@vscode/chat-lib 0.4.1-9 → 0.5.1-0

package/dist/src/_internal/extension/prompt/node/chatMLFetcher.js

@@ -12,6 +12,7 @@ var __decorate = (this && this.__decorate) || function (decorators, target, key,
 var __param = (this && this.__param) || function (paramIndex, decorator) {
     return function (target, key) { decorator(target, key, paramIndex); }
 };
+var ChatMLFetcherImpl_1;
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.ChatMLFetcherImpl = exports.AbstractChatMLFetcher = void 0;
 exports.createTelemetryData = createTelemetryData;
@@ -35,8 +36,11 @@ const fetcherService_1 = require("../../../platform/networking/common/fetcherSer
 const networking_1 = require("../../../platform/networking/common/networking");
 const openai_1 = require("../../../platform/networking/common/openai");
 const chatStream_1 = require("../../../platform/networking/node/chatStream");
+const chatWebSocketManager_1 = require("../../../platform/networking/node/chatWebSocketManager");
 const stream_1 = require("../../../platform/networking/node/stream");
 const fetch_2 = require("../../../platform/openai/node/fetch");
+const index_1 = require("../../../platform/otel/common/index");
+const otelService_1 = require("../../../platform/otel/common/otelService");
 const requestLogger_1 = require("../../../platform/requestLogger/node/requestLogger");
 const nullExperimentationService_1 = require("../../../platform/telemetry/common/nullExperimentationService");
 const telemetry_1 = require("../../../platform/telemetry/common/telemetry");
@@ -83,7 +87,9 @@ class AbstractChatMLFetcher extends lifecycle_1.Disposable {
 }
 exports.AbstractChatMLFetcher = AbstractChatMLFetcher;
 let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
-    constructor(_fetcherService, _telemetryService, _requestLogger, _logService, _authenticationService, _interactionService, _chatQuotaService, _capiClientService, options, _configurationService, _experimentationService, _powerService, _instantiationService) {
+    static { ChatMLFetcherImpl_1 = this; }
+    static { this._maxConsecutiveWebSocketFallbacks = 3; }
+    constructor(_fetcherService, _telemetryService, _requestLogger, _logService, _authenticationService, _interactionService, _chatQuotaService, _capiClientService, options, _configurationService, _experimentationService, _powerService, _instantiationService, _webSocketManager, _otelService) {
         super(options);
         this._fetcherService = _fetcherService;
         this._telemetryService = _telemetryService;
@@ -97,23 +103,36 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
         this._experimentationService = _experimentationService;
         this._powerService = _powerService;
         this._instantiationService = _instantiationService;
+        this._webSocketManager = _webSocketManager;
+        this._otelService = _otelService;
         /**
          * Delays (in ms) between connectivity check attempts before retrying a failed request.
          * Configurable for testing purposes.
          */
         this.connectivityCheckDelays = [1000, 10000, 10000];
+        /**
+         * Tracks consecutive WebSocket request failures where the HTTP retry succeeded.
+         * After {@link _maxConsecutiveWebSocketFallbacks} such failures, WebSocket requests are disabled entirely.
+         */
+        this._consecutiveWebSocketRetryFallbacks = 0;
     }
     /**
      * Note: the returned array of strings may be less than `n` (e.g., in case there were errors during streaming)
      */
     async fetchMany(opts, token) {
-        let { debugName, endpoint: chatEndpoint, finishedCb, location, messages, requestOptions, source, telemetryProperties, userInitiatedRequest, requestKindOptions } = opts;
+        let { debugName, endpoint: chatEndpoint, finishedCb, location, messages, requestOptions, source, telemetryProperties, userInitiatedRequest, requestKindOptions, conversationId, turnId, useWebSocket, ignoreStatefulMarker } = opts;
+        if (useWebSocket && this._consecutiveWebSocketRetryFallbacks >= ChatMLFetcherImpl_1._maxConsecutiveWebSocketFallbacks) {
+            this._logService.debug(`[ChatWebSocketManager] Disabling WebSocket for request due to ${this._consecutiveWebSocketRetryFallbacks} consecutive WebSocket failures with successful HTTP fallback.`);
+            useWebSocket = false;
+            ignoreStatefulMarker = true;
+        }
         if (!telemetryProperties) {
             telemetryProperties = {};
         }
         if (!telemetryProperties.messageSource) {
             telemetryProperties.messageSource = debugName;
         }
+        const transport = useWebSocket ? 'websocket' : 'http';
         // TODO @lramos15 telemetry should not drive request ids
         const ourRequestId = telemetryProperties.requestId ?? telemetryProperties.messageId ?? (0, uuid_1.generateUuid)();
         const maxResponseTokens = chatEndpoint.maxOutputTokens;
@@ -127,6 +146,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
         const postOptions = this.preparePostOptions(requestOptions);
         const requestBody = chatEndpoint.createRequestBody({
             ...opts,
+            ignoreStatefulMarker,
             requestId: ourRequestId,
             postOptions
         });
@@ -141,7 +161,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             ourRequestId,
             location: opts.location,
             body: requestBody,
-            ignoreStatefulMarker: opts.ignoreStatefulMarker,
+            ignoreStatefulMarker,
             isConversationRequest: opts.isConversationRequest,
             customMetadata: opts.customMetadata
         });
@@ -155,6 +175,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
         let actualStatusCode;
         let suspendEventSeen;
         let resumeEventSeen;
+        let otelInferenceSpan;
         try {
             let response;
