chore: improve types and atmn

packages/atmn/src/commands/push/push.ts

@@ -349,6 +349,32 @@ function normalizeFeatureForCompare(f: Feature): Record<string, unknown> {
 			}));
 	}
 
+	if (f.type === "ai_credit_system") {
+		const ai = f as Extract<Feature, { type: "ai_credit_system" }>;
+		if (ai.modelMarkups && Object.keys(ai.modelMarkups).length > 0) {
+			result.modelMarkups = Object.fromEntries(
+				Object.entries(ai.modelMarkups)
+					.sort(([a], [b]) => a.localeCompare(b))
+					.map(([modelId, entry]) => [
+						modelId,
+						{
+							markup: entry.markup,
+							inputCost: entry.inputCost,
+							outputCost: entry.outputCost,
+						},
+					]),
+			);
+		}
+		if (ai.defaultMarkup != null) result.defaultMarkup = ai.defaultMarkup;
+		if (ai.providerMarkups && Object.keys(ai.providerMarkups).length > 0) {
+			result.providerMarkups = Object.fromEntries(
+				Object.entries(ai.providerMarkups).sort(([a], [b]) =>
+					a.localeCompare(b),
+				),
+			);
+		}
+	}
+
 	return result;
 }
 

packages/atmn/src/lib/transforms/apiToSdk/feature.ts

@@ -60,6 +60,8 @@ export const featureTransformer = createTransformer<RawApiFeature, Feature>({
 				...BASE_COMPUTE,
 				type: () => "ai_credit_system" as const,
 				modelMarkups: (api) => mapModelMarkups(api),
+				defaultMarkup: (api) => api.default_markup ?? undefined,
+				providerMarkups: (api) => api.provider_markups ?? undefined,
 			},
 		},
 

packages/atmn/src/lib/transforms/sdkToApi/feature.ts

@@ -16,6 +16,8 @@ export interface ApiFeatureParams {
 		input_cost?: number;
 		output_cost?: number;
 	}>;
+	default_markup?: number;
+	provider_markups?: Record<string, { markup: number }>;
 }
 
 export function transformFeatureToApi(feature: Feature): ApiFeatureParams {
@@ -44,17 +46,25 @@ export function transformFeatureToApi(feature: Feature): ApiFeatureParams {
 		}));
 	}
 
-	if (feature.type === "ai_credit_system" && feature.modelMarkups) {
-		base.model_markups = Object.fromEntries(
-			Object.entries(feature.modelMarkups).map(([modelId, entry]) => [
-				modelId,
-				{
-					markup: entry.markup,
-					input_cost: entry.inputCost,
-					output_cost: entry.outputCost,
-				},
-			])
-		);
+	if (feature.type === "ai_credit_system") {
+		if (feature.modelMarkups) {
+			base.model_markups = Object.fromEntries(
+				Object.entries(feature.modelMarkups).map(([modelId, entry]) => [
+					modelId,
+					{
+						markup: entry.markup,
+						input_cost: entry.inputCost,
+						output_cost: entry.outputCost,
+					},
+				])
+			);
+		}
+		if (feature.defaultMarkup !== undefined) {
+			base.default_markup = feature.defaultMarkup;
+		}
+		if (feature.providerMarkups) {
+			base.provider_markups = feature.providerMarkups;
+		}
 	}
 
 	return base;

packages/atmn/src/lib/transforms/sdkToCode/feature.ts

@@ -42,9 +42,17 @@ export function buildFeatureCode(feature: Feature, varNameOverride?: string): st
 		lines.push(`\tcreditSchema: ${formatValue(feature.creditSchema)},`);
 	}
 
