feat(ai-chat): route Auto model by task signals

src/features/workspaces/ai/ai-inspector-view-model.test.ts

@@ -0,0 +1,28 @@
+import { describe, expect, it } from "vitest";
+
+import type { AIInspectorEvent } from "#/features/workspaces/ai/ai-inspector";
+import { getAIInspectorRunViews } from "#/features/workspaces/ai/ai-inspector-view-model";
+
+describe("AI inspector run views", () => {
+	it("surfaces the requested and routed model for auto turns", () => {
+		const [run] = getAIInspectorRunViews([
+			{
+				id: "event-1",
+				runId: "run-1",
+				sequence: 1,
+				createdAt: 1,
+				type: "turn.started",
+				payload: {
+					modelId: "claude-haiku",
+					requestedModelId: "auto",
+					routingReason: "simple-chat",
+					system: "system prompt",
+				},
+			} satisfies AIInspectorEvent,
+		]);
+
+		expect(run.modelId).toBe("claude-haiku");
+		expect(run.requestedModelId).toBe("auto");
+		expect(run.routingReason).toBe("simple-chat");
+	});
+});

src/features/workspaces/ai/ai-inspector-view-model.ts

@@ -60,6 +60,8 @@ function buildRunView(runId: string, events: AIInspectorEvent[]) {
 			case "turn.started":
 				run.startedAt = event.createdAt;
 				run.modelId = getString(payload.modelId);
+				run.requestedModelId = getString(payload.requestedModelId);
+				run.routingReason = getString(payload.routingReason);
 				run.system = getString(payload.system);
 				run.thread = payload.thread;
 				run.body = payload.body;

src/features/workspaces/ai/ai-inspector-view-types.ts

@@ -64,6 +64,8 @@ export interface AIInspectorRunView {
 	finishedAt?: number;
 	status: "running" | "completed" | "failed";
 	modelId?: string;
+	requestedModelId?: string;
+	routingReason?: string;
 	system?: string;
 	thread?: unknown;
 	body?: unknown;

src/features/workspaces/ai/ai-thread-inspector-recorder.ts

@@ -45,6 +45,8 @@ export class AIThreadInspectorRecorder {
 	async recordTurnStarted(input: {
 		ctx: TurnContext;
 		modelId: string;
+		requestedModelId: string;
+		routingReason?: string;
 		system: string;
 		thread: AIThreadContext;
 		tools: unknown;
@@ -56,6 +58,8 @@ export class AIThreadInspectorRecorder {
 			continuation: input.ctx.continuation,
 			messages: summarizeInspectorMessages(input.ctx.messages),
 			modelId: input.modelId,
+			requestedModelId: input.requestedModelId,
+			routingReason: input.routingReason,
 			system: input.system,
 			thread: input.thread,
 			tools: await summarizeInspectorTools(input.tools),

src/features/workspaces/ai/ai-thread-posthog-recorder.ts

@@ -138,6 +138,8 @@ export class AIThreadPostHogRecorder {
 	recordTurnStarted(input: {
 		ctx: TurnContext;
 		modelId: WorkspaceAiChatModelId;
+		requestedModelId: WorkspaceAiChatModelId;
+		routingReason?: string;
 		thread: AIThreadContext;
 		tools?: unknown;
 	}) {
@@ -166,6 +168,8 @@ export class AIThreadPostHogRecorder {
 			properties: {
 				...turnTelemetryProperties(turn),
 				model_id: input.modelId,
+				requested_model_id: input.requestedModelId,
+				routing_reason: input.routingReason ?? null,
 				continuation: Boolean(input.ctx.continuation),
 			},
 			...this.serverEventRuntime,

src/features/workspaces/ai/ai-thread-tcc-recorder.ts

@@ -78,6 +78,8 @@ export class AIThreadTccRecorder {
 		ctx: TurnContext;
 		env: Cloudflare.Env;
 		modelId: WorkspaceAiChatModelId;
+		requestedModelId: WorkspaceAiChatModelId;
+		routingReason?: string;
 		system: string;
 		thread: AIThreadContext;
 		tools?: unknown;
@@ -94,6 +96,8 @@ export class AIThreadTccRecorder {
 		const metadata = createTccMetadata({
 			gatewayModel,
 			modelId: input.modelId,
+			requestedModelId: input.requestedModelId,
+			routingReason: input.routingReason,
 			thread: input.thread,
 		});
 
