copilot: support 'extra_body' for BYOK (top_k, chat_template_kwargs, …

extensions/copilot/src/extension/byok/common/byokProvider.ts

@@ -68,6 +68,7 @@ export interface BYOKModelCapabilities {
 	editTools?: EndpointEditToolName[];
 	requestHeaders?: Record<string, string>;
 	modelOptions?: IChatModelRequestOptions;
+	extraBody?: Record<string, Object>,
 	supportedEndpoints?: ModelSupportedEndpoint[];
 	zeroDataRetentionEnabled?: boolean;
 	supportsReasoningEffort?: string[];
@@ -172,6 +173,7 @@ export function resolveModelInfo(modelId: string, providerName: string, knownMod
 		supported_endpoints: knownModelInfo?.supportedEndpoints,
 		zeroDataRetentionEnabled: knownModelInfo?.zeroDataRetentionEnabled,
 		modelOptions: knownModelInfo?.modelOptions,
+		extraBody: knownModelInfo?.extraBody,
 		reasoningEffortFormat: knownModelInfo?.reasoningEffortFormat
 	};
 	if (knownModelInfo?.requestHeaders && Object.keys(knownModelInfo.requestHeaders).length > 0) {

extensions/copilot/src/extension/byok/common/test/byokProvider.spec.ts

@@ -102,6 +102,23 @@ describe('resolveModelInfo', () => {
 		});
 	});
 
+	it('propagates configured extra body properties into chat-endpoint inputs', () => {
+		const info = resolveModelInfo('m1', 'TestProvider', undefined, {
+			...baseCapabilities,
+			extraBody: {
+				chat_template_kwargs: {
+					enable_thinking: false,
+				},
+			},
+		});
+
+		expect(info.extraBody).toEqual({
+			chat_template_kwargs: {
+				enable_thinking: false,
+			},
+		});
+	});
+
 	it('honors an explicit contextWindow as the source of truth for the context window', () => {
 		// A model documented as: Context Length 1M, Max Output 384K. The user can now
 		// declare the real capability directly instead of back-computing maxInputTokens.

extensions/copilot/src/extension/byok/node/openAIEndpoint.ts

@@ -281,18 +281,21 @@ export class OpenAIEndpoint extends ChatEndpoint {
 				body.previous_response_id = undefined;
 			}
 			this._applyReasoningEffort(body, options);
+			this._applyExtraBody(body, options);
 			return this._applyConfiguredModelOptions(body, options);
 		} else if (this.useMessagesApi) {
 			// Delegate to base ChatEndpoint for Messages API dispatch
 			const body = super.createRequestBody(options);
 			this._applyReasoningEffort(body, options);
+			this._applyExtraBody(body, options);
 			return this._applyConfiguredModelOptions(body, options);
 		} else {
 			const body = createCapiRequestBody(options, this.model, this.getCompletionsCallback());
 			if (body.messages && isKimiFamily(this)) {
 				body.messages = normalizeKimiToolCallIds(body.messages);
 			}
 			this._applyReasoningEffort(body, options);
+			this._applyExtraBody(body, options);
 			return this._applyConfiguredModelOptions(body, options);
 		}
 	}
@@ -362,6 +365,24 @@ export class OpenAIEndpoint extends ChatEndpoint {
 		}
 	}
 
+	/**
+	 * Forwards optional JSON properties, similar to OpenAI's 'extra_body' API.
+	 * Typical parameters include: 'enable_thinking' when using QwenCloud, 'top_k' against a vLLM backend or
+	 * 'chat_template_kwargs' for SGLang, etc. Those properties are fully user provided and don't follow
+	 * any predefined model.
+	 */
+	private _applyExtraBody(body: IEndpointBody, options: ICreateEndpointBodyOptions): void {
+		const extraBody = this.modelMetadata.extraBody;
+		if (!extraBody) {
+			return;
+		}
+
+		for (const [key, value] of Object.entries(extraBody)) {
+			// The options can be arbitrary objects, e.g. 'chat_template_kwargs'
+			(body as Record<string, Object>)[key] = value;
+		}
+	}
+
 	override interceptBody(body: IEndpointBody | undefined): void {
 		super.interceptBody(body);
 		// TODO @lramos15 - We should do this for all models and not just here