-	// Add modelMarkups for ai_credit_system features
-	if (feature.type === "ai_credit_system" && feature.modelMarkups) {
-		lines.push(`\tmodelMarkups: ${formatValue(feature.modelMarkups)},`);
+	// Add markup config for ai_credit_system features
+	if (feature.type === "ai_credit_system") {
+		if (feature.modelMarkups) {
+			lines.push(`\tmodelMarkups: ${formatValue(feature.modelMarkups)},`);
+		}
+		if (feature.defaultMarkup !== undefined) {
+			lines.push(`\tdefaultMarkup: ${feature.defaultMarkup},`);
+		}
+		if (feature.providerMarkups) {
+			lines.push(`\tproviderMarkups: ${formatValue(feature.providerMarkups)},`);
+		}
 	}
 
 	lines.push(`});`);

server/src/internal/agent/rules/actions/generateAndUpdateAgentRules.ts

@@ -11,13 +11,17 @@ export const generateAndUpdateAgentRules = async ({
 	endTime?: string;
 	startTime?: string;
 }) => {
-	const generated = await generateAgentRules({ ctx, endTime, startTime });
+	const [generated, existing] = await Promise.all([
+		generateAgentRules({ ctx, endTime, startTime }),
+		agentRulesRepo.get({ db: ctx.db, orgId: ctx.org.id }),
+	]);
+	// Generation only derives entity/credit rules; never overwrite user-written notes.
 	const rules = await agentRulesRepo.upsert({
 		db: ctx.db,
 		metadata: generated.metadata,
 		orgId: ctx.org.id,
 		orgSlug: ctx.org.slug,
-		rules: generated.rules,
+		rules: { ...generated.rules, notes: existing.notes },
 	});
 
 	return {

server/src/internal/balances/track/utils/buildAiCreditCostProperty.ts

@@ -14,7 +14,7 @@ export const buildAiCreditCostProperty = ({
 	featureDeductions: FeatureDeduction[];
 	entries: Array<{ featureId: string; amount: number }>;
 }): Record<string, number> | undefined => {
-	const aiDeduction = featureDeductions.find((d) => d.tokenUsage);
+	const aiDeduction = featureDeductions.find((d) => d.tokens);
 	if (!aiDeduction) return;
 
 	const creditCost: Record<string, number> = {};

server/src/internal/balances/track/utils/getTokenTrackParams.ts

@@ -139,12 +139,14 @@ export const getTokenTrackParams = async ({
 		{
 			feature: aiCreditFeature,
 			deduction: 1,
-			tokenUsage: {
-				modelName: input.model_id,
-				inputTokens: input.input_tokens,
-				outputTokens: input.output_tokens,
+			tokens: {
+				usage: {
+					modelName: input.model_id,
+					inputTokens: input.input_tokens,
+					outputTokens: input.output_tokens,
+				},
+				cost,
 			},
-			precomputedCreditCost: cost,
 		},
 	];
 

server/src/internal/balances/utils/deduction/computeCreditCosts.ts

@@ -8,8 +8,8 @@ export type CreditCostLookup = (entitlementId: string) => number;
 
 /**
  * Computes the credit cost for each customer entitlement and returns a lookup
- * function. Uses precomputedCreditCost when available (token tracking),
- * otherwise calls getCreditCost per entitlement (credit system schema lookups).
+ * function. Token deductions carry their USD cost from the API layer; all other
+ * costs come from credit system schema ratios.
  */
 export const computeCreditCosts = async ({
 	cusEnts,
@@ -20,33 +20,23 @@ export const computeCreditCosts = async ({
 }): Promise<CreditCostLookup> => {
 	const costMap = new Map<string, number>();
 
-	const tokens = deduction.tokenUsage
-		? {
-				input: deduction.tokenUsage.inputTokens,
-				output: deduction.tokenUsage.outputTokens,
-			}
-		: undefined;
-
 	await Promise.all(
 		cusEnts.map(async (ce) => {
-			// Precomputed cost (from /track/tokens) is in the AI credit feature's
-			// native unit (USD). It applies 1:1 to that feature's own entitlement,
-			// but parent credit systems still need their schema ratio applied —
-			// fall through to getCreditCost with amount = precomputed cost.
+			// A token deduction's cost is in the AI feature's native unit (USD): it
+			// applies 1:1 to its own entitlement, while parent credit systems apply
+			// their schema ratio to it via getCreditCost's amount.
 			if (
-				deduction.precomputedCreditCost != null &&
+				deduction.tokens &&
 				ce.entitlement.feature.id === deduction.feature.id
 			) {
-				costMap.set(ce.id, deduction.precomputedCreditCost);
+				costMap.set(ce.id, deduction.tokens.cost);
 				return;
 			}
 