@@ -454,6 +458,8 @@ function getTccApiKey(env: Cloudflare.Env) {
 function createTccMetadata(input: {
 	gatewayModel: string;
 	modelId: WorkspaceAiChatModelId;
+	requestedModelId: WorkspaceAiChatModelId;
+	routingReason?: string;
 	thread: AIThreadContext;
 }) {
 	return {
@@ -466,6 +472,8 @@ function createTccMetadata(input: {
 		gateway_model: input.gatewayModel,
 		model_id: input.modelId,
 		mutation_mode: input.thread.promptScope.canMutate ? "mutate" : "view",
+		requested_model_id: input.requestedModelId,
+		...(input.routingReason ? { routing_reason: input.routingReason } : {}),
 		workspace_id: input.thread.workspaceId,
 	} satisfies Record<string, string>;
 }

src/features/workspaces/ai/ai-thread-telemetry-recorder.ts

@@ -40,7 +40,11 @@ export class AIThreadTelemetryRecorder {
 
 	async recordTurnStarted(input: {
 		ctx: TurnContext;
+		/** Model the turn actually runs on (the routed model for Auto turns). */
 		modelId: WorkspaceAiChatModelId;
+		/** Model the user picked; "auto" when the router chose modelId. */
+		requestedModelId: WorkspaceAiChatModelId;
+		routingReason?: string;
 		system: string;
 		thread: AIThreadContext;
 		tools: unknown;

src/features/workspaces/ai/ai-thread.ts

@@ -31,6 +31,10 @@ import {
 	getWorkspaceAiGatewayProviderOptions,
 	getWorkspaceAiLanguageModel,
 } from "#/features/workspaces/ai/ai-thread-runtime";
+import {
+	extractWorkspaceAiRoutingSignals,
+	routeWorkspaceAiAutoModel,
+} from "#/features/workspaces/ai/model-router";
 import {
 	DEFAULT_WORKSPACE_AI_CHAT_MODEL_ID,
 	getWorkspaceAiChatModel,
@@ -69,7 +73,11 @@ type AIThreadRunSettlement =
 	  };
 
 interface AIThreadUsageContext {
+	/** Model actually run and metered; equals requestedModelId unless Auto routed. */
 	modelId: WorkspaceAiChatModelId;
+	/** Model the user picked; "auto" when the router chose modelId. */
+	requestedModelId: WorkspaceAiChatModelId;
+	routingReason?: string;
 	thread: AIThreadContext;
 }
 
@@ -173,9 +181,14 @@ export function createAIThreadClass(getUserAIStore: () => typeof UserAIStore) {
 				void this.keepAliveWhile(() => this._maybeGenerateThreadTitle());
 			}
 
-			const modelId = resolveWorkspaceAiChatModelId(ctx.body?.modelId);
+			const requestedModelId = resolveWorkspaceAiChatModelId(ctx.body?.modelId);
+			const route = this._resolveAutoModelRoute(requestedModelId, ctx);
+			const modelId = route?.modelId ?? requestedModelId;
 
 			if (!ctx.continuation) {
+				// Enforcement is currently disabled inside this check. When it lands
+				// it must gate on the routed model's billing tier (modelId here), not
+				// the requested picker id, so Auto turns cannot dodge premium gating.
 				const access = await checkWorkspaceAiMessageAccess({
 					env: this.env,
 					modelId,
@@ -188,6 +201,8 @@ export function createAIThreadClass(getUserAIStore: () => typeof UserAIStore) {
 
 				this.activeUsageContext = {
 					modelId,
+					requestedModelId,
+					routingReason: route?.reason,
 					thread,
 				};
 			}
@@ -213,6 +228,8 @@ export function createAIThreadClass(getUserAIStore: () => typeof UserAIStore) {
 			await this.telemetry.recordTurnStarted({
 				ctx,
 				modelId,
+				requestedModelId,
+				routingReason: route?.reason,
 				system,
 				thread,
 				tools: activeTools,
@@ -293,6 +310,39 @@ export function createAIThreadClass(getUserAIStore: () => typeof UserAIStore) {
 			return this.telemetry.getInspectorSnapshot(this.name);
 		}
 
+		// Auto resolves to a concrete standard-tier model at turn time; every
+		// other picker id passes through untouched. Continuation turns reuse the
+		// id routed when the run started so a run never switches models
+		// mid-flight; if that state is gone (e.g. recovery), we route again from
+		// the current turn's signals.
+		private _resolveAutoModelRoute(
+			requestedModelId: WorkspaceAiChatModelId,
+			ctx: TurnContext,
+		): { modelId: WorkspaceAiChatModelId; reason: string } | undefined {
+			if (requestedModelId !== "auto") {
+				return undefined;
+			}
+
+			const activeContext = ctx.continuation ? this.activeUsageContext : undefined;
+
+			if (activeContext && activeContext.requestedModelId !== "auto") {
+				// The run started from an explicit picker choice; a continuation
+				// whose body lost the selection must not be re-routed.
+				return undefined;
+			}
+
+			if (activeContext) {
+				return {
+					modelId: activeContext.modelId,
+					reason: activeContext.routingReason ?? "turn-continuation",
+				};
+			}
+
+			const routed = routeWorkspaceAiAutoModel(extractWorkspaceAiRoutingSignals(ctx));
+
+			return { modelId: routed.modelId, reason: routed.reason };
+		}
+
 		private async _getThreadContext() {
 			const directory = await this.parentAgent(getUserAIStore());
 			return directory.getThreadContext(this.name);