extensions/copilot/src/extension/byok/vscode-node/azureProvider.ts

@@ -139,6 +139,7 @@ export class AzureBYOKModelProvider extends AbstractCustomOAIBYOKModelProvider {
 			thinking: modelConfiguration?.thinking,
 			streaming: modelConfiguration?.streaming,
 			requestHeaders: modelConfiguration?.requestHeaders,
+			extraBody: modelConfiguration?.extraBody,
 			editTools: model.capabilities?.editTools?.filter(isEndpointEditToolName),
 			zeroDataRetentionEnabled: modelConfiguration?.zeroDataRetentionEnabled
 		};

extensions/copilot/src/extension/byok/vscode-node/customEndpointProvider.ts

@@ -102,6 +102,7 @@ interface _CustomEndpointModelConfig {
 	editTools?: EndpointEditToolName[];
 	requestHeaders?: Record<string, string>;
 	modelOptions?: IChatModelRequestOptions;
+	extraBody?: Record<string, Object>,
 	zeroDataRetentionEnabled?: boolean;
 	supportsReasoningEffort?: string[];
 	reasoningEffortFormat?: 'chat-completions' | 'responses' | 'messages';
@@ -165,6 +166,7 @@ export class CustomEndpointBYOKModelProvider extends AbstractOpenAICompatibleLMP
 			streaming: modelConfiguration?.streaming,
 			requestHeaders: modelConfiguration?.requestHeaders,
 			modelOptions: modelConfiguration?.modelOptions,
+			extraBody: modelConfiguration?.extraBody,
 			zeroDataRetentionEnabled: modelConfiguration?.zeroDataRetentionEnabled,
 			supportsReasoningEffort: modelConfiguration?.supportsReasoningEffort,
 			reasoningEffortFormat: modelConfiguration?.reasoningEffortFormat

extensions/copilot/src/extension/byok/vscode-node/customOAIProvider.ts

@@ -64,6 +64,8 @@ interface _CustomOAIModelConfig {
 	streaming?: boolean;
 	editTools?: EndpointEditToolName[];
 	requestHeaders?: Record<string, string>;
+	/** This is called 'extra_body' in OpenAI api */
+	extraBody?: Record<string, Object>,
 	zeroDataRetentionEnabled?: boolean;
 	supportsReasoningEffort?: string[];
 	reasoningEffortFormat?: 'chat-completions' | 'responses' | 'messages';
@@ -149,6 +151,7 @@ export abstract class AbstractCustomOAIBYOKModelProvider extends AbstractOpenAIC
 			thinking: modelConfiguration?.thinking ?? false,
 			streaming: modelConfiguration?.streaming,
 			requestHeaders: modelConfiguration?.requestHeaders,
+			extraBody: modelConfiguration?.extraBody,
 			zeroDataRetentionEnabled: modelConfiguration?.zeroDataRetentionEnabled,
 			supportsReasoningEffort: modelConfiguration?.supportsReasoningEffort,
 			reasoningEffortFormat: modelConfiguration?.reasoningEffortFormat

extensions/copilot/src/extension/byok/vscode-node/test/customEndpointProvider.spec.ts

@@ -453,6 +453,66 @@ describe('CustomEndpointBYOKModelProvider', () => {
 			]);
 		});
 