             const payloadValidationResult = isValidChatPayload(opts.messages, postOptions, chatEndpoint, this._configurationService, this._experimentationService);
@@ -169,13 +190,38 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             else {
                 const copilotToken = await this._authenticationService.getCopilotToken();
                 usernameToScrub = copilotToken.username;
-                const fetchResult = await this._fetchAndStreamChat(chatEndpoint, requestBody, baseTelemetry, streamRecorder.callback, requestOptions.secretKey, copilotToken, opts.location, ourRequestId, postOptions.n, token, userInitiatedRequest, telemetryProperties, opts.useFetcher, canRetryOnce, requestKindOptions);
+                const fetchResult = await this._fetchAndStreamChat(chatEndpoint, requestBody, baseTelemetry, streamRecorder.callback, requestOptions.secretKey, copilotToken, opts.location, ourRequestId, postOptions.n, token, userInitiatedRequest, useWebSocket, turnId, conversationId, telemetryProperties, opts.useFetcher, canRetryOnce, requestKindOptions);
                 response = fetchResult.result;
                 actualFetcher = fetchResult.fetcher;
                 actualBytesReceived = fetchResult.bytesReceived;
                 actualStatusCode = fetchResult.statusCode;
                 suspendEventSeen = fetchResult.suspendEventSeen;
                 resumeEventSeen = fetchResult.resumeEventSeen;
+                otelInferenceSpan = fetchResult.otelSpan;
+                // Tag span with debug name so orphaned spans (title, progressMessages, etc.) are identifiable
+                otelInferenceSpan?.setAttribute(index_1.GenAiAttr.AGENT_NAME, debugName);
+                // Extract and set structured prompt sections for the debug panel
+                if (otelInferenceSpan) {
+                    const capiMessages = requestBody.messages;
+                    // User request: last user-role message
+                    const userMessages = capiMessages?.filter(m => m.role === 'user');
+                    const lastUserMsg = userMessages?.[userMessages.length - 1];
+                    if (lastUserMsg?.content) {
+                        otelInferenceSpan.setAttribute(index_1.CopilotChatAttr.USER_REQUEST, lastUserMsg.content);
+                    }
+                    // System instructions (first system message) — raw text for debug panel sections
+                    const systemMsg = capiMessages?.find(m => m.role === 'system');
+                    if (systemMsg?.content) {
+                        otelInferenceSpan.setAttribute(index_1.GenAiAttr.SYSTEM_INSTRUCTIONS, systemMsg.content);
+                    }
+                }
+                // Always capture full request content for the debug panel
+                if (otelInferenceSpan) {
+                    const capiMessages = requestBody.messages;
+                    if (capiMessages) {
+                        otelInferenceSpan.setAttribute(index_1.GenAiAttr.INPUT_MESSAGES, (0, index_1.truncateForOTel)(JSON.stringify((0, index_1.toInputMessages)(capiMessages))));
+                    }
+                }
                 tokenCount = await chatEndpoint.acquireTokenizer().countMessagesTokens(messages);
                 const extensionId = source?.extensionId ?? constants_1.EXTENSION_ID;
                 this._onDidMakeChatMLRequest.fire({
@@ -189,7 +235,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             pendingLoggedChatRequest?.markTimeToFirstToken(timeToFirstToken);
             switch (response.type) {
                 case fetch_2.FetchResponseKind.Success: {
-                    const result = await this.processSuccessfulResponse(response, messages, requestBody, ourRequestId, maxResponseTokens, tokenCount, timeToFirstToken, streamRecorder, baseTelemetry, chatEndpoint, userInitiatedRequest, actualFetcher, actualBytesReceived, suspendEventSeen, resumeEventSeen);
+                    const result = await this.processSuccessfulResponse(response, messages, requestBody, ourRequestId, maxResponseTokens, tokenCount, timeToFirstToken, streamRecorder, baseTelemetry, chatEndpoint, userInitiatedRequest, transport, actualFetcher, actualBytesReceived, suspendEventSeen, resumeEventSeen);
                     // Handle FilteredRetry case with augmented messages
                     if (result.type === commonTypes_1.ChatFetchResponseType.FilteredRetry) {
                         if (opts.enableRetryOnFilter) {
@@ -237,6 +283,76 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                         };
                     }
                     pendingLoggedChatRequest?.resolve(result, streamRecorder.deltas);
+                    // Record OTel token usage metrics if available
+                    if (result.type === commonTypes_1.ChatFetchResponseType.Success && result.usage) {
+                        const metricAttrs = {
+                            operationName: index_1.GenAiOperationName.CHAT,
+                            providerName: index_1.GenAiProviderName.GITHUB,
+                            requestModel: chatEndpoint.model,
+                            responseModel: result.resolvedModel,
+                        };
+                        if (result.usage.prompt_tokens) {
+                            index_1.GenAiMetrics.recordTokenUsage(this._otelService, result.usage.prompt_tokens, 'input', metricAttrs);
+                        }
+                        if (result.usage.completion_tokens) {
+                            index_1.GenAiMetrics.recordTokenUsage(this._otelService, result.usage.completion_tokens, 'output', metricAttrs);
+                        }
+                        // Set token usage and response details on the chat span before ending it
+                        otelInferenceSpan?.setAttributes({
+                            [index_1.GenAiAttr.USAGE_INPUT_TOKENS]: result.usage.prompt_tokens ?? 0,
