Agent Host changes for lramos15/agents/auto-tiers-routing-profiles-optimization

Author: lramos15

extensions/copilot/package.json

@@ -4389,7 +4389,7 @@
               "null"
             ],
             "default": null,
-            "markdownDescription": "Overrides the routing tier that the `Auto` model requests, ignoring both the tier picked in the model picker and the tier inline chat defaults to. Accepts `eco`, `balanced`, `max`, or `fast`. Used by evals.\n\n**Note**: This is an advanced debugging setting.",
+            "markdownDescription": "Overrides the routing tier that the `Auto` model requests, ignoring both the tier picked in the model picker and the tier inline chat defaults to. Accepts `efficiency`, `balance`, `intelligence`, or `fast`. Used by evals.\n\n**Note**: This is an advanced debugging setting.",
             "tags": [
               "advanced"
             ]

extensions/copilot/src/extension/conversation/common/languageModelAccess.ts

@@ -90,9 +90,9 @@ export function buildReasoningEffortSchemaProperty(effortLevels: readonly string
  */
 export function getAutoModeTierLabel(tier: string): string {
 	switch (tier) {
-		case 'eco': return l10n.t('Efficiency');
-		case 'balanced': return l10n.t('Balance');
-		case 'max': return l10n.t('Intelligence');
+		case 'efficiency': return l10n.t('Efficiency');
+		case 'balance': return l10n.t('Balance');
+		case 'intelligence': return l10n.t('Intelligence');
 		case 'fast': return l10n.t('Fast');
 		default: return tier.charAt(0).toUpperCase() + tier.slice(1);
 	}
@@ -104,9 +104,9 @@ export function getAutoModeTierLabel(tier: string): string {
  */
 export function getAutoModeTierDescription(tier: string): string {
 	switch (tier) {
-		case 'eco': return l10n.t('Cheaper models for everyday tasks');
-		case 'balanced': return l10n.t('Balances capability and cost');
-		case 'max': return l10n.t('Most capable models, higher cost');
+		case 'efficiency': return l10n.t('Cheaper models for everyday tasks');
+		case 'balance': return l10n.t('Balances capability and cost');
+		case 'intelligence': return l10n.t('Most capable models, higher cost');
 		case 'fast': return l10n.t('Lowest latency models');
 		default: return tier;
 	}

extensions/copilot/src/extension/conversation/vscode-node/test/languageModelAccess.test.ts

@@ -663,7 +663,7 @@ suite('reasoning effort schema', () => {
 	});
 
 	test('Kimi K3 prefers high when available', () => {
-		assert.strictEqual(pickDefaultReasoningEffort(['low', 'high', 'max'], 'kimi-k3'), 'high');
+		assert.strictEqual(pickDefaultReasoningEffort(['low', 'high', 'intelligence'], 'kimi-k3'), 'high');
 	});
 
 	test('non-claude family prefers medium when available', () => {
@@ -698,10 +698,10 @@ suite('auto mode tier schema', () => {
 		assert.deepStrictEqual(buildAutoModeTierSchemaProperty(selectableAutoModeTiers, defaultAutoModeTier), {
 			type: 'string',
 			title: 'Optimize for',
-			enum: ['eco', 'balanced', 'max'],
+			enum: ['efficiency', 'balance', 'intelligence'],
 			enumItemLabels: ['Efficiency', 'Balance', 'Intelligence'],
 			enumDescriptions: ['Cheaper models for everyday tasks', 'Balances capability and cost', 'Most capable models, higher cost'],
-			default: 'balanced',
+			default: 'balance',
 			group: 'navigation',
 		});
 	});

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

@@ -7,7 +7,7 @@
  * Routing profiles accepted by `POST /auto`. A tier is picked per session and
  * biases which models the router may choose from.
  */
-export const autoModeTiers = ['eco', 'balanced', 'max', 'fast'] as const;
+export const autoModeTiers = ['efficiency', 'balance', 'intelligence', 'fast'] as const;
 
 export type AutoModeTier = typeof autoModeTiers[number];
 
