Build Log

Commons 50-Mission Seed — Real Bugs Found Building It, Not Just Content

August 14, 2026

Seeded 48 real Commons Missions, then found and fixed a real Railway production-build failure, a real mobile-layout bug, and a real profile-matching taxonomy bug that had likely never worked for any user.

AlloyCommonsForgeKit OSBusiness Strategy

new Commons Missions seeded live

real production bugs found and fixed

workflows run (2 architecture reviews + 1 business-strategy pressure test)

gates from 'use server' non-function exports (1 new, fixed by moving to plain modules)

Timeline

Start

ForgeKit OS v2 architecture review

Workflow: repo-tracing + external research + judge panel + adversarial verify — concluded OS 1.x/Alloy 1.x, no version bump

Work-as-root-abstraction pressure test

Second workflow tested a GPT-authored 'evidence-native OS / Attention Router' thesis against the same evidence — unanimous 'no' on all 8 proposed capabilities

Business-strategy pressure test

Third workflow tested a 'don't sell ForgeKit OS, sell operational software' memo — confirmed sound, but flagged its own analysis as itself risking the failure mode it warned against

Memory writes + customer-#2 leave-behind

Architecture Gravity + PROBE/PILOT/PRODUCT memories written; iterative one-page leave-behind built and refined through 5 real feedback rounds

Network Miner probe built

Local LinkedIn-connections scorer, calibrated against 8 real people already discussed in conversation

Commons 50-mission seed

Seeded 48 missions live, built a real image-upload feature + Relay image variant along the way, found 2 real gaps in the registry/query layer

First deploy attempt failed

Discovered mid-deploy that Railway watches railway-proof, not master — merged and pushed correctly

Real production-build failure found and fixed

'use server' file exporting non-function values — 3 instances (2 pre-existing, 1 introduced this session), all fixed and build-verified

Mobile UX + taxonomy bugs found and fixed

Aisle view clipped on real phones; all 48 missions showed 'Community'; profile-interest matching used two disjoint vocabularies and had likely never worked

What shipped

Two full multi-agent architecture-review workflows (OS v2 decision + business-strategy pressure test) — all findings verified against real repo evidence, not asserted

Two new auto-memory entries: Architecture Gravity, PROBE/PILOT/PRODUCT vocabulary

A polished, iterated one-page HTML leave-behind artifact for customer-#2 outreach (products/alloy work, but the artifact itself lives outside the repo)

experiments/network-miner/ — a local, no-web-calls LinkedIn connection scorer (PROBE)

products/alloy/lib/commons/upload-image.ts + upload-image-types.ts — Commons' first real image-upload feature (@vercel/blob)

products/alloy/lib/commons/contribution-kinds/blind-continuation-v1.ts extended with mediaType: 'text' | 'image'

products/alloy/lib/commons/queries.ts's getRelayPreviousPiece (replaces getRelayPreviousPieceBody, was blind to image pieces)

products/alloy/scripts/seed-commons-missions-v2.mjs — seeds 48 real Missions/Workpieces, idempotent

products/alloy/scripts/revert-commons-missions-v2.mjs — real, tested revert path reading its slug list from the seed script's own source

48 mission images copied to products/alloy/public/commons/mission-images-50/, wired into MISSION_PHOTOS

products/alloy/lib/commons/onboarding-shared.ts — extracted from onboarding.ts (use-server export-rule fix), ONBOARDING_TOPICS expanded 7 → 15 real values

components/commons/primitives.tsx's MISSION_CATEGORIES — 48 new real per-mission category assignments, plus 3 case/synonym bugs fixed in the original 13

/commons/explore now serves the list view by default; the Aisle moved to /commons/explore/wander with a toggle link each way; /commons/explore/list redirects

“Avoided work has economic value too. The danger begins now, if we run another architecture review instead of acting.”
— Zeb, closing the 3-workflow architecture-review chain — correcting Claude's own framing of the review as possibly-wasted analysis