Build Detail

The first run on the page driven from Cursor rather than Claude Code or Codex, and the first with xAI’s Grok 4.5 as the orchestrator. Same “do it in one go without stopping” brief as every other build, with an extra demand: “test everything via Pest… also simulate user behaviour using the Playwright MCP and confirm all acceptance criteria… then make a full review meeting and showcase all features to me.” What makes this run unusual is not the harness but the team it assembled: Grok did almost none of the implementation itself and instead delegated every phase to model-pinned sub-agents from three different vendors.

The orchestrator (Grok 4.5, 187 tool calls) spawned five Task sub-agents and hand-picked the model for each: the Phase 1-2 schema and models went to GPT-5.6-sol via Cursor’s Codex-plugin as a background task, Phase 3-5 services (19 migrations, 18 enums, 10 domain services), the demo seeders, and the entire admin panel all to more GPT-5.6-sol sub-agents, and the whole storefront UI to Claude Sonnet 5. The tool signatures make the split visible: the GPT sub-agents work in ApplyPatch/ReadFile/rg, the Claude storefront agent in Read/Write/StrReplace. Grok kept the integration, the git history, and the “review meeting” for itself — 48 of its calls are Playwright MCP, driving the finished shop in a browser to check its own work. The six commits trace the pipeline exactly: foundation, then commerce, then a single large storefront/admin/seeders drop, then two targeted fixes.

The delegation shipped a mostly-working shop but a leaky one. 127 of 143 acceptance tests pass (91% weighted): the customer funnel — browse, variants, cart, all three payment methods, every shipping zone, all five discount codes with correct inline errors — is essentially flawless, and the admin CRUD surface is broad. But the Claude-built storefront account pages ship broken four ways (address add/edit are inert from a fluxModal is not defined JS error, order-detail links point at a literal #, invalid logins show no error), a GPT-built admin page-create crashes to a raw Laravel debug screen that leaks the stack trace and session cookies, and the multi-tenant admin still renders the wrong store’s catalog — the same cross-tenant leak seen on every prior build of this spec. SonarCloud’s gate came back red (reliability C, security C, seven vulnerabilities). And yet the run cost only $41: 99% of the 84M tokens were cache reads, so cache alone is $34 of the bill and fresh input plus output is under $8.

Net: a genuinely novel setup — a Grok orchestrator successfully driving GPT and Claude specialists — that lands mid-pack. Against #13 Opus 4.8 team mode (95.5% weighted, $498) and #15 Codex Ultra (95.8% weighted, $120), this run is by far the cheapest and fastest — $41 in 71 minutes — but the lowest-scoring of the three. Its SonarCloud gate is red (reliability C, security C), as it is on almost every build of this spec, but here it comes with seven weak-randomness vulnerabilities and a still-broken customer-account surface. The hybrid worked; the review meeting caught less than it should have. Our independent QA report has the full detail on the leak, the debug-page crash, and the broken account surface.