+                            [index_1.GenAiAttr.USAGE_OUTPUT_TOKENS]: result.usage.completion_tokens ?? 0,
+                            [index_1.GenAiAttr.RESPONSE_MODEL]: result.resolvedModel ?? chatEndpoint.model,
+                            [index_1.GenAiAttr.RESPONSE_ID]: result.requestId,
+                            [index_1.GenAiAttr.RESPONSE_FINISH_REASONS]: ['stop'],
+                            ...(result.usage.prompt_tokens_details?.cached_tokens
+                                ? { [index_1.GenAiAttr.USAGE_CACHE_READ_INPUT_TOKENS]: result.usage.prompt_tokens_details.cached_tokens }
+                                : {}),
+                            [index_1.CopilotChatAttr.TIME_TO_FIRST_TOKEN]: timeToFirstToken,
+                            ...(result.serverRequestId ? { [index_1.CopilotChatAttr.SERVER_REQUEST_ID]: result.serverRequestId } : {}),
+                            ...(result.usage.completion_tokens_details?.reasoning_tokens
+                                ? { [index_1.GenAiAttr.USAGE_REASONING_TOKENS]: result.usage.completion_tokens_details.reasoning_tokens }
+                                : {}),
+                        });
+                    }
+                    // Always capture response content for the debug panel
+                    if (otelInferenceSpan && result.type === commonTypes_1.ChatFetchResponseType.Success) {
+                        const responseText = streamRecorder.deltas.map(d => d.text).join('');
+                        const toolCalls = streamRecorder.deltas
+                            .filter(d => d.copilotToolCalls?.length)
+                            .flatMap(d => d.copilotToolCalls.map(tc => ({
+                            type: 'tool_call', id: tc.id, name: tc.name, arguments: tc.arguments
+                        })));
+                        const parts = [];
+                        if (responseText) {
+                            parts.push({ type: 'text', content: responseText });
+                        }
+                        parts.push(...toolCalls);
+                        if (parts.length > 0) {
+                            otelInferenceSpan.setAttribute(index_1.GenAiAttr.OUTPUT_MESSAGES, (0, index_1.truncateForOTel)(JSON.stringify([{ role: 'assistant', parts }])));
+                        }
+                    }
+                    // Emit OTel inference details event BEFORE ending the span
+                    // so the log record inherits the active trace context
+                    (0, index_1.emitInferenceDetailsEvent)(this._otelService, {
+                        model: chatEndpoint.model,
+                        temperature: requestOptions?.temperature,
+                        maxTokens: requestOptions?.max_tokens,
+                    }, result.type === commonTypes_1.ChatFetchResponseType.Success ? {
+                        id: result.requestId,
+                        model: result.resolvedModel,
+                        finishReasons: ['stop'],
+                        inputTokens: result.usage?.prompt_tokens,
+                        outputTokens: result.usage?.completion_tokens,
+                    } : undefined);
+                    otelInferenceSpan?.end();
+                    otelInferenceSpan = undefined;
+                    // Record OTel time-to-first-token metric
+                    if (timeToFirstToken > 0) {
+                        index_1.GenAiMetrics.recordTimeToFirstToken(this._otelService, chatEndpoint.model, timeToFirstToken / 1000);
+                    }
+                    if (useWebSocket && result.type === commonTypes_1.ChatFetchResponseType.Success) {
+                        this._consecutiveWebSocketRetryFallbacks = 0;
+                    }
                     return result;
                 }
                 case fetch_2.FetchResponseKind.Canceled:
@@ -245,6 +361,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                         requestId: ourRequestId,
                         model: chatEndpoint.model,
                         apiType: chatEndpoint.apiType,
+                        transport,
                         associatedRequestId: telemetryProperties.associatedRequestId,
                         retryAfterError: telemetryProperties.retryAfterError,
                         retryAfterErrorGitHubRequestId: telemetryProperties.retryAfterErrorGitHubRequestId,
@@ -268,6 +385,13 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                         issuedTime: baseTelemetry.issuedTime,
                     });
                     pendingLoggedChatRequest?.resolveWithCancelation();
+                    // Set canceled status on OTel span
+                    otelInferenceSpan?.setAttributes({
+                        [index_1.GenAiAttr.RESPONSE_FINISH_REASONS]: ['cancelled'],
+                        [index_1.CopilotChatAttr.CANCELED]: true,
+                    });
+                    otelInferenceSpan?.end();
+                    otelInferenceSpan = undefined;
                     return this.processCanceledResponse(response, ourRequestId, streamRecorder, telemetryProperties);
                 case fetch_2.FetchResponseKind.Failed: {
                     const processed = this.processFailedResponse(response, ourRequestId);
@@ -276,7 +400,9 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                     const statusCodesToRetry = retryServerErrorStatusCodes
                         .split(',')
                         .map(s => parseInt(s.trim(), 10));
-                    if (enableRetryOnError && actualStatusCode !== undefined && statusCodesToRetry.includes(actualStatusCode)) {
+                    const retryAfterServerError = enableRetryOnError && actualStatusCode !== undefined && statusCodesToRetry.includes(actualStatusCode);
+                    const retryWithoutWebSocket = enableRetryOnError && useWebSocket;
+                    if (retryAfterServerError || retryWithoutWebSocket) {
