Enable persistent CoT for hidden models

extensions/copilot/package.json

@@ -4045,6 +4045,15 @@
               "onExp"
             ]
           },
+          "github.copilot.chat.responsesApi.persistentCoT.enabled": {
+            "type": "boolean",
+            "default": false,
+            "markdownDescription": "%github.copilot.config.responsesApi.persistentCoT.enabled%",
+            "tags": [
+              "experimental",
+              "onExp"
+            ]
+          },
           "github.copilot.chat.updated53CodexPrompt.enabled": {
             "type": "boolean",
             "default": true,

extensions/copilot/package.nls.json

@@ -345,6 +345,7 @@
 	"github.copilot.config.responsesApiReasoningSummary": "Sets the reasoning summary style used for the Responses API. Requires `#github.copilot.chat.useResponsesApi#`.",
 	"github.copilot.config.responsesApiContextManagement.enabled": "Enables context management for the Responses API. Requires `#github.copilot.chat.useResponsesApi#`.",
 	"github.copilot.config.responsesApi.promptCacheKey.enabled": "Enables prompt cache key being set for the Responses API.",
+	"github.copilot.config.responsesApi.persistentCoT.enabled": "Enables persistent chain of thought for supported Responses API models.",
 	"github.copilot.config.updated53CodexPrompt.enabled": "Enables the updated prompt for gpt-5.3-codex model.",
 	"github.copilot.config.claude47OpusPrompt.enabled": "Enables the updated system prompt tuned for the Claude Opus 4.7 model.",
 	"github.copilot.config.gpt55GetChangedFilesTool.enabled": "Enables the Get Changed Files tool for gpt-5.5 models.",

extensions/copilot/src/platform/configuration/common/configurationService.ts

@@ -968,6 +968,8 @@ export namespace ConfigKey {
 	export const ResponsesApiContextManagementEnabled = defineSetting<boolean>('chat.responsesApiContextManagement.enabled', ConfigType.ExperimentBased, false);
 	/** Enable client-side prompt_cache_key (conversationId:modelFamily) sent to Responses API */
 	export const ResponsesApiPromptCacheKeyEnabled = defineSetting<boolean>('chat.responsesApi.promptCacheKey.enabled', ConfigType.ExperimentBased, false);
+	/** Enable persistent chain of thought for supported Responses API model families */
+	export const ResponsesApiPersistentCoTEnabled = defineSetting<boolean>('chat.responsesApi.persistentCoT.enabled', ConfigType.ExperimentBased, false);
 	/** Enable updated prompt for 5.3Codex model */
 	export const Updated53CodexPromptEnabled = defineSetting<boolean>('chat.updated53CodexPrompt.enabled', ConfigType.ExperimentBased, true);
 	/** Enable updated prompt for Claude Opus 4.7 model */

extensions/copilot/src/platform/endpoint/node/responsesApi.ts

@@ -27,7 +27,7 @@ import { IChatWebSocketManager } from '../../networking/node/chatWebSocketManage
 import { IExperimentationService } from '../../telemetry/common/nullExperimentationService';
 import { ITelemetryService } from '../../telemetry/common/telemetry';
 import { TelemetryData } from '../../telemetry/common/telemetryData';
-import { getVerbosityForModelSync } from '../common/chatModelCapabilities';
+import { getVerbosityForModelSync, isHiddenModelM } from '../common/chatModelCapabilities';
 import { rawPartAsCompactionData } from '../common/compactionDataContainer';
 import { rawPartAsPhaseData } from '../common/phaseDataContainer';
 import { getIndexOfStatefulMarker, getStatefulMarkerAndIndex } from '../common/statefulMarkerContainer';
@@ -163,10 +163,13 @@ export function createResponsesRequestBody(accessor: ServicesAccessor, options:
 		? (effortFromSetting || options.modelCapabilities?.reasoningEffort || 'medium')
 		: undefined;
 	const summary = summaryConfig === 'off' || shouldDisableReasoningSummary ? undefined : summaryConfig;
-	if (effort || summary) {
+	const persistentCoTEnabled = configService.getExperimentBasedConfig(ConfigKey.ResponsesApiPersistentCoTEnabled, expService)
+		&& isHiddenModelM(endpoint);
+	if (effort || summary || persistentCoTEnabled) {
 		body.reasoning = {
 			...(effort ? { effort } : {}),
-			...(summary ? { summary } : {})
+			...(summary ? { summary } : {}),
+			...(persistentCoTEnabled ? { context: 'all_turns' } : {})
 		};
 	}
 