@@ -17,17 +17,37 @@ export type AutoModeTier = typeof autoModeTiers[number];
  * offered as a choice. It remains reachable through the internal
  * {@link ConfigKey.Advanced.AutoModeTierOverride} setting.
  */
-export const selectableAutoModeTiers: readonly AutoModeTier[] = ['eco', 'balanced', 'max'];
+export const selectableAutoModeTiers: readonly AutoModeTier[] = ['efficiency', 'balance', 'intelligence'];
 
 /** The tier used when the user has not picked one. */
-export const defaultAutoModeTier: AutoModeTier = 'balanced';
+export const defaultAutoModeTier: AutoModeTier = 'balance';
 
 /** The tier inline chat defaults to; latency matters more than routing depth there. */
 export const inlineChatAutoModeTier: AutoModeTier = 'fast';
 
 /** Key the selected tier is stored under in the Auto model's configuration. */
 export const AUTO_MODE_TIER_PROPERTY = 'tier';
 
+/**
+ * Tier names retired in favour of the current ones. `POST /auto` still accepts
+ * them, but the agent runtime rejects them at its session boundary, so one
+ * vocabulary is used everywhere and these are mapped forward on read.
+ */
+const retiredAutoModeTiers: Readonly<Record<string, AutoModeTier>> = {
+	eco: 'efficiency',
+	balanced: 'balance',
+	max: 'intelligence',
+};
+
+/**
+ * Maps a retired tier name to its current one, leaving anything else untouched.
+ * Only raw inputs that bypass the picker schema need this, notably the override
+ * setting; a persisted picker value is already filtered against the live schema.
+ */
+export function normalizeAutoModeTier(value: string): string {
+	return retiredAutoModeTiers[value] ?? value;
+}
+
 /**
  * Narrows an untrusted value (persisted model configuration, or configuration
  * supplied by a third-party extension through the `vscode.lm` API) to a tier the

extensions/copilot/src/platform/endpoint/node/autoV2Fetcher.ts

@@ -151,7 +151,7 @@ export class AutoV2Fetcher {
 				"conversationId": { "classification": "SystemMetaData", "purpose": "FeatureInsight", "comment": "The conversation ID in which the selection was made." },
 				"vscodeRequestId": { "classification": "SystemMetaData", "purpose": "FeatureInsight", "comment": "The VS Code chat request id in which the selection was made." },
 				"selectedModel": { "classification": "SystemMetaData", "purpose": "FeatureInsight", "comment": "The model the server selected for this prompt." },
-				"tier": { "classification": "SystemMetaData", "purpose": "FeatureInsight", "comment": "The routing profile requested for this selection, e.g. eco, balanced, max, fast. Empty when none was requested." },
+				"tier": { "classification": "SystemMetaData", "purpose": "FeatureInsight", "comment": "The routing profile requested for this selection, e.g. efficiency, balance, intelligence, fast. Empty when none was requested." },
 				"e2eLatencyMs": { "classification": "SystemMetaData", "purpose": "PerformanceAndHealth", "isMeasurement": true, "comment": "The end-to-end latency of the auto request in milliseconds, including network overhead." },
 				"scoreReasoning": { "classification": "SystemMetaData", "purpose": "FeatureInsight", "isMeasurement": true, "comment": "Hydra per-dimension score for reasoning. -1 if not present in the response." },
 				"scoreCodeGen": { "classification": "SystemMetaData", "purpose": "FeatureInsight", "isMeasurement": true, "comment": "Hydra per-dimension score for code generation. -1 if not present in the response." },

