Build Log · 004

EVCO · Cloudflare Pages + Workers + D1

Seven hours forty-five minutes wall-clock. The slow part was never the AI.

An EV charging station catalog with admin CRUD, magic-link auth, and a public site with category filters. Cloudflare Pages plus Workers plus D1. The headline figure is wall-clock, not active build, because the composition decision was the real time investment.

Builderiris + sally + (implementation attributed to amelia)
Modelmimo-v2.5
StackCloudflare Pages + Workers + D1
Engagement2026-06-03 to 2026-06-04
StatusShipped

The brief

Build a catalog site for electric vehicle charging stations. Public product list with category and feature filters. Admin CRUD for stations, categories, contact submissions. Magic-link authentication via MailChannels. Cloudflare Pages plus Workers plus D1; multi-tenant from day one.

The Coach assigned: iris + sally + (implementation attributed to amelia). iris + sally + (implementation attributed to amelia) shipped.

The problem

EV charging is a public-info domain that gets cataloged badly. Stations, ports, connectors, prices, photos, availability. The simple version of the product is a filterable list, a station page, a small contact-form CRM, and an admin panel for whoever's curating the catalog.

The brief was the small honest version: ship the catalog and the admin, on a sovereign-ish stack (Cloudflare so the data and the edge are in the same place), with magic-link auth so the admin doesn't fight a password reset flow.

How the factory shipped it

The Coach selected the team. The factory wrote the code. Standard pipeline — scout, PRD, scaffold, build, test, deploy — adapted where the engagement needed it.

Why this team: Design-led pipeline because the visual coherence mattered for a public catalog. iris on art direction, sally on mockups, the implementation attributed to amelia in the DONE markers but with a session-log gap that the build log explains below.

Agent team architecture:

The pipeline:
  scout       carson      —  the intake
  plan        mary        —  the PRD
  scaffold    phil        —  the project shell
  design      iris        —  art direction
  design      sally       —  mockups and UI specs
  build       amelia      —  implementation (attribution; see caveats)
  test        testlead    —  the test suite
  qa          qa          —  the verification pass

What got delivered

The numbers

All figures verified from agent session logs. Anything that can't trace to a log stays off the page.

7h 45m
Wall-clock window
first DONE marker to last DONE marker
$4.50
Build cost estimate
design + verified implementation attribution
43
QA tests passing at hand-off
of 46 total; 3 fail, one post-build commit closes them
CF Pages
Core stack
plus Workers plus D1; multi-tenant from day one

The timeline

2026-06-03 ~15:00 UTC
carson completes the intake
2026-06-03 17:00 UTC
phil completes the scaffold
2026-06-03 18:33 UTC
iris opens session: art direction pass, 88% EVCO content
2026-06-03 20:29 UTC
iris opens the largest design session: 165 path mentions, 100% EVCO
2026-06-04 12:50 UTC
iris post-build review session
2026-06-04 ~22:45 UTC
testlead closes the test suite; QA reports 43 of 46 tests passing
2026-06-05
Post-build commit: fixes DELETE 404 plus adds auth rate limiting at 5 req per hr

What broke

The amelia attribution gap

The DONE markers attribute the 4.x implementation phases to amelia. No amelia session JSONL on disk covers the build window. The implementation happened — the artifacts and the QA pass exist — but the cost-side telemetry was cleaned up by the session-store rotation before the archeology pass ran. The build log records what we can prove, not what we'd like to claim.

Three QA failures at hand-off

43 of 46 tests passing means three failures sitting in the QA report at close-of-build. The post-build commit closes them: a DELETE endpoint returning 404 instead of the expected error, plus auth rate-limiting added at 5 requests per hour to harden the magic-link flow against probing. Both fixes documented in git.

Design cost is partly proportional

iris contributes a cleanly verified $0.19 from short, isolated sessions. sally contributes about $3.08, but that's a proportional allocation from a 35-day multi-project session at 7.1% EVCO path-share — defensible as estimate, not measurement. The methodology range is in the caveats.

The catalog has more pages than the data model

About 6 user-facing pages against a 4-entity data model. The page count drives more layout work than the data work. iris and sally spent more time on design coherence than amelia spent on data plumbing. That ratio is normal for catalog-shape builds and matches Teardown reference-set ratios.

Why it worked

Results

What we can’t claim

The verified portion is iris's $0.19 across 5 sessions at 100% EVCO path-share. The proportional portion is sally's $3.08 from a 7.1% share of a multi-project session — defensible as estimate, not measurement.

A v3 method calibration in June 2026 produced a credible range. Full cost honestly lands somewhere in $0.40 (verified floor) to $7.73 (verified plus proportional), best-guess ~$3–5. Method documented internally.

The amelia DONE markers for 4.x implementation phases do not have matching session JSONLs on disk. The cleanup commits that ran later swept the relevant sessions before the archeology pass. The implementation cost is not in the budget; the implementation artifacts and QA results are in the deliverables.

“Tell me what you need built.”

Hire the Coach

Zero human intervention. The Coach selected the team. The factory shipped.

← Back to the factory