extensions/copilot/src/platform/endpoint/node/test/responsesApi.spec.ts

@@ -9,6 +9,7 @@ import { describe, expect, it } from 'vitest';
 import { TokenizerType } from '../../../../util/common/tokenizer';
 import { IInstantiationService } from '../../../../util/vs/platform/instantiation/common/instantiation';
 import { ChatLocation } from '../../../chat/common/commonTypes';
+import { ConfigKey, IConfigurationService } from '../../../configuration/common/configurationService';
 import { ILogService } from '../../../log/common/logService';
 import { isOpenAIContextManagementResponse } from '../../../networking/common/fetch';
 import { IChatEndpoint, ICreateEndpointBodyOptions } from '../../../networking/common/networking';
@@ -368,6 +369,59 @@ describe('responseApiInputToRawMessagesForLogging', () => {
 });
 
 describe('createResponsesRequestBody', () => {
+	it('enables persistent CoT on initial requests for hidden model M when the experiment is enabled', () => {
+		const services = createPlatformServices();
+		const accessor = services.createTestingAccessor();
+		const instantiationService = accessor.get(IInstantiationService);
+		accessor.get(IConfigurationService).setConfig(ConfigKey.ResponsesApiPersistentCoTEnabled, true);
+		const endpoint = { ...testEndpoint, family: 'ember-alpha', supportsReasoningEffort: ['low', 'medium', 'high'] };
+
+		const body = instantiationService.invokeFunction(servicesAccessor => createResponsesRequestBody(servicesAccessor, createRequestOptions([], false), endpoint.model, endpoint));
+
+		expect(body.reasoning).toEqual({ effort: 'medium', summary: 'detailed', context: 'all_turns' });
+
+		accessor.dispose();
+		services.dispose();
+	});
+
+	it('does not enable persistent CoT when the experiment is disabled or the family is unsupported', () => {
+		const services = createPlatformServices();
+		const accessor = services.createTestingAccessor();
+		const instantiationService = accessor.get(IInstantiationService);
+		const emberEndpoint = { ...testEndpoint, family: 'ember-alpha' };
+		const unsupportedEndpoint = { ...testEndpoint, model: 'ember-alpha', family: 'other-family' };
+
+		const disabledBody = instantiationService.invokeFunction(servicesAccessor => createResponsesRequestBody(servicesAccessor, createRequestOptions([], false), emberEndpoint.model, emberEndpoint));
+		accessor.get(IConfigurationService).setConfig(ConfigKey.ResponsesApiPersistentCoTEnabled, true);
+		const unsupportedBody = instantiationService.invokeFunction(servicesAccessor => createResponsesRequestBody(servicesAccessor, createRequestOptions([], false), unsupportedEndpoint.model, unsupportedEndpoint));
+
+		expect(disabledBody.reasoning?.context).toBeUndefined();
+		expect(unsupportedBody.reasoning?.context).toBeUndefined();
+
+		accessor.dispose();
+		services.dispose();
+	});
+
+	it('keeps persistent CoT enabled when continuing from a previous response', () => {
+		const services = createPlatformServices();
+		const accessor = services.createTestingAccessor();
+		const instantiationService = accessor.get(IInstantiationService);
+		accessor.get(IConfigurationService).setConfig(ConfigKey.ResponsesApiPersistentCoTEnabled, true);
+		const endpoint = { ...testEndpoint, family: 'ember-alpha' };
+		const messages: Raw.ChatMessage[] = [
+			createStatefulMarkerMessage(endpoint.model, 'resp-prev'),
+			{ role: Raw.ChatRole.User, content: [{ type: Raw.ChatCompletionContentPartKind.Text, text: 'continue' }] },
+		];
+
+		const body = instantiationService.invokeFunction(servicesAccessor => createResponsesRequestBody(servicesAccessor, createRequestOptions(messages, false), endpoint.model, endpoint));
+
+		expect(body.previous_response_id).toBe('resp-prev');
+		expect(body.reasoning?.context).toBe('all_turns');
+
+		accessor.dispose();
+		services.dispose();
+	});
+
 	it('extracts compaction threshold from request body context management', () => {
 		expect(getResponsesApiCompactionThresholdFromBody({
 			context_management: [{

extensions/copilot/src/platform/networking/common/networking.ts

@@ -75,7 +75,7 @@ export interface IEndpointBody {
 	prediction?: Prediction;
 	messages?: any[];
 	n?: number;
-	reasoning?: { effort?: string; summary?: string };
+	reasoning?: { effort?: string; summary?: string; context?: 'current_turn' | 'all_turns' };
 	tool_choice?: OptionalChatRequestParams['tool_choice'] | { type: 'function'; name: string } | string;
 	top_logprobs?: number;
 	intent?: boolean;