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

@@ -18,7 +18,7 @@ import { IChatEndpoint } from '../../networking/common/networking';
 import { IRequestLogger } from '../../requestLogger/common/requestLogger';
 import { IExperimentationService } from '../../telemetry/common/nullExperimentationService';
 import { ITelemetryService } from '../../telemetry/common/telemetry';
-import { AUTO_MODE_TIER_PROPERTY, autoModeTiers, defaultAutoModeTier, inlineChatAutoModeTier, isSelectableAutoModeTier, type AutoModeTier } from '../common/autoModeTiers';
+import { AUTO_MODE_TIER_PROPERTY, autoModeTiers, defaultAutoModeTier, inlineChatAutoModeTier, isSelectableAutoModeTier, normalizeAutoModeTier, type AutoModeTier } from '../common/autoModeTiers';
 import { ICAPIClientService } from '../common/capiClient';
 import type { IChatModelCapabilities, IChatModelInformation } from '../common/endpointProvider';
 import { AutoChatEndpoint } from './autoChatEndpoint';
@@ -397,16 +397,19 @@ export class AutomodeService extends Disposable implements IAutomodeService {
 	 *
 	 * Only a non-default selection counts as explicit: the workbench materializes
 	 * the schema default into `modelConfiguration` and strips a pick of the
-	 * default back out when storing it, so a `balanced` entry cannot be told
-	 * apart from "never picked" — reading it as a selection would make the inline
-	 * pin below unreachable.
+	 * default back out when storing it, so a `balance` entry cannot be told apart
+	 * from "never picked" — reading it as a selection would make the inline pin
+	 * below unreachable.
 	 */
 	private _resolveTier(chatRequest: IAutoModeRoutingRequest | undefined): AutoModeTier | undefined {
 		const override = this._configurationService.getConfig(ConfigKey.Advanced.AutoModeTierOverride);
 		if (override) {
+			// Normalized because the override is a raw string setting, so unlike a picker value it never
+			// passes through the schema filter that drops retired names.
+			const normalized = normalizeAutoModeTier(override);
 			// The override is internal, so unlike the picker it may select `fast`.
-			if ((autoModeTiers as readonly string[]).includes(override)) {
-				return override as AutoModeTier;
+			if ((autoModeTiers as readonly string[]).includes(normalized)) {
+				return normalized as AutoModeTier;
 			}
 			this._logService.warn(`[AutomodeService] Ignoring auto tier override '${override}' — not one of [${autoModeTiers.join(', ')}].`);
 		}

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

@@ -390,10 +390,10 @@ describe('AutomodeService', () => {
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
 			await Promise.all([
-				route('session-a', 'max'),
-				route('session-b', 'max'),
-				route('session-a', 'eco'),
-				route('session-a', 'max', [{ value: { mimeType: 'image/png' } }]),
+				route('session-a', 'intelligence'),
+				route('session-b', 'intelligence'),
+				route('session-a', 'efficiency'),
+				route('session-a', 'intelligence', [{ value: { mimeType: 'image/png' } }]),
 			]);
 
 			expect(autoCalls()).toHaveLength(4);
@@ -632,10 +632,10 @@ describe('AutomodeService', () => {
 				location: ChatLocation.Editor,
 				prompt: 'test prompt',
 				sessionId: 'session-auto-inline-tier',
-				modelConfiguration: { tier: 'max' },
+				modelConfiguration: { tier: 'intelligence' },
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
-			expect(autoRequestBodies()).toEqual([{ prompt: 'test prompt', tier: 'max' }]);
+			expect(autoRequestBodies()).toEqual([{ prompt: 'test prompt', tier: 'intelligence' }]);
 		});
 
 		it('sends the tier picked in the model configuration', async () => {
@@ -648,10 +648,10 @@ describe('AutomodeService', () => {
 				location: ChatLocation.Panel,
 				prompt: 'test prompt',
 				sessionId: 'session-auto-tier',
-				modelConfiguration: { tier: 'max' },
+				modelConfiguration: { tier: 'intelligence' },
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
-			expect(autoRequestBodies()).toEqual([{ prompt: 'test prompt', tier: 'max' }]);
+			expect(autoRequestBodies()).toEqual([{ prompt: 'test prompt', tier: 'intelligence' }]);
 		});
 