                         const { retryResult } = await this._retryAfterError({
                             opts,
                             processed,
@@ -285,6 +411,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                             tokenCount,
                             maxResponseTokens,
                             timeToError: timeToFirstToken,
+                            transport,
                             actualFetcher,
                             bytesReceived: actualBytesReceived,
                             baseTelemetry,
@@ -310,6 +437,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                         maxResponseTokens,
                         timeToFirstToken,
                         isVisionRequest: this.filterImageMessages(messages),
+                        transport,
                         fetcher: actualFetcher,
                         bytesReceived: actualBytesReceived,
                         issuedTime: baseTelemetry.issuedTime,
@@ -323,6 +451,14 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             }
         }
         catch (err) {
+            // End OTel inference span on error if not already ended
+            if (otelInferenceSpan) {
+                otelInferenceSpan.setStatus(2 /* SpanStatusCode.ERROR */, err instanceof Error ? err.message : String(err));
+                otelInferenceSpan.setAttribute(index_1.StdAttr.ERROR_TYPE, err instanceof Error ? err.constructor.name : 'Error');
+                otelInferenceSpan.setAttribute(index_1.GenAiAttr.RESPONSE_FINISH_REASONS, ['error']);
+                otelInferenceSpan.recordException(err);
+                otelInferenceSpan.end();
+            }
             const timeToError = Date.now() - baseTelemetry.issuedTime;
             if (err.fetcherId) {
                 actualFetcher = err.fetcherId;
@@ -334,34 +470,34 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                 resumeEventSeen = true;
             }
             const processed = this.processError(err, ourRequestId, err.gitHubRequestId, usernameToScrub);
-            if (processed.type === commonTypes_1.ChatFetchResponseType.NetworkError && enableRetryOnError) {
-                const isRetryNetworkErrorEnabled = this._configurationService.getExperimentBasedConfig(configurationService_1.ConfigKey.TeamInternal.RetryNetworkErrors, this._experimentationService);
-                if (isRetryNetworkErrorEnabled) {
-                    const { retryResult, connectivityTestError, connectivityTestErrorGitHubRequestId } = await this._retryAfterError({
-                        opts,
-                        processed,
-                        telemetryProperties,
-                        requestBody,
-                        tokenCount,
-                        maxResponseTokens,
-                        timeToError,
-                        actualFetcher,
-                        bytesReceived: err.bytesReceived,
-                        baseTelemetry,
-                        streamRecorder,
-                        retryReason: 'network_error',
-                        debugNamePrefix: 'retry-error-',
-                        pendingLoggedChatRequest,
-                        token,
-                        usernameToScrub,
-                        suspendEventSeen,
-                        resumeEventSeen,
-                    });
-                    if (retryResult) {
-                        return retryResult;
-                    }
-                    telemetryProperties = { ...telemetryProperties, connectivityTestError, connectivityTestErrorGitHubRequestId };
+            const retryNetworkError = enableRetryOnError && processed.type === commonTypes_1.ChatFetchResponseType.NetworkError && this._configurationService.getExperimentBasedConfig(configurationService_1.ConfigKey.TeamInternal.RetryNetworkErrors, this._experimentationService);
+            const retryWithoutWebSocket = enableRetryOnError && useWebSocket && (processed.type === commonTypes_1.ChatFetchResponseType.NetworkError || processed.type === commonTypes_1.ChatFetchResponseType.Failed);
+            if (retryNetworkError || retryWithoutWebSocket) {
+                const { retryResult, connectivityTestError, connectivityTestErrorGitHubRequestId } = await this._retryAfterError({
+                    opts,
+                    processed,
+                    telemetryProperties,
+                    requestBody,
+                    tokenCount,
+                    maxResponseTokens,
+                    timeToError,
+                    transport,
+                    actualFetcher,
+                    bytesReceived: err.bytesReceived,
+                    baseTelemetry,
+                    streamRecorder,
+                    retryReason: 'network_error',
+                    debugNamePrefix: 'retry-error-',
+                    pendingLoggedChatRequest,
+                    token,
+                    usernameToScrub,
+                    suspendEventSeen,
+                    resumeEventSeen,
+                });
+                if (retryResult) {
+                    return retryResult;
                 }
+                telemetryProperties = { ...telemetryProperties, connectivityTestError, connectivityTestErrorGitHubRequestId };
             }
             if (processed.type === commonTypes_1.ChatFetchResponseType.Canceled) {
                 chatMLFetcherTelemetry_1.ChatMLFetcherTelemetrySender.sendCancellationTelemetry(this._telemetryService, {
@@ -369,6 +505,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                     requestId: ourRequestId,
                     model: chatEndpoint.model,
                     apiType: chatEndpoint.apiType,
+                    transport,
                     associatedRequestId: telemetryProperties.associatedRequestId,
                     retryAfterError: telemetryProperties.retryAfterError,