 			const creditCost = await getCreditCost({
 				featureId: deduction.feature.id,
 				creditSystem: ce.entitlement.feature,
-				amount: deduction.precomputedCreditCost,
-				modelName: deduction.tokenUsage?.modelName,
-				tokens,
+				amount: deduction.tokens?.cost,
 			});
 			costMap.set(ce.id, creditCost);
 		}),

server/src/internal/balances/utils/types/featureDeduction.ts

@@ -7,13 +7,18 @@ export type TokenUsage = {
 	outputTokens: number;
 };
 
+/** Token usage and its USD cost are priced together at the API layer — one cannot exist without the other. */
+export type TokenDeduction = {
+	usage: TokenUsage;
+	cost: number;
+};
+
 export type FeatureDeduction = {
 	feature: Feature;
 	deduction: number;
 	targetBalance?: number;
-	tokenUsage?: TokenUsage;
-	/** Pre-computed dollar cost; if set, the deduction layer skips its own getCreditCost call. */
-	precomputedCreditCost?: number;
+	/** Present only for track_tokens deductions; standard deductions omit it. */
+	tokens?: TokenDeduction;
 	lock?: LockParams;
 	lockReceipt?: LockReceipt;
 	lockReceiptKey?: string;

server/src/internal/features/creditSystemUtils.ts

@@ -96,17 +96,22 @@ export const getCreditCost = async ({
 		return amount;
 	}
 	if (isAiCreditSystem(creditSystem.type)) {
-		if (!tokens || !modelName) {
-			throw new RecaseError({
-				message: "modelName and tokens must be provided for AI credit systems",
-				code: ErrCode.InvalidRequest,
-				statusCode: 400,
+		if (tokens && modelName) {
+			return await getModelCreditCost({
+				modelName,
+				creditSystem,
+				...tokens,
 			});
 		}
-		return await getModelCreditCost({
-			modelName,
-			creditSystem,
-			...tokens,
+		// No token context (plain /track values, balance updates, queued replays):
+		// the feature's own balance is already in USD, so the value maps 1:1.
+		if (featureId === creditSystem.id) {
+			return amount;
+		}
+		throw new RecaseError({
+			message: "modelName and tokens must be provided for AI credit systems",
+			code: ErrCode.InvalidRequest,
+			statusCode: 400,
 		});
 	}
 	// If tracking the credit system feature itself, 1:1 mapping

server/src/internal/features/utils/getModelPricing.ts

@@ -7,9 +7,13 @@ const CACHE_KEY = "models_dev_pricing";
 const STALE_KEY = `${CACHE_KEY}_stale`;
 const TTL_PRIMARY = 60 * 60 * 3;
 const TTL_STALE = 60 * 60 * 24 * 3;
+// Runs inside the track request path — a hanging models.dev must not hang tracks.
+const FETCH_TIMEOUT_MS = 5000;
 
 const fetchFromSource = async (): Promise<ModelPricingData> => {
-	const response = await fetch("https://models.dev/api.json");
+	const response = await fetch("https://models.dev/api.json", {
+		signal: AbortSignal.timeout(FETCH_TIMEOUT_MS),
+	});
 	if (!response.ok) {
 		throw new InternalError({
 			message: `models.dev returned ${response.status}`,

server/tests/integration/balances/track/basic/track-tokens-orbs.test.ts

@@ -10,13 +10,14 @@ import { Decimal } from "decimal.js";
 