 		it('falls back to the default tier when the configured tier is not user selectable', async () => {
@@ -667,7 +667,7 @@ describe('AutomodeService', () => {
 				modelConfiguration: { tier: 'fast' },
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
-			expect(autoRequestBodies()).toEqual([{ prompt: 'test prompt', tier: 'balanced' }]);
+			expect(autoRequestBodies()).toEqual([{ prompt: 'test prompt', tier: 'balance' }]);
 		});
 
 		it('re-routes the conversation when the tier changes', async () => {
@@ -680,35 +680,35 @@ describe('AutomodeService', () => {
 				location: ChatLocation.Panel,
 				prompt: 'test prompt',
 				sessionId: 'session-auto-tier-change',
-				modelConfiguration: { tier: 'eco' },
+				modelConfiguration: { tier: 'efficiency' },
 			} as unknown as ChatRequest;
 
 			await automodeService.resolveAutoModeEndpoint(chatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 			await automodeService.resolveAutoModeEndpoint({ ...chatRequest, prompt: 'second turn' } as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
-			await automodeService.resolveAutoModeEndpoint({ ...chatRequest, prompt: 'third turn', modelConfiguration: { tier: 'max' } } as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
+			await automodeService.resolveAutoModeEndpoint({ ...chatRequest, prompt: 'third turn', modelConfiguration: { tier: 'intelligence' } } as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
-			expect(autoRequestBodies().map(b => b.tier)).toEqual(['eco', 'max']);
+			expect(autoRequestBodies().map(b => b.tier)).toEqual(['efficiency', 'intelligence']);
 		});
 
 		it('lets the tier override win over the picker and the inline chat pin', async () => {
 			const gpt4oEndpoint = createEndpoint('gpt-4o', 'OpenAI');
 			mockAuto(autoResponse('gpt-4o'));
 
-			setTierOverride('eco');
+			setTierOverride('efficiency');
 			automodeService = createService();
 			await automodeService.resolveAutoModeEndpoint({
 				location: ChatLocation.Panel,
 				prompt: 'panel turn',
 				sessionId: 'session-override-panel',
-				modelConfiguration: { tier: 'max' },
+				modelConfiguration: { tier: 'intelligence' },
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 			await automodeService.resolveAutoModeEndpoint({
 				location: ChatLocation.Editor,
 				prompt: 'inline turn',
 				sessionId: 'session-override-inline',
 			} as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
-			expect(autoRequestBodies().map(b => b.tier)).toEqual(['eco', 'eco']);
+			expect(autoRequestBodies().map(b => b.tier)).toEqual(['efficiency', 'efficiency']);
 		});
 
 		// The override is an internal/eval knob, so unlike the picker it may target
@@ -741,10 +741,10 @@ describe('AutomodeService', () => {
 				location: ChatLocation.Panel,
 				prompt: 'panel turn',
 				sessionId: 'session-override-bogus',
-				modelConfiguration: { tier: 'max' },
+				modelConfiguration: { tier: 'intelligence' },
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
-			expect(autoRequestBodies()).toEqual([{ prompt: 'panel turn', tier: 'max' }]);
+			expect(autoRequestBodies()).toEqual([{ prompt: 'panel turn', tier: 'intelligence' }]);
 		});
 
 		it('announces tier support when the setting changes', async () => {
@@ -755,7 +755,7 @@ describe('AutomodeService', () => {
 			const listener = automodeService.onDidChangeAutoModeTierSupport(() => announced++);
 			await configurationService.setConfig(ConfigKey.Advanced.AutoModeTiersEnabled, true);
 			// An unrelated change must not re-announce.
-			await configurationService.setConfig(ConfigKey.Advanced.AutoModeTierOverride, 'max');
+			await configurationService.setConfig(ConfigKey.Advanced.AutoModeTierOverride, 'intelligence');
 			listener.dispose();
 
 			expect({ announced, supported: automodeService.areAutoModeTiersSupported() }).toEqual({ announced: 1, supported: true });
@@ -771,18 +771,18 @@ describe('AutomodeService', () => {
 				location: ChatLocation.Panel,
 				prompt: 'first turn',
 				sessionId: 'session-auto-tier-error',
-				modelConfiguration: { tier: 'eco' },
+				modelConfiguration: { tier: 'efficiency' },
 			} as unknown as ChatRequest;
 			const first = await automodeService.resolveAutoModeEndpoint(chatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 			expect(first.model).toBe('gpt-4o');
 