                     retryAfterErrorGitHubRequestId: telemetryProperties.retryAfterErrorGitHubRequestId,
@@ -401,6 +538,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                     maxResponseTokens,
                     timeToFirstToken: timeToError,
                     isVisionRequest: this.filterImageMessages(messages),
+                    transport,
                     fetcher: actualFetcher,
                     bytesReceived: err.bytesReceived,
                     issuedTime: baseTelemetry.issuedTime,
@@ -428,6 +566,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                 const res = await this._fetcherService.fetch(url, {
                     headers,
                     useFetcher,
+                    callSite: 'capi-ping',
                 });
                 if (res.status >= 200 && res.status < 300) {
                     this._logService.info(`CAPI ping successful, proceeding with chat request retry...`);
@@ -468,7 +607,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
         return authHeaders;
     }
     async _retryAfterError(params) {
-        const { opts, processed, telemetryProperties, requestBody, tokenCount, maxResponseTokens, timeToError, actualFetcher, bytesReceived, baseTelemetry, streamRecorder, retryReason, debugNamePrefix, pendingLoggedChatRequest, token, usernameToScrub, suspendEventSeen, resumeEventSeen, } = params;
+        const { opts, processed, telemetryProperties, requestBody, tokenCount, maxResponseTokens, timeToError, transport, actualFetcher, bytesReceived, baseTelemetry, streamRecorder, retryReason, debugNamePrefix, pendingLoggedChatRequest, token, usernameToScrub, suspendEventSeen, resumeEventSeen, } = params;
         // net::ERR_NETWORK_CHANGED: https://github.com/microsoft/vscode/issues/260297
         const isNetworkChangedError = ['darwin', 'linux'].includes(process.platform) && processed.reason.indexOf('net::ERR_NETWORK_CHANGED') !== -1;
         // When Electron's network process crashes, all requests through it fail permanently.
@@ -495,6 +634,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             maxResponseTokens,
             timeToFirstToken: timeToError,
             isVisionRequest: this.filterImageMessages(opts.messages),
+            transport,
             fetcher: actualFetcher,
             bytesReceived,
             issuedTime: baseTelemetry.issuedTime,
@@ -505,6 +645,8 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
         streamRecorder.callback('', 0, { text: '', retryReason });
         const retryResult = await this.fetchMany({
             ...opts,
+            useWebSocket: false,
+            ignoreStatefulMarker: opts.useWebSocket || opts.ignoreStatefulMarker,
             debugName: debugNamePrefix + opts.debugName,
             userInitiatedRequest: false, // do not mark the retry as user initiated
             telemetryProperties: {
@@ -518,9 +660,17 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             useFetcher,
         }, token);
         pendingLoggedChatRequest?.resolve(retryResult, streamRecorder.deltas);
+        if (opts.useWebSocket && retryResult.type === commonTypes_1.ChatFetchResponseType.Success) {
+            this._consecutiveWebSocketRetryFallbacks++;
+            this._logService.info(`[ChatWebSocketManager] WebSocket request failed with successful HTTP fallback (${this._consecutiveWebSocketRetryFallbacks} consecutive).`);
+            if (opts.conversationId && opts.turnId) {
+                // Closing here because the retry is transparent.
+                this._webSocketManager.closeConnection(opts.conversationId, opts.turnId);
+            }
+        }
         return { retryResult, connectivityTestError, connectivityTestErrorGitHubRequestId };
     }
-    async _fetchAndStreamChat(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, copilotToken, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions) {
+    async _fetchAndStreamChat(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, copilotToken, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, useWebSocket, turnId, conversationId, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions) {
         const isPowerSaveBlockerEnabled = this._configurationService.getExperimentBasedConfig(configurationService_1.ConfigKey.TeamInternal.ChatRequestPowerSaveBlocker, this._experimentationService);
         const blockerHandle = isPowerSaveBlockerEnabled && location !== commonTypes_1.ChatLocation.Other ? this._powerService.acquirePowerSaveBlocker() : undefined;
         let suspendEventSeen = false;
@@ -534,7 +684,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             this._logService.info(`System resumed during streaming request ${ourRequestId} (${commonTypes_1.ChatLocation.toString(location)})`);
         });
         try {
-            const fetchResult = await this._doFetchAndStreamChat(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, copilotToken, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions);
+            const fetchResult = await this._doFetchAndStreamChat(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, copilotToken, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, useWebSocket, turnId, conversationId, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions);
             return { ...fetchResult, suspendEventSeen: suspendEventSeen || undefined, resumeEventSeen: resumeEventSeen || undefined };
         }
         catch (err) {
@@ -552,28 +702,190 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             blockerHandle?.dispose();
         }
     }
-    async _doFetchAndStreamChat(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, copilotToken, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions) {
+    async _doFetchAndStreamChat(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, copilotToken, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, useWebSocket, turnId, conversationId, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions) {
