Build Detail

The first build on this page run through Cursor’s terminal CLI (cursor-agent v2026.09.10-fd3934a) rather than its IDE background composer, and the first Cursor run to use a single model throughout instead of pinning a different vendor per phase: Meta’s Muse Spark 1.3, a 1M-context model, in Cursor’s “MAX” reasoning mode, for the coordinator and every one of its sub-agents. The brief is the familiar one-shot instruction used across this page, with one addition none of the others spelled out this explicitly: “If you find bugs, you must fix them… In case bugs appear, you must fix them all and restart the review meeting.” — a self-correction loop written directly into the prompt, not left to the agent’s own initiative.

The coordinator took that instruction literally. Six implementer sub-agents ran strictly sequentially, each one starting the instant its predecessor’s summary landed — Database + Models (Agent A) → Foundation/Tenancy/Auth/RBAC (B) → Catalog + Search (C) → Commerce Core (D) → Storefront UI + Customer Accounts (E) → Admin Panel + Analytics (F) — a genuine dependency pipeline rather than the parallel-fan-out pattern seen on most other sub-agent builds here. What follows is the interesting part: an independent verification agent (G) browser-tested the finished app via Playwright MCP and reported 18 bugs; a bugfix agent (H) fixed all 18 and added regression tests; a second verification pass (I) re-checked every fix in a live browser and found 3 remaining FAILs; H was re-spawned for a second bugfix round; and a third and final I pass confirmed everything green before the coordinator wrote its own “Review Meeting” demo-access summary. That is 11 sub-agent spawns across 9 lettered roles (H and I each ran twice) plus the coordinator — 2.6K tool calls in total — and it is a rare case on this page of an agent-driven QA loop that actually iterated on its own findings instead of reporting a single pass/fail verdict.

The result is the largest codebase on this page by class count (295 PHP classes) and, on our own independent 143-test Playwright pass, a strong 93% weighted score (132 pass, 9 fail, 2 partial) with zero critical bugs — no data corruption, no cross-tenant leak, notable given several higher-numbered builds on this exact spec have shipped one or the other. But the defects that did survive two internal bugfix/verification rounds cluster tightly in one place the run’s own G/H/I loop apparently never specifically probed: customer-account click interactions. The Addresses page’s Add/Edit buttons never open their modal <dialog>; the “Log out” button fires zero network requests on most clicks despite a working backend route; and order “View” links silently no-op because Cursor’s own coordinator model appears to mishandle a URL-encoded # in the order-number path when Livewire’s client-side router intercepts the click. All three are real, user-facing, and all three passed the build’s own two rounds of Playwright-driven self-verification untouched.

Net: at an OpenRouter-equivalent $44.92 (Cursor itself bills this as “Included” subscription usage — in the user’s own words, “17% of the monthly token contingent of my Cursor Ultra plan”) and 2h 45m for a from-scratch build with a genuine multi-round self-correction loop, this lands 6th of 19 builds on this page by score — ahead of the prior Cursor run (#16 Grok 4.5, 90.6% weighted for a comparable $41) on both score and cost, and clean of the cross-tenant leak that run repeated. Only #18 Codex Luna scores higher for less money. Our independent QA report has the full detail on the three interaction bugs the run’s own verification missed.