+		it('applies configured extra body parameters to Responses and Messages API bodies', () => {
+			const results = [
+				{
+					supportedEndpoints: [ModelSupportedEndpoint.Responses],
+					url: 'https://api.example.com/v1/responses',
+				},
+				{
+					supportedEndpoints: [ModelSupportedEndpoint.Messages],
+					url: 'https://api.example.com/v1/messages',
+				},
+			].map(({ supportedEndpoints, url }) => {
+				const metadata: IChatModelInformation = {
+					...makeMetadata(supportedEndpoints),
+					extraBody: {
+						chat_template_kwargs: {
+							enable_thinking: false
+						}
+					},
+				};
+				const endpoint = instaService.createInstance(CustomEndpointOAIEndpoint,
+					metadata,
+					'test-api-key',
+					url);
+				const body = endpoint.createRequestBody({
+					debugName: 'test',
+					messages: [{
+						role: Raw.ChatRole.User,
+						content: [{ type: Raw.ChatCompletionContentPartKind.Text, text: 'Hello' }]
+					}],
+					requestId: `test-req-${endpoint.apiType}-model-options`,
+					postOptions: {
+						stream: true,
+					},
+					finishedCb: undefined,
+					location: ChatLocation.Other,
+				});
+
+				return {
+					apiType: endpoint.apiType,
+					chat_template_kwargs: (body as any).chat_template_kwargs,
+				};
+			});
+
+			expect(results).toEqual([
+				{
+					apiType: 'responses',
+					chat_template_kwargs: {
+						enable_thinking: false
+					}
+				},
+				{
+					apiType: 'messages',
+					chat_template_kwargs: {
+						enable_thinking: false
+					}
+				},
+			]);
+		});
+
+
 		it('replaces default Bearer with user-supplied Authorization header on Chat Completions endpoints', () => {
 			const metadata = makeMetadata(undefined);
 			metadata.requestHeaders = { 'Authorization': 'Bearer user-token' };

extensions/copilot/src/extension/prompt/vscode-node/requestLoggerImpl.ts

@@ -15,7 +15,6 @@ import { ILogService } from '../../../platform/log/common/logService';
 import { messageToMarkdown } from '../../../platform/log/common/messageStringify';
 import { ContextManagementResponse } from '../../../platform/networking/common/anthropic';
 import { IResponseDelta, isOpenAiFunctionTool } from '../../../platform/networking/common/fetch';
-import { IEndpointBody } from '../../../platform/networking/common/networking';
 import { CapturingToken } from '../../../platform/requestLogger/common/capturingToken';
 import { ChatRequestScheme, ILoggedElementInfo, ILoggedRequestInfo, ILoggedToolCall, LoggedInfo, LoggedInfoKind, LoggedRequest, LoggedRequestKind, resolveMarkdownContent } from '../../../platform/requestLogger/common/requestLogger';
 import { AbstractRequestLogger } from '../../../platform/requestLogger/node/requestLogger';
@@ -602,12 +601,26 @@ export class RequestLogger extends AbstractRequestLogger {
 		result.push(`# ${entry.debugName} - ${id}`);
 		result.push(``);
 
-		// Just some other options to track
-		// TODO Probably we should just extract every item on the body and format it as below, instead of doing this one-by-one
-		const otherOptions: Record<string, string | number | boolean> = {};
-		for (const opt of ['temperature', 'stream', 'store', 'reasoning_effort'] satisfies (keyof IEndpointBody)[]) {
-			if (entry.chatParams.body?.[opt] !== undefined) {
-				otherOptions[opt] = entry.chatParams.body[opt];
+		// Options that aren't already shown all end up in 'otherOptions' for debug purposes.
+		// This includes 'temperature', 'stream', 'store', 'reasoning_effort', everything that is in 'extraBody', etc.
+		const otherOptions: Record<string, string | number | boolean | object> = {};
+		if (entry.chatParams.body) {
+			for (const [optKey, optValue] of Object.entries(entry.chatParams.body)) {
+				if ([
+					'input',
+					'max_completion_tokens',
+					'max_output_tokens',
+					'max_tokens',
+					'messages',
+					'model',
+					'prediction',
+					'reasoning',
+					'tools',
+				].includes(optKey)) {
+					continue;
+				}
+
+				otherOptions[optKey] = optValue;
 			}
 		}
 

extensions/copilot/src/platform/endpoint/common/endpointProvider.ts

@@ -143,6 +143,7 @@ export type IChatModelInformation = IModelAPIResponse & {
 	urlOrRequestMetadata?: string | RequestMetadata;
 	requestHeaders?: Readonly<Record<string, string>>;
 	modelOptions?: Readonly<IChatModelRequestOptions>;
+	extraBody?: Readonly<Record<string, Object>>;
 	zeroDataRetentionEnabled?: boolean;
 	/**
 	 * BYOK-only override that forces the body shape used when forwarding the reasoning effort to the model.