 // ═══════════════════════════════════════════════════════════════════
 // TRACK-TOKENS-ORBS: AI credit system nested inside a parent credit system
-// Verifies that a single /track/tokens call deducts USD from the AI credit
-// feature AND deducts the ratio-mapped amount from any parent credit
-// system whose schema references it.
+//
+// Parent credit systems are overflow pools (same semantics as classic
+// metered → credits deduction order): a token track drains the AI credit
+// balance first, and only the overflow is ratio-mapped onto the parent.
 // ═══════════════════════════════════════════════════════════════════
 
 test.concurrent(
-	`${chalk.yellowBright("track-tokens-orbs: AI credit system inside parent credit system deducts both balances")}`,
+	`${chalk.yellowBright("track-tokens-orbs-1: AI balance covers the cost — parent orbs untouched")}`,
 	async () => {
 		const aiCreditsItem = items.free({
 			featureId: TestFeature.AiCredits,
@@ -24,7 +25,7 @@ test.concurrent(
 		});
 		const orbsItem = items.free({
 			featureId: TestFeature.Orbs,
-			includedUsage: 50_000, // 50,000 orbs
+			includedUsage: 50_000, // orbs schema: 1000 orbs per $1 of AI usage
 		});
 		const freeProd = products.base({
 			id: "free",
@@ -32,7 +33,7 @@ test.concurrent(
 		});
 
 		const { customerId, autumnV1, autumnV2 } = await initScenario({
-			customerId: "track-tokens-orbs",
+			customerId: "track-tokens-orbs-1",
 			setup: [
 				s.customer({ testClock: false }),
 				s.products({ list: [freeProd] }),
@@ -49,9 +50,6 @@ test.concurrent(
 			.div(1_000_000)
 			.toNumber(); // 0.125
 
-		// Orbs schema: 1000 orbs per $1 of AI usage
-		const expectedOrbsCost = new Decimal(expectedUsdCost).mul(1000).toNumber(); // 125
-
 		const trackRes: TrackResponseV2 = await autumnV2.post("/track_tokens", {
 			customer_id: customerId,
 			feature_id: TestFeature.AiCredits,
@@ -71,10 +69,61 @@ test.concurrent(
 			usage: expectedUsdCost,
 		});
 
-		// Parent orbs balance dropped by USD cost × 1000
+		// AI balance covered the full cost, so the parent overflow pool is untouched
 		expect(customer.features[TestFeature.Orbs]).toMatchObject({
-			balance: new Decimal(50_000).minus(expectedOrbsCost).toNumber(),
-			usage: expectedOrbsCost,
+			balance: 50_000,
+			usage: 0,
+		});
+	},
+);
+
+test.concurrent(
+	`${chalk.yellowBright("track-tokens-orbs-2: cost exceeding AI balance overflows into parent orbs at the schema ratio")}`,
+	async () => {
+		const aiCreditsItem = items.free({
+			featureId: TestFeature.AiCredits,
+			includedUsage: 100, // $100 of AI usage
+		});
+		const orbsItem = items.free({
+			featureId: TestFeature.Orbs,
+			includedUsage: 50_000,
+		});
+		const freeProd = products.base({
+			id: "free",
+			items: [aiCreditsItem, orbsItem],
+		});
+
+		const { customerId, autumnV1, autumnV2 } = await initScenario({
+			customerId: "track-tokens-orbs-2",
+			setup: [
+				s.customer({ testClock: false }),
+				s.products({ list: [freeProd] }),
+			],
+			actions: [s.attach({ productId: freeProd.id })],
+		});
+
+		// (5 * 24M) / 1M = $120 > the $100 AI balance
+		const trackRes: TrackResponseV2 = await autumnV2.post("/track_tokens", {
+			customer_id: customerId,
+			feature_id: TestFeature.AiCredits,
+			model_id: "custom/internal-model",
+			input_tokens: 24_000_000,
+			output_tokens: 0,
+		});
+		expect(trackRes.value).toBeCloseTo(120, 10);
+
+		const customer = await autumnV1.customers.get<ApiCustomerV3>(customerId);
+
+		// AI pool fully drained
+		expect(customer.features[TestFeature.AiCredits]).toMatchObject({
+			balance: 0,
+			usage: 100,
+		});
+
+		// $20 overflow lands on orbs at 1000 orbs per $1
+		expect(customer.features[TestFeature.Orbs]).toMatchObject({
+			balance: new Decimal(50_000).minus(20_000).toNumber(),
+			usage: 20_000,
 		});
 	},
 );

