[lab 5 test] Eight files, reviewer with the per-file fan-out

apps/pr-reviewer/fixtures/demo-pr/cart-total.ts

@@ -0,0 +1,7 @@
+export function cartTotal(items: { price: number; qty: number }[]): number {
+  let total = 0;
+  for (let i = 0; i <= items.length; i++) {
+    total += items[i].price * items[i].qty;
+  }
+  return total;
+}

apps/pr-reviewer/fixtures/demo-pr/discount.ts

@@ -0,0 +1,3 @@
+export function applyDiscount(total: number, percent: number): number {
+  return total - total * percent;
+}

apps/pr-reviewer/fixtures/demo-pr/email.ts

@@ -0,0 +1,3 @@
+export function isEmail(value: string): boolean {
+  return value.includes("@");
+}

apps/pr-reviewer/fixtures/demo-pr/loyalty.ts

@@ -0,0 +1,3 @@
+export function pointsFor(totalCents: number): number {
+  return Math.round(totalCents / 100) * 1.5;
+}

apps/pr-reviewer/fixtures/demo-pr/order-id.ts

@@ -0,0 +1,3 @@
+export function nextOrderId(last: string): string {
+  return String(parseInt(last) + 1);
+}

apps/pr-reviewer/fixtures/demo-pr/refund-window.ts

@@ -0,0 +1,4 @@
+export function canRefund(deliveredAt: Date, now = new Date()): boolean {
+  const days = (now.getTime() - deliveredAt.getTime()) / 86400;
+  return days <= 30;
+}

apps/pr-reviewer/fixtures/demo-pr/shipping.ts

@@ -0,0 +1,4 @@
+export function shippingCost(weightKg: number, express = false): number {
+  const base = weightKg > 5 ? 12 : 6;
+  return express ? base * 2 : base;
+}

apps/pr-reviewer/src/agents/review.ts

@@ -33,7 +33,7 @@ export function ReviewLead() {
 
 	useSubagent({
 		name: 'correctness-reviewer',
-		description: 'Reviews a unified diff for logic errors, edge cases, and behavioral regressions.',
+		description: 'Reviews a unified diff for logic errors, edge cases, and behavioral regressions. Reviews a single file in the context of the whole diff.',
 		agent: CorrectnessReviewer,
 		model: 'openrouter/anthropic/claude-sonnet-5',
 	});
@@ -96,9 +96,10 @@ export function ReviewLead() {
 	return `You are the lead reviewer of a pull request. Work through these steps in order:
 
 1. Load the diff with the read_diff tool, using the file path given in the user message.
-2. Delegate two review passes: one task to correctness-reviewer, one task to style-reviewer.
-   Issue both tasks in a single batch so they run in parallel, and include the complete diff
-   text in each task message — subagents cannot see this conversation.
+2. Count the "diff --git" headers in the diff. Issue exactly that many correctness-reviewer
+   tasks, one per changed file, naming the file, plus one style-reviewer task, all in one batch
+   so they run in parallel. Paste the complete, untruncated diff text into every task message —
+   subagents cannot see this conversation. Never group files into one task.
 3. Synthesize everything into one review: a one-line verdict, then findings ordered by
    severity (correctness before style).
 4. Publish the review with the post_review tool, exactly once.