fix: inject thinking param for OpenAI-compatible APIs with reasoning_effort

src/api/providers/__tests__/openai.spec.ts

@@ -68,7 +68,7 @@ vi.mock("axios", () => ({
 	},
 }))
 
-import { OpenAiHandler, getOpenAiModels } from "../openai"
+import { OpenAiHandler, getOpenAiModels, createThinkingAwareFetch } from "../openai"
 import { ApiHandlerOptions } from "../../../shared/api"
 import { Anthropic } from "@anthropic-ai/sdk"
 import { openAiModelInfoSaneDefaults } from "@roo-code/types"
@@ -143,6 +143,7 @@ describe("OpenAiHandler", () => {
 				expect.objectContaining({
 					baseURL: "https://api.openai.com/v1",
 					apiKey: "test-api-key",
+					fetch: expect.any(Function),
 				}),
 			)
 		})
@@ -978,3 +979,90 @@ describe("getOpenAiModels", () => {
 		expect(result).toEqual(["gpt-4", "gpt-3.5-turbo"])
 	})
 })
+
+describe("createThinkingAwareFetch", () => {
+	const originalFetch = globalThis.fetch
+
+	afterEach(() => {
+		globalThis.fetch = originalFetch
+	})
+
+	it("should inject thinking: { type: 'enabled' } when reasoning_effort is present", async () => {
+		let capturedBody: string | undefined
+
+		globalThis.fetch = vi.fn(async (_input: RequestInfo | URL, init?: RequestInit) => {
+			capturedBody = init?.body as string
+			return new Response(JSON.stringify({}), { status: 200 })
+		}) as any
+
+		const wrappedFetch = createThinkingAwareFetch()
+		const body = JSON.stringify({ model: "some-model", reasoning_effort: "high" })
+		await wrappedFetch("https://example.com/v1/chat/completions", { method: "POST", body })
+
+		const parsed = JSON.parse(capturedBody!)
+		expect(parsed.thinking).toEqual({ type: "enabled" })
+		expect(parsed.reasoning_effort).toBe("high")
+	})
+
+	it("should not inject thinking when reasoning_effort is absent", async () => {
+		let capturedBody: string | undefined
+
+		globalThis.fetch = vi.fn(async (_input: RequestInfo | URL, init?: RequestInit) => {
+			capturedBody = init?.body as string
+			return new Response(JSON.stringify({}), { status: 200 })
+		}) as any
+
+		const wrappedFetch = createThinkingAwareFetch()
+		const body = JSON.stringify({ model: "some-model", messages: [] })
+		await wrappedFetch("https://example.com/v1/chat/completions", { method: "POST", body })
+
+		const parsed = JSON.parse(capturedBody!)
+		expect(parsed.thinking).toBeUndefined()
+	})
+
+	it("should not overwrite existing thinking parameter", async () => {
+		let capturedBody: string | undefined
+
+		globalThis.fetch = vi.fn(async (_input: RequestInfo | URL, init?: RequestInit) => {
+			capturedBody = init?.body as string
+			return new Response(JSON.stringify({}), { status: 200 })
+		}) as any
+
+		const wrappedFetch = createThinkingAwareFetch()
+		const body = JSON.stringify({
+			model: "some-model",
+			reasoning_effort: "high",
+			thinking: { type: "disabled", budget_tokens: 0 },
+		})
+		await wrappedFetch("https://example.com/v1/chat/completions", { method: "POST", body })
+
+		const parsed = JSON.parse(capturedBody!)
+		expect(parsed.thinking).toEqual({ type: "disabled", budget_tokens: 0 })
+	})
+
+	it("should pass through non-JSON bodies unchanged", async () => {
+		let capturedBody: string | undefined
+
+		globalThis.fetch = vi.fn(async (_input: RequestInfo | URL, init?: RequestInit) => {
+			capturedBody = init?.body as string
+			return new Response("ok", { status: 200 })
+		}) as any
+
+		const wrappedFetch = createThinkingAwareFetch()
+		const body = "not-json-body"
+		await wrappedFetch("https://example.com/v1/chat/completions", { method: "POST", body })
+
+		expect(capturedBody).toBe("not-json-body")
+	})
+
+	it("should pass through requests with no body", async () => {
+		globalThis.fetch = vi.fn(async () => {
+			return new Response("ok", { status: 200 })
+		}) as any
+
+		const wrappedFetch = createThinkingAwareFetch()
+		await wrappedFetch("https://example.com/v1/models")
+
+		expect(globalThis.fetch).toHaveBeenCalledWith("https://example.com/v1/models", undefined)
+	})
+})

src/api/providers/openai.ts

@@ -81,6 +81,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 				baseURL,
 				apiKey,
 				headers,
+				fetch: createThinkingAwareFetch(),
 			})
 			this.languageModelFactory = (modelId: string) => provider.chat(modelId)
 		}
@@ -373,6 +374,34 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 	}
 }
 
+/**
+ * Creates a fetch wrapper that injects `thinking: { type: "enabled" }` into
+ * request bodies that already contain `reasoning_effort`. This fixes
+ * OpenAI-compatible APIs (e.g. VolcEngine) that default `thinking` to
+ * `disabled` when it's not explicitly provided, causing a 400 error when
+ * combined with `reasoning_effort`.
+ *
+ * Standard OpenAI endpoints silently ignore unrecognised body fields,
+ * so this is safe for all OpenAI-compatible providers.
+ */
+/** @internal Exported for testing. */
+export function createThinkingAwareFetch(): typeof globalThis.fetch {
+	return async (input: RequestInfo | URL, init?: RequestInit) => {
+		if (init?.body && typeof init.body === "string") {
+			try {
+				const json = JSON.parse(init.body)
+				if (json.reasoning_effort && !json.thinking) {
+					json.thinking = { type: "enabled" }
+					init = { ...init, body: JSON.stringify(json) }
+				}
+			} catch {
+				// Not JSON – pass through untouched.
+			}
+		}
+		return globalThis.fetch(input, init)
+	}
+}
+
 export async function getOpenAiModels(baseUrl?: string, apiKey?: string, openAiHeaders?: Record<string, string>) {
 	try {
 		if (!baseUrl) {