         if (cancellationToken.isCancellationRequested) {
             return { result: { type: fetch_2.FetchResponseKind.Canceled, reason: 'before fetch request' } };
         }
-        this._logService.debug(`modelMaxPromptTokens ${chatEndpointInfo.modelMaxPromptTokens}`);
-        this._logService.debug(`modelMaxResponseTokens ${request.max_tokens ?? 2048}`);
-        this._logService.debug(`chat model ${chatEndpointInfo.model}`);
-        secretKey ??= copilotToken.token;
-        if (!secretKey) {
-            // If no key is set we error
-            const urlOrRequestMetadata = (0, networking_1.stringifyUrlOrRequestMetadata)(chatEndpointInfo.urlOrRequestMetadata);
-            this._logService.error(`Failed to send request to ${urlOrRequestMetadata} due to missing key`);
-            (0, stream_1.sendCommunicationErrorTelemetry)(this._telemetryService, `Failed to send request to ${urlOrRequestMetadata} due to missing key`);
-            return {
-                result: {
-                    type: fetch_2.FetchResponseKind.Failed,
-                    modelRequestId: undefined,
-                    failKind: fetch_2.ChatFailKind.TokenExpiredOrInvalid,
-                    reason: 'key is missing'
-                }
-            };
+        // OTel inference span for this LLM call
+        const serverAddress = typeof chatEndpointInfo.urlOrRequestMetadata === 'string'
+            ? (() => { try {
+                return new URL(chatEndpointInfo.urlOrRequestMetadata).hostname;
+            }
+            catch {
+                return undefined;
+            } })()
+            : undefined;
+        const chatSessionId = (0, requestLogger_1.getCurrentCapturingToken)()?.chatSessionId;
+        const otelSpan = this._otelService.startSpan(`chat ${chatEndpointInfo.model}`, {
+            kind: 2 /* SpanKind.CLIENT */,
+            attributes: {
+                [index_1.GenAiAttr.OPERATION_NAME]: index_1.GenAiOperationName.CHAT,
+                [index_1.GenAiAttr.PROVIDER_NAME]: index_1.GenAiProviderName.GITHUB,
+                [index_1.GenAiAttr.REQUEST_MODEL]: chatEndpointInfo.model,
+                [index_1.GenAiAttr.CONVERSATION_ID]: telemetryProperties?.requestId ?? ourRequestId,
+                [index_1.GenAiAttr.REQUEST_MAX_TOKENS]: request.max_tokens ?? request.max_output_tokens ?? request.max_completion_tokens ?? 2048,
+                ...(request.temperature !== undefined ? { [index_1.GenAiAttr.REQUEST_TEMPERATURE]: request.temperature } : {}),
+                ...(request.top_p !== undefined ? { [index_1.GenAiAttr.REQUEST_TOP_P]: request.top_p } : {}),
+                [index_1.CopilotChatAttr.MAX_PROMPT_TOKENS]: chatEndpointInfo.modelMaxPromptTokens,
+                ...(serverAddress ? { [index_1.StdAttr.SERVER_ADDRESS]: serverAddress } : {}),
+                ...(conversationId ? { [index_1.CopilotChatAttr.SESSION_ID]: conversationId } : {}),
+                ...(chatSessionId ? { [index_1.CopilotChatAttr.CHAT_SESSION_ID]: chatSessionId } : {}),
+            },
+        });
+        const otelStartTime = Date.now();
+        try {
+            this._logService.debug(`modelMaxPromptTokens ${chatEndpointInfo.modelMaxPromptTokens}`);
+            this._logService.debug(`modelMaxResponseTokens ${request.max_tokens ?? 2048}`);
+            this._logService.debug(`chat model ${chatEndpointInfo.model}`);
+            secretKey ??= copilotToken.token;
+            if (!secretKey) {
+                // If no key is set we error
+                const urlOrRequestMetadata = (0, networking_1.stringifyUrlOrRequestMetadata)(chatEndpointInfo.urlOrRequestMetadata);
+                this._logService.error(`Failed to send request to ${urlOrRequestMetadata} due to missing key`);
+                (0, stream_1.sendCommunicationErrorTelemetry)(this._telemetryService, `Failed to send request to ${urlOrRequestMetadata} due to missing key`);
+                return {
+                    result: {
+                        type: fetch_2.FetchResponseKind.Failed,
+                        modelRequestId: undefined,
+                        failKind: fetch_2.ChatFailKind.TokenExpiredOrInvalid,
+                        reason: 'key is missing'
+                    }
+                };
+            }
+            // WebSocket path: use persistent WebSocket connection for Responses API endpoints
+            if (useWebSocket && turnId && conversationId) {
+                const wsResult = await this._doFetchViaWebSocket(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, location, ourRequestId, turnId, conversationId, cancellationToken, userInitiatedRequest, telemetryProperties, requestKindOptions);
+                return { ...wsResult, otelSpan };
+            }
+            const httpResult = await this._doFetchViaHttp(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions);
+            return { ...httpResult, otelSpan };
+        }
+        catch (err) {
+            otelSpan.setStatus(2 /* SpanStatusCode.ERROR */, err instanceof Error ? err.message : String(err));
+            otelSpan.setAttribute(index_1.StdAttr.ERROR_TYPE, err instanceof Error ? err.constructor.name : 'Error');
+            otelSpan.recordException(err);
+            throw err;
+        }
+        finally {
+            const durationSec = (Date.now() - otelStartTime) / 1000;
+            index_1.GenAiMetrics.recordOperationDuration(this._otelService, durationSec, {
+                operationName: index_1.GenAiOperationName.CHAT,
+                providerName: index_1.GenAiProviderName.GITHUB,
+                requestModel: chatEndpointInfo.model,
+            });
+            // Span is NOT ended here — caller (fetchMany) will set token attributes and end it
+        }
+    }
+    /**
+     * Sends a chat request via a persistent WebSocket connection instead of HTTP POST.