server/tests/integration/balances/track/basic/track-tokens-replay.test.ts

@@ -0,0 +1,120 @@
+import { expect, test } from "bun:test";
+
+import type { ApiCustomerV3 } from "@autumn/shared";
+import { ApiVersion, ApiVersionClass } from "@autumn/shared";
+import { TestFeature } from "@tests/setup/v2Features.js";
+import { items } from "@tests/utils/fixtures/items.js";
+import { products } from "@tests/utils/fixtures/products.js";
+import { initScenario, s } from "@tests/utils/testInitUtils/initScenario.js";
+import chalk from "chalk";
+import { Decimal } from "decimal.js";
+import { runQueuedTrack } from "@/internal/balances/track/runQueuedTrack.js";
+
+// ═══════════════════════════════════════════════════════════════════
+// TRACK-TOKENS-REPLAY: queued replay + plain value tracks on AI credit features
+//
+// When Redis fails open, track_tokens queues only the TrackParams body — the
+// token context (FeatureDeduction.tokens) is not serialized. The
+// replay worker rebuilds deductions from {feature_id, value}, so the USD value
+// must deduct 1:1 from the AI credit balance, exactly like the original token
+// track would have. Parent credit systems are overflow pools: untouched while
+// the AI balance covers the deduction (same as live track_tokens behavior).
+// ═══════════════════════════════════════════════════════════════════
+
+test.concurrent(
+	`${chalk.yellowBright("track-tokens-replay-1: queued replay body deducts AI credits 1:1")}`,
+	async () => {
+		const aiCreditsItem = items.free({
+			featureId: TestFeature.AiCredits,
+			includedUsage: 100, // $100 of AI usage
+		});
+		const orbsItem = items.free({
+			featureId: TestFeature.Orbs,
+			includedUsage: 50_000, // orbs schema: 1000 orbs per $1 of AI usage
+		});
+		const freeProd = products.base({
+			id: "free",
+			items: [aiCreditsItem, orbsItem],
+		});
+
+		const { customerId, autumnV1, ctx } = await initScenario({
+			customerId: "track-tokens-replay-1",
+			setup: [
+				s.customer({ testClock: false }),
+				s.products({ list: [freeProd] }),
+			],
+			actions: [s.attach({ productId: freeProd.id })],
+		});
+
+		// The USD cost computed by the original track_tokens call; only this
+		// survives in the queued body.
+		const usdCost = 0.125;
+
+		await runQueuedTrack({
+			ctx: { ...ctx, apiVersion: new ApiVersionClass(ApiVersion.V2_1) },
+			body: {
+				customer_id: customerId,
+				feature_id: TestFeature.AiCredits,
+				value: usdCost,
+				idempotency_key: `replay-${crypto.randomUUID()}`,
+			},
+			apiVersion: ApiVersion.V2_1,
+		});
+
+		const customer = await autumnV1.customers.get<ApiCustomerV3>(customerId);
+
+		expect(customer.features[TestFeature.AiCredits]).toMatchObject({
+			balance: new Decimal(100).minus(usdCost).toNumber(),
+			usage: usdCost,
+		});
+		expect(customer.features[TestFeature.Orbs]).toMatchObject({
+			balance: 50_000,
+			usage: 0,
+		});
+	},
+);
+
+test.concurrent(
+	`${chalk.yellowBright("track-tokens-replay-2: plain /track with a USD value deducts an AI credit balance 1:1")}`,
+	async () => {
+		const aiCreditsItem = items.free({
+			featureId: TestFeature.AiCredits,
+			includedUsage: 100,
+		});
+		const orbsItem = items.free({
+			featureId: TestFeature.Orbs,
+			includedUsage: 50_000,
+		});
+		const freeProd = products.base({
+			id: "free",
+			items: [aiCreditsItem, orbsItem],
+		});
+
+		const { customerId, autumnV1, autumnV2 } = await initScenario({
+			customerId: "track-tokens-replay-2",
+			setup: [
+				s.customer({ testClock: false }),
+				s.products({ list: [freeProd] }),
+			],
+			actions: [s.attach({ productId: freeProd.id })],
+		});
+
+		const usdValue = 5;
+		await autumnV2.post("/track", {
+			customer_id: customerId,
+			feature_id: TestFeature.AiCredits,
+			value: usdValue,
+		});
+
+		const customer = await autumnV1.customers.get<ApiCustomerV3>(customerId);
+
+		expect(customer.features[TestFeature.AiCredits]).toMatchObject({
+			balance: new Decimal(100).minus(usdValue).toNumber(),
+			usage: usdValue,
+		});
+		expect(customer.features[TestFeature.Orbs]).toMatchObject({
+			balance: 50_000,
+			usage: 0,
+		});
+	},
+);