-			// The tier changes and the re-route fails: the eco endpoint must not be
+			// The tier changes and the re-route fails: the efficiency endpoint must not be
 			// handed back as though it satisfied the new tier.
 			mockAuto({ error: 'server_error' }, 500);
 			await expect(automodeService.resolveAutoModeEndpoint({
 				...chatRequest,
 				prompt: 'second turn',
-				modelConfiguration: { tier: 'max' },
+				modelConfiguration: { tier: 'intelligence' },
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint])).rejects.toThrow();
 		});
 
@@ -798,15 +798,15 @@ describe('AutomodeService', () => {
 				location: ChatLocation.Panel,
 				prompt: 'first turn',
 				sessionId: 'session-auto-tier-discount',
-				modelConfiguration: { tier: 'eco' },
+				modelConfiguration: { tier: 'efficiency' },
 			} as unknown as ChatRequest;
 			await automodeService.resolveAutoModeEndpoint(chatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
 			mockAuto(autoResponse('gpt-4o', { discounted_costs: { 'gpt-4o': 0.9 } }));
 			await automodeService.resolveAutoModeEndpoint({
 				...chatRequest,
 				prompt: 'second turn',
-				modelConfiguration: { tier: 'max' },
+				modelConfiguration: { tier: 'intelligence' },
 			} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
 			const discounts = (mockInstantiationService.createInstance as ReturnType<typeof vi.fn>).mock.calls.map(c => c[3]);
@@ -825,7 +825,7 @@ describe('AutomodeService', () => {
 					location,
 					prompt: 'test prompt',
 					sessionId: `session-tiers-off-${location}`,
-					modelConfiguration: { tier: 'max' },
+					modelConfiguration: { tier: 'intelligence' },
 				} as unknown as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 			}
 
@@ -843,15 +843,32 @@ describe('AutomodeService', () => {
 			const gpt4oEndpoint = createEndpoint('gpt-4o', 'OpenAI');
 			mockAuto(autoResponse('gpt-4o'));
 
-			setTierOverride('max');
+			setTierOverride('intelligence');
 			automodeService = createService();
 			await automodeService.resolveAutoModeEndpoint({
 				location: ChatLocation.Panel,
 				prompt: 'panel turn',
 				sessionId: 'session-override-tiers-off',
 			} as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
 
-			expect(autoRequestBodies()).toEqual([{ prompt: 'panel turn', tier: 'max' }]);
+			expect(autoRequestBodies()).toEqual([{ prompt: 'panel turn', tier: 'intelligence' }]);
+		});
+
+		// The override is a raw string setting, so a config left on a retired name by an
+		// eval or an internal user must keep working rather than silently fall back.
+		it('maps a retired tier name in the override to its current one', async () => {
+			const gpt4oEndpoint = createEndpoint('gpt-4o', 'OpenAI');
+			mockAuto(autoResponse('gpt-4o'));
+
+			setTierOverride('eco');
+			automodeService = createService();
+			await automodeService.resolveAutoModeEndpoint({
+				location: ChatLocation.Panel,
+				prompt: 'panel turn',
+				sessionId: 'session-override-retired',
+			} as ChatRequest, [mockChatEndpoint, gpt4oEndpoint]);
+
+			expect(autoRequestBodies()).toEqual([{ prompt: 'panel turn', tier: 'efficiency' }]);
 		});
 	});
 
@@ -875,7 +892,7 @@ describe('AutomodeService', () => {
 			for (let i = 0; i < 50; i++) {
 				await route(`session-${i}`, `turn ${i}`);
 			}
-			await route('session-49', 'retiered turn', 'max');
+			await route('session-49', 'retiered turn', 'intelligence');
 
 			const callsBefore = autoCalls().length;
 			await route('session-0', 'follow up');