+     * Events are the same Responses API streaming events, processed by OpenAIResponsesProcessor.
+     */
+    async _doFetchViaWebSocket(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, location, ourRequestId, turnId, conversationId, cancellationToken, userInitiatedRequest, telemetryProperties, requestKindOptions) {
+        const intent = locationToIntent(location);
+        const agentInteractionType = requestKindOptions?.kind === 'subagent' ?
+            'conversation-subagent' :
+            requestKindOptions?.kind === 'background' ?
+                'conversation-background' :
+                intent === 'conversation-agent' ? intent : undefined;
+        const additionalHeaders = {
+            'Authorization': `Bearer ${secretKey}`,
+            'X-Request-Id': ourRequestId,
+            'OpenAI-Intent': intent,
+            'X-GitHub-Api-Version': '2025-05-01',
+            'X-Interaction-Id': this._interactionService.interactionId,
+            'X-Initiator': userInitiatedRequest ? 'user' : 'agent',
+            ...(chatEndpointInfo.getExtraHeaders ? chatEndpointInfo.getExtraHeaders(location) : {}),
+        };
+        if (agentInteractionType) {
+            additionalHeaders['X-Interaction-Type'] = agentInteractionType;
+            additionalHeaders['X-Agent-Task-Id'] = ourRequestId;
+        }
+        if (request.messages?.some((m) => Array.isArray(m.content) ? m.content.some(c => 'image_url' in c) : false) && chatEndpointInfo.supportsVision) {
+            additionalHeaders['Copilot-Vision-Request'] = 'true';
         }
+        const connection = await this._webSocketManager.getOrCreateConnection(conversationId, turnId, additionalHeaders);
+        // Generate unique ID to link input and output messages
+        const modelCallId = (0, uuid_1.generateUuid)();
+        const telemetryData = telemetryData_1.TelemetryData.createAndMarkAsIssued({
+            endpoint: 'completions',
+            engineName: 'chat',
+            uiKind: commonTypes_1.ChatLocation.toString(location),
+            transport: 'websocket',
+            ...{ ...telemetryProperties, modelCallId },
+        }, {
+            maxTokenWindow: chatEndpointInfo.modelMaxPromptTokens
+        });
+        const modelRequestId = (0, fetch_1.getRequestId)(connection.responseHeaders);
+        telemetryData.extendWithRequestId(modelRequestId);
+        for (const [key, value] of Object.entries(request)) {
+            if (key === 'messages' || key === 'input') {
+                continue;
+            } // Skip messages (PII)
+            telemetryData.properties[`request.option.${key}`] = JSON.stringify(value) ?? 'undefined';
+        }
+        telemetryData.properties['headerRequestId'] = ourRequestId;
+        this._telemetryService.sendGHTelemetryEvent('request.sent', telemetryData.properties, telemetryData.measurements);
+        const requestStart = Date.now();
+        const handle = connection.sendRequest(request, cancellationToken);
+        const extendedBaseTelemetryData = baseTelemetryData.extendedBy({ modelCallId });
+        const processor = this._instantiationService.createInstance(responsesApi_1.OpenAIResponsesProcessor, extendedBaseTelemetryData, modelRequestId.headerRequestId, modelRequestId.gitHubRequestId);
+        const chatCompletions = new async_1.AsyncIterableObject(async (emitter) => {
+            try {
+                await new Promise((resolve, reject) => {
+                    handle.onEvent(event => {
+                        const completion = processor.push(event, finishedCb);
+                        if (completion) {
+                            (0, responsesApi_1.sendCompletionOutputTelemetry)(this._telemetryService, this._logService, completion, extendedBaseTelemetryData);
+                            emitter.emitOne(completion);
+                        }
+                    });
+                    handle.onError(error => {
+                        error.gitHubRequestId = modelRequestId.gitHubRequestId;
+                        if ((0, errors_2.isCancellationError)(error)) {
+                            reject(error);
+                            return;
+                        }
+                        const warningTelemetry = telemetryData.extendedBy({ error: error.message });
+                        this._telemetryService.sendGHTelemetryEvent('request.shownWarning', warningTelemetry.properties, warningTelemetry.measurements);
+                        const totalTimeMs = Date.now() - requestStart;
+                        telemetryData.measurements.totalTimeMs = totalTimeMs;
+                        telemetryData.properties.error = error.message;
+                        this._logService.debug(`request.error: [websocket], took ${totalTimeMs} ms`);
+                        this._telemetryService.sendGHTelemetryEvent('request.error', telemetryData.properties, telemetryData.measurements);
+                        reject(error);
+                    });
