Build Detail

The first Kimi Code run on the page, and the first billed on a flat subscription rather than pay-per-token: Kimi K3 Max at Moonshot’s $39/month Allegretto plan, thinking effort set to max. Same “do it in one go without stopping… test everything via Pest… simulate user behaviour using the Playwright MCP… then make a full review meeting” brief as every other build. What sets this run apart is not the harness’s ambition but its economics: “this build required 157% of the weekly limit and multiple 5h limits,” as the operator put it, and the session spans a full week of wall-clock time because of it.

Kimi split the spec into a strict 12-phase sub-agent pipeline (foundation → catalog → themes → cart → payments → accounts → three admin-UI phases → search → analytics → apps/API → seeders → polish), each phase a single Agent call handed a full written report back to the main thread before the next was spawned — closer to Codex’s sequential goal-mode than to Claude’s team mode. For the final QA pass it switched tools entirely: one AgentSwarm call fanned out 5 parallel browser-test shards covering all 143 acceptance tests. What makes the transcript unusual is how often it had to restart: 7 separate provider-imposed stalls interrupt the run, from short 429 rate-limit backoffs up to a 132-hour wait for the weekly quota to reset after Phase 11, and the first AgentSwarm attempt was killed outright by a 403 usage-limit error and had to be resumed hours later, after a second same-day quota window passed. Each time, a plain “continue” was enough to pick the same phase back up mid-task.

The result lands solidly in the upper half of the page: 134 of 143 acceptance tests pass (94.1% weighted), the fourth-highest score here behind #14 Fable 5 (99.7%), #15 Codex Ultra (95.8%), and #13 Opus 4.8 team mode (95.5%). The customer funnel, admin CRUD, discounts, and multi-tenant storefront isolation are essentially solid. But the failures cluster in one place: the entire customer-account section has a systemic click-handling bug — order-history “View” links, address add/edit, and logout all fail to respond to clicks despite correct hrefs and working routes underneath (direct navigation and a manual form.submit() both succeed), pointing at one shared front-end regression rather than four separate ones. On top of that, the familiar cross-tenant admin catalog leak survives yet again, and the admin order-status filter and product-archive toggle are both broken. SonarCloud’s gate failed on reliability alone (C) with security and maintainability both clean (A) — a milder profile than most builds here, and consistent with 226 classes at the second-highest average maintainability index (96.4) on the page, behind only #13’s 104.5.

Net: a solid mid-pack-to-upper build, kept out of the top three by one systemic front-end bug that costs four test failures at once. The number that actually differentiates this run is time, not tokens — 20h 21m of genuine active work stretched across 7 days 5h47m of wall clock by repeated provider quota exhaustion, at an equivalent OpenRouter cost of $190 for a subscription that nominally costs $39/month. Against #14 Fable 5 (99.7% weighted, $423, 6h58m) this run is cheaper but noticeably less complete, and the quota stalls make it operationally the least predictable build on the page — a capable model bottlenecked by its own billing plan as much as by its own competence.