 import { sleep, waitFor, waitUntilAborted } from "../utils"
 
+interface UndiciModule {
+	fetch: typeof fetch
+}
+
+interface UndiciRequestInit extends RequestInit {
+	dispatcher?: unknown
+}
+
 const DEEPSEEK_API_KEY = process.env.DEEPSEEK_API_KEY
 
 type DeepSeekModelId = "deepseek-v4-flash" | "deepseek-v4-pro"
@@ -71,10 +79,11 @@ function getRequestBody(init?: RequestInit):
 }
 
 function installDeepSeekRequestCapture(capture: CapturedDeepSeekRequest[], baseUrl: string): () => void {
-	const originalFetch = globalThis.fetch
+	const undiciModule = undici as unknown as UndiciModule
+	const originalFetch = undiciModule.fetch
 	const targetOrigin = new URL(baseUrl).origin
 
-	globalThis.fetch = async function (input: RequestInfo | URL, init?: RequestInit): Promise<Response> {
+	undiciModule.fetch = async function (input: RequestInfo | URL, init?: RequestInit): Promise<Response> {
 		const url = getRequestUrl(input)
 
 		if (isUrlWithOrigin(url, targetOrigin) && isChatCompletionsUrl(url)) {
@@ -106,11 +115,12 @@ function installDeepSeekRequestCapture(capture: CapturedDeepSeekRequest[], baseU
 			capture.push(request)
 		}
 
-		return originalFetch.call(globalThis, input, init as RequestInit)
-	} as typeof globalThis.fetch
+		// Передаём init как есть (включая dispatcher из OpenAiHandler) оригинальному undici.fetch
+		return originalFetch.call(undiciModule, input, init as UndiciRequestInit)
+	}
 
 	return () => {
-		globalThis.fetch = originalFetch
+		undiciModule.fetch = originalFetch
 	}
 }
 

src/api/providers/__tests__/kimi-code.spec.ts

@@ -10,6 +10,16 @@ const { mockGetAccessToken, mockForceRefreshAccessToken, mockGetModels } = vi.ho
 	mockGetModels: vi.fn(),
 }))
 
+vi.mock("undici", async (importOriginal) => {
+	const actual = await importOriginal<typeof import("undici")>()
+	return {
+		...actual,
+		fetch: vi.fn().mockImplementation(async (url: RequestInfo | URL, init?: RequestInit) => {
+			return globalThis.fetch(url, init)
+		}),
+	}
+})
+
 vi.mock("../../../integrations/kimi-code/oauth", () => ({
 	kimiCodeOAuthManager: {
 		getAccessToken: mockGetAccessToken,

src/api/providers/openai.ts

@@ -1,6 +1,7 @@
 import { Anthropic } from "@anthropic-ai/sdk"
 import OpenAI, { AzureOpenAI } from "openai"
 import axios from "axios"
+import { Agent, fetch as undiciFetch, Dispatcher } from "undici"
 
 import {
 	type ModelInfo,
@@ -52,34 +53,79 @@
 			...(this.options.openAiHeaders || {}),
 		}
 
+		function resolveTimeoutMs(configuredMs: number | undefined): number {
+			if (configuredMs === undefined || configuredMs === 0) {
+				return 0
+				// return 60 * 60 * 1000;
+			}
+			return configuredMs
+		}
+
+		const timeoutMs = resolveTimeoutMs(this.timeoutMs)
+
+		// VS Code bundles its own undici with a 5-minute `bodyTimeout` default.
+		// For streaming LLM requests, that default terminates the connection.
+		// We bypass the VS Code-bundled undici by injecting our own Agent-backed
+		// fetch into the OpenAI SDK.
+		const agent = new Agent({
+			headersTimeout: timeoutMs,
+			bodyTimeout: timeoutMs,
+			keepAliveTimeout: timeoutMs,
+			keepAliveMaxTimeout: timeoutMs,
+			connect: {
+				timeout: Math.min(timeoutMs, 60_000),
+			},
+		})
+
+		interface UndiciRequestInit extends RequestInit {
+			dispatcher?: Dispatcher
+		}
+
+		const customFetch: typeof fetch = (url, init) => {
+			const undiciInit = { ...init, dispatcher: agent } as UndiciRequestInit
+			// return globalThis.fetch(url, undiciInit as RequestInit)
+			const fetchImpl = undiciFetch as unknown as (
+				url: RequestInfo | URL,
+				init: UndiciRequestInit,
+			) => Promise<Response>
+
+			return fetchImpl(url, undiciInit)
+		}
+
+		const timeoutConfig = {
+			timeout: timeoutMs,
+		}
+
 		if (isAzureAiInference) {
-			// Azure AI Inference Service (e.g., for DeepSeek) uses a different path structure
 			this.client = new OpenAI({
 				baseURL,
 				apiKey,
 				defaultHeaders: headers,
 				defaultQuery: { "api-version": this.options.azureApiVersion || "2024-05-01-preview" },
-				timeout: this.timeoutMs,
+				...timeoutConfig,
 			})
 		} else if (isAzureOpenAi) {
 			// Azure API shape slightly differs from the core API shape:
 			// https://github.com/openai/openai-node?tab=readme-ov-file#microsoft-azure-openai
+
 			const azureBaseURL = `${baseURL.replace(/\/openai\/?$/i, "").replace(/\/$/, "")}/openai`
 			this.client = new AzureOpenAI({
 				baseURL: azureBaseURL,
 				apiKey,
 				apiVersion: this.options.azureApiVersion || azureOpenAiDefaultApiVersion,
 				defaultHeaders: headers,
-				timeout: this.timeoutMs,
+				...timeoutConfig,
 			})
 		} else {
 			this.client = new OpenAI({
 				baseURL,
 				apiKey,
 				defaultHeaders: headers,
-				timeout: this.timeoutMs,
+				...timeoutConfig,
 			})
 		}
+
+		;(this.client as unknown as { fetch: typeof fetch }).fetch = customFetch
 	}
 
 	override async *createMessage(