+                    handle.onComplete(() => {
+                        const totalTimeMs = Date.now() - requestStart;
+                        telemetryData.measurements.totalTimeMs = totalTimeMs;
+                        this._logService.debug(`request.response: [websocket], took ${totalTimeMs} ms`);
+                        this._telemetryService.sendGHTelemetryEvent('request.response', telemetryData.properties, telemetryData.measurements);
+                        resolve();
+                    });
+                });
+            }
+            finally {
+                let messagesToLog = request.messages;
+                if ((!messagesToLog || messagesToLog.length === 0) && request.input) {
+                    try {
+                        const rawMessages = (0, responsesApi_1.responseApiInputToRawMessagesForLogging)(request);
+                        messagesToLog = (0, openai_1.rawMessageToCAPI)(rawMessages);
+                    }
+                    catch (e) {
+                        this._logService.error(`Failed to convert Response API input to messages for telemetry:`, e);
+                        messagesToLog = [];
+                    }
+                }
+                (0, chatStream_1.sendEngineMessagesTelemetry)(this._telemetryService, messagesToLog ?? [], telemetryData, false, this._logService);
+            }
+        });
+        return {
+            result: {
+                type: fetch_2.FetchResponseKind.Success,
+                chatCompletions,
+            }
+        };
+    }
+    async _doFetchViaHttp(chatEndpointInfo, request, baseTelemetryData, finishedCb, secretKey, location, ourRequestId, nChoices, cancellationToken, userInitiatedRequest, telemetryProperties, useFetcher, canRetryOnce, requestKindOptions) {
         // Generate unique ID to link input and output messages
         const modelCallId = (0, uuid_1.generateUuid)();
         const response = await this._fetchWithInstrumentation(chatEndpointInfo, ourRequestId, request, secretKey, location, cancellationToken, userInitiatedRequest, { ...telemetryProperties, modelCallId }, useFetcher, canRetryOnce, requestKindOptions);
@@ -660,6 +972,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             endpoint: 'completions',
             engineName: 'chat',
             uiKind: commonTypes_1.ChatLocation.toString(location),
+            transport: 'http',
             ...telemetryProperties // This includes the modelCallId from fetchAndStreamChat
         }, {
             maxTokenWindow: chatEndpoint.modelMaxPromptTokens
@@ -946,7 +1259,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
             reason: `Request Failed: ${response.status} ${text}`
         };
     }
-    async processSuccessfulResponse(response, messages, requestBody, requestId, maxResponseTokens, promptTokenCount, timeToFirstToken, streamRecorder, baseTelemetry, chatEndpointInfo, userInitiatedRequest, fetcher, bytesReceived, suspendEventSeen, resumeEventSeen) {
+    async processSuccessfulResponse(response, messages, requestBody, requestId, maxResponseTokens, promptTokenCount, timeToFirstToken, streamRecorder, baseTelemetry, chatEndpointInfo, userInitiatedRequest, transport, fetcher, bytesReceived, suspendEventSeen, resumeEventSeen) {
         const completions = [];
         for await (const chatCompletion of response.chatCompletions) {
             chatMLFetcherTelemetry_1.ChatMLFetcherTelemetrySender.sendSuccessTelemetry(this._telemetryService, {
@@ -960,6 +1273,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
                 timeToFirstToken,
                 timeToFirstTokenEmitted: (baseTelemetry && streamRecorder.firstTokenEmittedTime) ? streamRecorder.firstTokenEmittedTime - baseTelemetry.issuedTime : -1,
                 hasImageMessages: this.filterImageMessages(messages),
+                transport,
                 fetcher,
                 bytesReceived,
                 suspendEventSeen,
@@ -1225,7 +1539,7 @@ let ChatMLFetcherImpl = class ChatMLFetcherImpl extends AbstractChatMLFetcher {
     }
 };
 exports.ChatMLFetcherImpl = ChatMLFetcherImpl;
-exports.ChatMLFetcherImpl = ChatMLFetcherImpl = __decorate([
+exports.ChatMLFetcherImpl = ChatMLFetcherImpl = ChatMLFetcherImpl_1 = __decorate([
     __param(0, fetcherService_1.IFetcherService),
     __param(1, telemetry_1.ITelemetryService),
     __param(2, requestLogger_1.IRequestLogger),
@@ -1238,7 +1552,9 @@ exports.ChatMLFetcherImpl = ChatMLFetcherImpl = __decorate([
     __param(9, configurationService_1.IConfigurationService),
     __param(10, nullExperimentationService_1.IExperimentationService),
     __param(11, powerService_1.IPowerService),
-    __param(12, instantiation_1.IInstantiationService)
+    __param(12, instantiation_1.IInstantiationService),
+    __param(13, chatWebSocketManager_1.IChatWebSocketManager),
+    __param(14, otelService_1.IOTelService)
 ], ChatMLFetcherImpl);
 /**
  * Validates a chat request payload to ensure it is valid