server/tests/unit/agent/generate-and-update-agent-rules.test.ts

@@ -0,0 +1,82 @@
+import { beforeEach, describe, expect, mock, test } from "bun:test";
+import type { AutumnContext } from "@/honoUtils/HonoEnv.js";
+
+const generatedRules = {
+	entity_rules: { attach_to_entities: true, entity_feature_id: "deployments" },
+	credit_rules: { credit_feature_id: "credits" },
+	notes: "",
+};
+
+const mockState = {
+	existingNotes: "",
+	upsertCalls: [] as Record<string, unknown>[],
+};
+
+mock.module(
+	"@/internal/agent/workflows/generateAgentRules/generateAgentRules.js",
+	() => ({
+		generateAgentRules: async () => ({
+			rules: generatedRules,
+			metadata: { generated_from: "axiom" },
+			unconfigured: false,
+		}),
+	}),
+);
+
+mock.module("@/internal/agent/rules/repos/index.js", () => ({
+	agentRulesRepo: {
+		get: async () => ({
+			entity_rules: { attach_to_entities: false, entity_feature_id: "" },
+			credit_rules: { credit_feature_id: "" },
+			notes: mockState.existingNotes,
+			metadata: {},
+			org_id: "org_test",
+			org_slug: "test",
+			updated_at: null,
+		}),
+		upsert: async (args: { rules: typeof generatedRules }) => {
+			mockState.upsertCalls.push(args);
+			return { ...args.rules, metadata: {}, org_id: "org_test" };
+		},
+	},
+}));
+
+const { generateAndUpdateAgentRules } = await import(
+	"@/internal/agent/rules/actions/generateAndUpdateAgentRules.js"
+);
+
+const ctx = {
+	db: {},
+	org: { id: "org_test", slug: "test" },
+} as unknown as AutumnContext;
+
+describe("generateAndUpdateAgentRules", () => {
+	beforeEach(() => {
+		mockState.existingNotes = "";
+		mockState.upsertCalls = [];
+	});
+
+	test("preserves existing user notes when applying generated rules", async () => {
+		mockState.existingNotes = "Always attach add-ons at the customer level.";
+
+		const result = await generateAndUpdateAgentRules({ ctx });
+
+		expect(mockState.upsertCalls).toHaveLength(1);
+		expect(mockState.upsertCalls[0]).toMatchObject({
+			rules: {
+				entity_rules: generatedRules.entity_rules,
+				credit_rules: generatedRules.credit_rules,
+				notes: "Always attach add-ons at the customer level.",
+			},
+		});
+		expect(result.notes).toBe("Always attach add-ons at the customer level.");
+	});
+
+	test("keeps notes empty when none were saved", async () => {
+		await generateAndUpdateAgentRules({ ctx });
+
+		expect(mockState.upsertCalls[0]).toMatchObject({
+			rules: { notes: "" },
+		});
+	});
+});

