Memoize getFirstNonCopilotModel in scenario automation endpoint provider

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

@@ -8,32 +8,47 @@ import { ConfigKey } from '../../../platform/configuration/common/configurationS
 import { ChatEndpointFamily } from '../../../platform/endpoint/common/endpointProvider';
 import { ExtensionContributedChatEndpoint } from '../../../platform/endpoint/vscode-node/extChatEndpoint';
 import { IChatEndpoint } from '../../../platform/networking/common/networking';
+import { Delayer } from '../../../util/vs/base/common/async';
+import { MicrotaskDelay } from '../../../util/vs/base/common/symbols';
 import { ProductionEndpointProvider } from './endpointProviderImpl';
 
 export class ScenarioAutomationEndpointProviderImpl extends ProductionEndpointProvider {
+	/**
+	 * Cached first-non-copilot model. Resolved lazily on first use and invalidated when the
+	 * registered chat-model set changes. Without this cache, `getChatEndpoint` would call
+	 * `lm.selectChatModels()` (empty selector) on every invocation — which fans out across
+	 * all registered vendors and re-resolves each one. In long automation runs that run at
+	 * several Hz for the entire turn, this can dominate renderer/CDP traffic.
+	 */
+	private _firstNonCopilotModelPromise: Promise<LanguageModelChat | undefined> | undefined;
+	private _invalidateDelayer: Delayer<void> | undefined;
+	private _changeListenerInstalled = false;
+
 	override async getChatEndpoint(requestOrFamilyOrModel: LanguageModelChat | ChatRequest | ChatEndpointFamily): Promise<IChatEndpoint> {
 		const isProxyingCAPI = !!this._configService.getConfig(ConfigKey.Shared.DebugOverrideCAPIUrl) || !!this._configService.getConfig(ConfigKey.Shared.DebugOverrideProxyUrl);
 		if (this._authService.copilotToken?.isNoAuthUser && !isProxyingCAPI) {
 			// When using no auth in scenario automation, we want to force using a custom model / non-copilot for all requests
 			const getFirstNonCopilotModel = async () => {
-				const allModels = await lm.selectChatModels();
-				const firstNonCopilotModel = allModels.find(m => m.vendor !== 'copilot');
+				const firstNonCopilotModel = await this._resolveFirstNonCopilotModel();
 				if (firstNonCopilotModel) {
-					this._logService.trace(`Using custom contributed chat model`);
+					this._logService.trace(`ScenarioAutomation: using BYOK model ${firstNonCopilotModel.vendor}/${firstNonCopilotModel.id}`);
 					return this._instantiationService.createInstance(ExtensionContributedChatEndpoint, firstNonCopilotModel);
 				} else {
+					this._logService.error(`ScenarioAutomation: no non-copilot models registered`);
 					throw new Error('No custom contributed chat models found.');
 				}
 			};
 
 			// Check if we have a hard-coded family which indicates a copilot model
 			if (typeof requestOrFamilyOrModel === 'string') {
+				this._logService.trace(`ScenarioAutomation: redirecting family '${requestOrFamilyOrModel}' to BYOK`);
 				return getFirstNonCopilotModel();
 			}
 
 			// Check if a copilot model was explicitly requested in the picker
 			const model = 'model' in requestOrFamilyOrModel ? requestOrFamilyOrModel.model : requestOrFamilyOrModel;
 			if (model.vendor === 'copilot') {
+				this._logService.trace(`ScenarioAutomation: redirecting copilot model '${model.id}' to BYOK`);
 				return getFirstNonCopilotModel();
 			}
 		}
@@ -44,10 +59,46 @@ export class ScenarioAutomationEndpointProviderImpl extends ProductionEndpointPr
 			// In scenario automation, some model families (e.g. copilot-utility-small → gpt-4o-mini) may
 			// not be available via the capi proxy. Fall back to copilot-utility.
 			if (typeof requestOrFamilyOrModel === 'string') {
-				this._logService.trace(`ScenarioAutomation: failed to resolve model family '${requestOrFamilyOrModel}', falling back to copilot-utility`);
+				this._logService.warn(`ScenarioAutomation: failed to resolve model family '${requestOrFamilyOrModel}', falling back to copilot-utility: ${error}`);
 				return super.getChatEndpoint('copilot-utility');
 			}
 			throw error;
 		}
 	}
+
+	private _resolveFirstNonCopilotModel(): Promise<LanguageModelChat | undefined> {
+		this._ensureChangeListener();
+		if (!this._firstNonCopilotModelPromise) {
+			this._firstNonCopilotModelPromise = (async () => {
+				try {
+					const allModels = await lm.selectChatModels();
+					const found = allModels.find(m => m.vendor !== 'copilot');
+					this._logService.info(`ScenarioAutomation: resolved BYOK model ${found ? `${found.vendor}/${found.id}` : '<none>'} from ${allModels.length} registered model(s)`);
+					return found;
+				} catch (err) {
+					this._logService.warn(`ScenarioAutomation: selectChatModels failed; clearing cache: ${err}`);
+					this._firstNonCopilotModelPromise = undefined;
+					throw err;
+				}
+			})();
+		}
+		return this._firstNonCopilotModelPromise;
+	}
+
+	private _ensureChangeListener(): void {
+		if (this._changeListenerInstalled) {
+			return;
+		}
+		this._changeListenerInstalled = true;
+		// Coalesce bursts of model-set changes (e.g. when a BYOK provider activates and
+		// publishes several utility-alias models in quick succession) into a single
+		// invalidation so we don't churn the cache.
+		this._invalidateDelayer = this._register(new Delayer<void>(MicrotaskDelay));
+		this._register(lm.onDidChangeChatModels(() => {
+			this._invalidateDelayer!.trigger(() => {
+				this._logService.info(`ScenarioAutomation: chat model set changed; invalidating cached BYOK model`);
+				this._firstNonCopilotModelPromise = undefined;
+			}).catch(() => { /* cancelled on dispose */ });
+		}));
+	}
 }