server/tests/unit/balances/track/handle-track-tokens.test.ts

@@ -19,7 +19,14 @@ const featureDeductions = [
 	{
 		feature: { id: "ai_credits" },
 		deduction: 1,
-		precomputedCreditCost: 3.5,
+		tokens: {
+			usage: {
+				modelName: "openai/gpt-4.1",
+				inputTokens: 100,
+				outputTokens: 50,
+			},
+			cost: 3.5,
+		},
 	},
 ];
 

server/tests/unit/features/get-credit-cost.test.ts

@@ -0,0 +1,70 @@
+import { describe, expect, test } from "bun:test";
+import {
+	ErrCode,
+	type Feature,
+	FeatureType,
+	FeatureUsageType,
+} from "@autumn/shared";
+import { getCreditCost } from "@/internal/features/creditSystemUtils.js";
+
+// Uses custom/* models so pricing resolves offline (no models.dev fetch).
+const CUSTOM_MODEL = "custom/foo";
+
+const aiCreditFeature: Feature = {
+	internal_id: "fe_ai_credits",
+	org_id: "org_test",
+	created_at: Date.now(),
+	env: "sandbox" as Feature["env"],
+	id: "ai_credits",
+	name: "AI Credits",
+	type: FeatureType.AiCreditSystem,
+	config: { schema: [], usage_type: FeatureUsageType.Single },
+	archived: false,
+	event_names: [],
+	model_markups: {
+		[CUSTOM_MODEL]: { markup: 0, input_cost: 1000, output_cost: 2000 },
+	},
+};
+
+describe("getCreditCost — AI credit system without token context", () => {
+	test("self feature with no tokens maps 1:1 (plain /track values, queued replays)", async () => {
+		const cost = await getCreditCost({
+			featureId: aiCreditFeature.id,
+			creditSystem: aiCreditFeature,
+			amount: 5.25,
+		});
+		expect(cost).toBe(5.25);
+	});
+
+	test("self feature with no tokens defaults to a per-unit cost of 1", async () => {
+		const cost = await getCreditCost({
+			featureId: aiCreditFeature.id,
+			creditSystem: aiCreditFeature,
+		});
+		expect(cost).toBe(1);
+	});
+
+	test("self feature WITH tokens still prices through the model (not 1:1)", async () => {
+		const cost = await getCreditCost({
+			featureId: aiCreditFeature.id,
+			creditSystem: aiCreditFeature,
+			modelName: CUSTOM_MODEL,
+			tokens: { input: 1000, output: 500 },
+		});
+		// (1000 * 1000 + 2000 * 500) / 1_000_000 = 2.0
+		expect(cost).toBeCloseTo(2.0, 10);
+	});
+
+	test("non-self feature with no tokens throws", async () => {
+		expect(
+			getCreditCost({
+				featureId: "some_other_feature",
+				creditSystem: aiCreditFeature,
+				amount: 5,
+			}),
+		).rejects.toMatchObject({
+			code: ErrCode.InvalidRequest,
+			message: expect.stringContaining("modelName and tokens"),
+		});
+	});
+});

server/tests/unit/features/get-model-pricing.test.ts

@@ -113,3 +113,20 @@ test("fetch failure with no stale cache throws InternalError", async () => {
 		message: "Failed to fetch models.dev pricing and no cache available",
 	});
 });
+
+test("fetch carries an abort timeout so a hanging models.dev cannot hang tracks", async () => {
+	let capturedSignal: AbortSignal | undefined;
+	globalThis.fetch = Object.assign(
+		async (_input: unknown, init?: RequestInit) => {
+			fetchCalls++;
+			capturedSignal = init?.signal ?? undefined;
+			return Response.json(pricingData);
+		},
+		{ preconnect: realFetch.preconnect },
+	) as typeof fetch;
+
+	await getModelsDevPricing();
+
+	expect(capturedSignal).toBeInstanceOf(AbortSignal);
+	expect(capturedSignal?.aborted).toBe(false);
+});